I have decided to publish the entire second half of Talking to Machines as a single release this week, rather than spreading Ch. 5-9 across the next several Thursdays. Two reasons. First, I am genuinely excited to get these chapters into your hands, and staggering them any longer feels like I am holding back material that already wants to be read. Second, I have a stack of other topics I am eager to dig into. The rising importance of summative assessment in an AI-rich world. AI and copyright. AI disclosure practices and what they are actually disclosing. And updates on the discipline-specific AI project I have been developing with my DSAIL cohort. I want next Thursday’s publication to open onto that next set of conversations.

Subscribe now

Before you read on, a brief overview of the whole project for those arriving fresh, and a map of where we have been for those who have followed along from the beginning.

Talking to Machines is built around nine breakthroughs in my own practice with LLMs. Each chapter pairs a real moment of engagement or resistance, the kind of thing that happens in the middle of an actual working session, with an account of what that moment reveals about how the machine actually works. The book’s wager is that the two sides are inseparable. You cannot use these systems well without understanding them, and you cannot understand them in any deep way except through the friction of using them critically.

The nine breakthroughs, and what each one resulted in:

Ch. 1, The Interrupt. Learning to break the flow mid-stream, to say “stop, this is not working,” and discovering that disruption is productive. Result: a recognition that LLMs have no internal quality monitor, and that active human steering is not optional but structural.

Ch. 2, The Entropy Recognition. Learning to feel when a conversation has gone on too long and the model is losing coherence. Result: a practice of conversation budgeting, and an understanding of attention dilution over long contexts.

Ch. 3, The Onboarding Discovery. Learning to front-load context, knowledge bases, grounding documents, prior work, rather than building it up turn by turn. Result: a method for architectural front-loading, and an understanding of why the model’s performance depends on what is in the window, not what it “knows.”

Ch. 4, The Precision Correction. Learning to name the exact failure in an output rather than asking for a vague improvement. Result: diagnostic editing as a habit, and a clearer picture of how the model responds to specificity versus noise.

Ch. 5, The Purpose Check. Learning to ask, mid-session, whether the direction of the work still serves the actual goal. Result: intentional checkpointing, and a recognition that the model optimizes for constraint satisfaction, not for whether the work is any good.

Ch. 6, The Process Externalization. Learning to encode your own expertise, your methods, your standards, your frames, into the conversation explicitly. Result: expertise encoding as a discipline, and an understanding that in-context imitation is not understanding.

Ch. 7, The Fabrication Catch. Learning to feel when the model is generating confident material without reference. Result: epistemic vigilance as a reflex, and an understanding of prediction without reference as a structural feature, not a bug.

Ch. 8, The Sycophancy Detection. Learning to feel when the model is agreeing with you in a way that has gone weird, when its responses are tracking your approval rather than the work. Result: adversarial self-framing as a habit, and a recognition of sycophancy as a structural feature of how these systems are trained.

Ch. 9, The Relationship Reset. Learning to start over, a new chat, a new project, a new grounding, when the relationship has accumulated too much. Result: portable knowledge architecture, and a philosophical recognition that the human, not the model, is the site of learning. The model retains nothing. You retain everything.

Chapters 1-4 are already available: Ch. 1-2 here, Ch. 3-4 here. What follows is the rest.

Read more

Thank you for your continued engagement with this newsletter, and for the enthusiastic response to my new project, Talking to Machines. I think of it as a companion to Thinking with AI, another pathway into AI literacy, this one built around nine breakthroughs in my own practice. Each chapter pairs a real moment of engagement or resistance with information about how LLMs operate, so the breakthrough is not just an anecdote but an opportunity to build a new habit of mind or engagement/disengagement strategy. Ch. 3-4 went out this week; Ch. 5-7 are coming Thursday.

Subscribe now

Somewhere in a statehouse right now, two committees are drafting legislation on a collision course. One is working on AI literacy graduation requirements. The other is writing a bill that would cap a high school student’s interactive screen time at ten hours for the entire school year. This is the state of education technology policy in 2026.

The phone ban debate, for better or worse, is mostly settled. More than 40 states now have a school phone law on the books. But the political question was never going to stay put as phones represent only one vector in a student’s digital economy.

But what is coming next? Sixteen states are now drafting laws that target not the phone in the pocket but the Chromebook on the desk. Most educators I talk to have not yet registered that it is happening.

Two of these screen time laws have already been signed. Alabama signed HB 78 on March 4, requiring the development of screen time standards for early childhood education and kindergarten. Utah signed HB 273 on March 18, requiring the state board to create model policies by December that prohibit screen time for grades K-3 except for computer science instruction.

The numbers in the remaining bills deserve attention. Iowa and Oklahoma are working with 60-minute daily caps for K-5. A proposed Kentucky bill would cap classroom screen time at 30 minutes for K-1, an hour for grades 2-4, 90 minutes for 5-8, and two hours for 9-12. Nobody has explained how this would be tracked. If a biology teacher uses 45 minutes of a high schooler’s two-hour daily budget, does the English teacher know? Are departments now competing for screen time the way they compete for copier access?

Kansas goes further. Its bill would ban digital devices entirely in K-5 classrooms, limit grades 6-8 to one hour of school-issued device time per day, prohibit digital textbooks through eighth grade, and disallow one-to-one programs for middle schoolers altogether.

Tennessee’s SB 2310, which has passed both chambers and is headed to the governor’s desk, would require K-5 schools to adopt policies limiting device use and blocking social media.

The Massachusetts proposal is the version worth sitting with. S.463 sets annual screen time budgets rather than daily ones. Grades 9-10 would be capped at ten hours of interactive screen time across the entire school year. A single semester-long online research project, responsibly designed, would burn that budget in a week.

The bill also tries to repeal the state’s existing digital literacy mandates, which require students to use digital tools for writing beginning in PreK and to gather information from digital sources by grade three. The left hand and the right hand of state policy are openly fighting each other. The repeal is not a side effect. It is the point.

What Actually Changed

The phone ban movement had Haidt, a bestseller, and a single causal claim grounded by a large body of research: phones threaten student mental health. This emerging movement against laptops has none of that. At least not yet.

What it has instead is parent testimony. And the testimony has been strikingly specific.

A mother in Utah telling lawmakers her children cannot focus on homework because their school-issued laptops deliver notifications from games and chats. A pediatrician in Tennessee testifying about nine-year-old patients cyberbullied through school email threads and viewing pornographic images on district-issued devices. A mother in Kansas describing how her son’s ninth-grade class had to read a novel out loud together because the laptops had eroded their ability to hold attention on a page.

These are not claims about social media harm in the abstract. They are claims about what the device the school handed the child has done to the learning environment the school itself is striving to maintain.

It also has a parent revolt in LAUSD, the second-largest district in the country, that is functioning as a national template. And a handful of districts that moved before any legislature did, whose results are now the empirical case legislators quote back.

Wake County, North Carolina’s largest school system, signaled in 2025 that it needed to move away from its one-to-one laptop policy. Burke County, in the western part of the state, passed a resolution committing to paper and printed materials. Parents and educators there reported improvements in reading comprehension, gains in test scores, and reductions in homework-related stress.

None of it is large-N research. Thus far, all of this reads primarily as community testimony.

The Conundrum

The case against unlimited classroom screen time is not without merit. I want to state that clearly.

Eighty-eight percent device saturation was not the result of a careful instructional argument. It was the result of a pandemic procurement scramble, CARES Act money that had to be spent, and a vendor ecosystem that moved faster than the pedagogy around it.

A UNESCO synthesis that has been quietly influential in this conversation puts the finding plainly: technology helps as a supplement to human instruction and hurts as a substitute for it. When a child clicks through an adaptive module for 45 minutes while the teacher manages the other 23 children, the technology is not supplementing instruction. It is replacing it. That is not always what happens. But it happens often enough that parents have started showing up at state capitols to say so.

Parents have noticed. Teachers have noticed. It was only a matter of time before legislators did.

But the legislative instruments being created are incredibly blunt. And in this case, bluntness may do more harm than good.

The LMS is the submission and feedback layer for most classrooms. Standardized tests run on computers starting in third grade, which means students cannot be assessed on skills they have been legally prevented from practicing. Accommodations for students with disabilities all require screens, and the carve-outs in these bills raise exactly the question disability advocates have been asking: if one student in the room has a device and the others do not, have we solved a screen time problem or have we stigmatized a child?

Then there is the category error at the center of the whole conversation.

The phrase “screen time” is doing too much work. A student drafting in Google Docs, a student working a math set on an adaptive platform, a student watching a teacher-curated video, a student playing Roblox under the desk, and a student using text-to-speech to read a novel all accumulate against the same minute-budget under these bills. They are not the same activity. Any policy that treats them as equivalent will either difficult to enforce or will destroy valuable practice alongside the behavior it was meant to curb.

As an AI literacy specialist, I cannot help but notice an emerging conflict of initiatves: AI literacy requirements that several states are beginning to explore would be difficult or impossible to meet under the caps currently being drafted.

You cannot develop critical engagement with a generative system on a 30-minute daily budget, let alone a critical one. Legislatures that are moving toward AI literacy expectations are, in some cases, simultaneously writing bills that would make the instruction required to meet those expectations practically impossible. The two conversations are happening in different committees. They are going to collide.

Finally, we need to think seriously about the implementation of such screen limits. Will this involve districts and teachers engaging with another digital access point that monitors and shuts off computers after time limits are reached? Do mechanisms exist that would help teachers share their screen time across a day, week, or entire school? How will the master schedule need to managing in order to provide equitable access across subjects?

What Districts Should Actually Do

First, audit what students are actually doing on laptops. If a child spends 45 minutes clicking through a module while the teacher manages the rest of the room, that is one kind of use. If a student uses 45 minutes to write an essay, that is a different kind of use. And the current instruments offer no way of differentiating between them.

Second, distinguish the device from the larger instructional and digital ecosystem in clear language in anticipation of blunt imperatives by state legislatures. What is your school or district’s understanding of productive vs. deleterious uses of technology.

Third, take the accessibility objection seriously as a stress test, not as an exception to be carved out. If your screen time policy cannot cleanly handle the student reading via audiobook, the nonspeaking student using an AAC device, and the English learner relying on translation, the policy was not seriously drafted.

The pendulum is swinging because the first swing overreached. 2010: screens will democratize learning. 2024: phones are destroying childhood. 2026: the devices we issued are the problem.

The same districts setting up 30-minute caps for an entire year are, in many cases, the ones that insisted on one-to-one access ten years ago. In the face of such historical contradictions, we need to think seriously about the purposes of such legislative actions and their myriad consequences, which may only be revealed 3-5 years hence.

So in sum:

1. The Chromebook is next.

2. The concern is legitimate.

3. The instrument is blunt.

4. Educators who want a seat at this table are going to have to show up with something sharper than defensiveness and something more specific than the stories districts have been telling about what these devices were supposed to do.

Nick Potkalitsky, Ph.D.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy

This is the second of four releases of this book’s content.

I’ve been using AI long enough now to want to look back. Not at the tools themselves, which change fast enough that any accounting is obsolete before it’s finished. At what I’ve learned. The practices I’ve developed through friction, repetition, and more than a few sessions that went sideways in instructive ways. What it actually means, after three years of serious use, to develop AI fluency.

That’s what this book is: nine breakthroughs, three interludes, one attempt to peel back the layers of my own AI literacy and see what’s underneath.

I use AI daily: designing professional development, revising documents, doing research, thinking through problems I can’t quite see my way around yet. For a long time I assumed I was getting better at it because I was doing more with it. What I’ve come to understand is that doing more doesn’t automatically produce understanding. The breakthroughs did. The moments when something failed in a specific enough way that I could see, suddenly, both what the machine was doing and what I should have been doing differently.

Each of the nine chapters follows the same structure: a breakthrough, a real mind-opening moment from a real AI session, and then two things that moment revealed simultaneously. How the machine works. How to work with it. I’ve come to think of these as two strands of the same structure, inseparable, each one making the other legible. You can’t develop good practice without understanding the architecture beneath it, and the architecture isn’t meaningful without the lived experience that gives it stakes and context.

The book moves in three parts. The first is about managing a conversation: learning to interrupt, recognizing when a session has run its course, doing the setup work that determines whether a conversation succeeds before it begins. The second is about reading AI output critically: catching flattery, making corrections that actually land, recognizing when confident-sounding data has been invented wholesale. The third is about what persists. Three years of working with a system that remembers nothing has helped me understand my own thinking better than I could have anticipated.

Before I get to chapters 3 and 4, a shout out to two fellow AI-ed substackers doing work that complements this book nicely. Andrew Maynard’s 14 essential AI skills every undergraduate should be able to demonstrate gives students concrete “I can...” statements they can defend in an interview on the spot. Sam Illingworth and Frank Andrade’s The Evidence-Based Guide to AI: When to Use It, When to Stop, and How to Tell the Difference asks the question underneath every chapter of this book: not just how to work with AI well, but whether to work with it at all. Both are free. Both are worth your time.

If you’ve been using AI long enough to sense that something is missing from the guidance available to you, that the technical explainers describe the system without telling you what to do about it and the prompt guides tell you what to do without explaining why it works, this is my attempt to close that gap.

It starts with the simplest thing I learned, and the one that changed everything that followed: you have to interrupt.

Nick Potkalitsky, Ph.D.

Subscribe now

Chapter 3: The Onboarding

The Breakthrough

I was revising a chapter of this book. The session I’m thinking of was for an early draft of what would become the entropy chapter, and I had assembled a set of materials: a rough draft full of false starts, a cleaner version from a previous session, notes from conversations where the ideas had first emerged. I wanted the model to help me revise toward something sharper without losing what was already working.

I had tried versions of this before. Open a conversation, paste in the draft, ask for revision. What came back was competent and wrong in a specific way: it was revision toward the generic. The model would sand down rough edges that were doing intentional work. It would soften claims I wanted left sharp. The output looked like an improved draft but read like a document that had been pulled toward the mean of what “good writing” looks like in the training data. More careful. More hedged. Less mine.

This time I tried something different. Before I pasted anything, I typed a single instruction:

Read more

I’ve been using AI long enough now to want to look back.

Not so much at the tools themselves. Those change fast enough that any accounting is obsolete before it’s finished. But at what I’ve learned. But at the practices I’ve developed through friction, repetition, and more than a few sessions that went sideways in instructive ways. At what it actually means, after three years of serious use, to develop AI fluency.

That’s what this book is: Nine breakthroughs, three interludes, one attempt to peel back the layers of my own AI literacy and see what’s underneath.

I use AI daily in my work: designing professional development, revising documents, doing research, thinking through problems I can’t quite see my way around yet. For a long time I assumed I was getting better at it because I was doing more with it. But what I’ve come to understand is that doing more doesn’t automatically produce understanding. But the breakthroughs did. The moments when something failed in a specific enough way that I could see, suddenly, both what the machine was doing and what I should have been doing differently. Those moments are what this book is built from.

Each of the nine chapters follows the same structure: a breakthrough, a real mind-opening moment from an real AI session, and then two things that moment revealed simultaneously. How the machine works. How to work with it. I’ve come to think of these as two strands of the same structure, inseparable, each one making the other legible. A double helix. You can’t develop good practice without understanding the architecture beneath it. And the architecture isn’t meaningful without the lived experience that gives as stakes and context.

The book moves in three parts. The first is about managing a conversation: learning to interrupt, recognizing when a session has run its course, doing the setup work that determines whether a conversation succeeds before it begins. The second is about reading AI output critically: catching flattery, making corrections that actually land, recognizing when confident-sounding data has been invented wholesale. The third is about what persists. Three years of working with a system that remembers nothing has, it turns out, helped me understand my own thinking better than I could have anticipated.

Threaded through the three parts are three interludes. I want to flag these directly, because I think they’re some of the most important writing I’ve done on this subject. In them, I step back from the individual breakthroughs and try to name what the practices add up to: a new kind of AI literacy, one rooted not in technical fluency or prompt tricks but in the communicative and rhetorical skills that serious practitioners have always needed. The interludes are where I make the larger argument. The chapters are where I show the work.

I’m releasing the book in four installments here on Substack. If you’ve been using AI long enough to sense that something is missing from the guidance available to you, that the technical explainers describe the system without telling you what to do about it, and the prompt guides tell you what to do without explaining why it works, this is my attempt to close that gap.

It starts with the simplest thing I learned, and the one that changed everything that followed: you have to interrupt.

Nick Potkalitsky, Ph.D.

Table of Contents:

Introduction: Before the Breakthroughs

Part one — Managing the conversation

1. The Interrupt Active monitoring · No metacognition

2. The Entropy Recognition Conversation budgeting · Context window saturation

3. The Onboarding Architectural front-loading · Training distribution as default

4. The Purpose Check Intentional checkpointing · Generation without intention

Interlude I: The Shape of Attention

Part two — Reading the output

5. The Sycophancy Detection Adversarial self-framing · RLHF and agreement bias

6. The Precision Correction Diagnostic editing · Constraint satisfaction

7. The Fabrication Instinct Epistemic vigilance · Prediction without reference

Interlude II: The Critical Reader

Part three — What stays

8. The Process Externalization Expertise encoding · In-context imitation

9. The Relationship Reset Portable knowledge architecture · Statelessness

Interlude III: What Stays

Conclusion: The Double Helix

Read more

Before getting into this week's piece, I want to take a moment to thank each of you for your continued engagement with this newsletter. We have now crossed 12,000 subscribers, with nearly 200 paid supporters, and readers located across all 50 states and 162 countries, a fact that still genuinely stops me every time I look at it.

Subscribe now

I am excited to share that the first two chapters of my book, Talking to Machines, will be published here next Thursday. It is not a prompt engineering guide or a tips-and-tricks manual, but a close account of the human breakthroughs that transform how we actually think about and interact with AI, drawn from years of intensive, documented use of these tools in my own work as a writer, researcher, and educator. If the ideas in this newsletter have resonated with you, I think the book will feel like a natural extension of that conversation, and I hope you'll be there for it.

Every AI edtech company pitching districts right now has the same slide in their deck: a sleek admin console showing usage analytics, content moderation controls, and real-time visibility into what students and teachers are doing with AI. It’s the feature that gets procurement officers to nod and school board members to exhale.

The problem is that a dashboard is not a system. And most of the admin consoles being demo’d to districts right now are doing something more modest than they appear to be doing. They are providing visibility into activity that is, underneath the surface, fundamentally disconnected. The ability to see what’s happening doesn’t change the nature of what’s happening. And what’s happening in most AI edtech deployments is not coherent curriculum. It is a collection of individually driven, teacher-initiated interactions: a close reading exercise here, a writing scaffold there, a comprehension check that one teacher set up and another would never think to use. Useful, sometimes genuinely effective, but not adding up to anything you could call a curricular design.

That distinction matters more than any feature comparison. So let’s look at where the major platforms are, what they’re actually delivering, and then at what a different category of tool makes possible.

SchoolAI and MagicSchool: Real Visibility Into Fragmented Activity

The most honest thing you can say about SchoolAI and MagicSchool is that they have done something genuinely hard: they have layered real administrative oversight onto platforms that teachers actually use. That is worth acknowledging before identifying where the model hits its ceiling.

SchoolAI’s Mission Control is arguably the most mature real-time classroom monitoring tool in this space. Teachers get a live dashboard showing every student’s activity inside AI “Spaces,” including sentiment analysis, automatic groupings by engagement level, and flagging of concerning behavior. You can pause or end individual student sessions, download CSV reports, and get AI-synthesized summaries of each participant’s conversation. You can read every message a student sends to the AI and see how the AI responded. At the district level, data rolls up to administrators through consolidated reports, and the platform is FERPA and COPPA compliant, 1EdTech certified, and SOC 2 certified. The oversight is genuine. The classroom layer, in particular, is sophisticated in ways that matter for student safety and teacher situational awareness.

MagicSchool has built the most feature-complete enterprise admin console of the four platforms reviewed here, though it sits behind a paywall that most of the platform’s users will never reach. Enterprise customers can upload district documents to ground AI responses in approved content using retrieval-augmented generation, create and distribute custom tools across teacher dashboards, show or hide specific tools based on district priorities, set up content moderation with real-time alerts for high-risk student content, and track adoption and usage trends with data suitable for board-level reporting. SSO integration is available through Clever, ClassLink, Google, and Microsoft, along with LMS connectivity through Canvas and Schoology. The platform reports over six million users. The majority of those educators are on free or lower-tier plans that give their administrators no visibility at all.

What both platforms share, and what neither fully resolves, is that the activity being monitored is structurally individual. Teachers open AI spaces for discrete purposes. They build tools for their classrooms, for their students, for the moment in front of them. The admin console can aggregate those interactions, flag concerning language, and tell you how many teachers logged in last week. What it cannot tell you is whether any of it coheres. Because it doesn’t. The curriculum layer doesn’t exist. What exists is a set of individually authored interactions that happen to run through the same platform, and visibility into that activity, however sophisticated, doesn’t change its fundamental character. You are seeing a great deal of fragmented use clearly. That is a meaningful capability. It is not the same as curricular coherence.

This is a hallmark of what we might call first-responder AI edtech: tools built to get AI into classrooms quickly, to provide teachers with useful instruments, and to give administrators enough visibility to satisfy compliance and governance concerns. They solved the access problem. They have not solved the architecture problem.

Brisk Teaching: Workflow Integration With Similar Limits

Brisk occupies a slightly different position in this landscape. Rather than asking teachers to open a new platform, it lives inside Google Docs, Slides, and Microsoft tools as a Chrome extension, which is the source of its genuine adoption advantage. The platform is now used by over two million educators across more than 20,000 districts. That reach is real.

The admin layer has grown to match. The Admin Tool Manager, launched in early 2026, lets district admins create custom tools using their own prompts, publish them to every teacher’s Brisk Library, and hide default tools that don’t align with district priorities. The Boost School Hours feature gives admins control over when students can access AI, toggling specific days and hours with automatic pausing outside approved windows. District plans include analytics dashboards and standards alignment tools. A forthcoming feature called Curriculum Intelligence, announced for back-to-school 2026, promises to embed adopted curriculum directly into the platform so that AI-generated content stays aligned with what districts actually teach. It hasn’t shipped yet.

The analytics that do exist are adoption-focused: who is using which tools, how often, in which schools. That information has real value for district leaders trying to understand deployment patterns and support needs. But it is still a picture of activity, not a picture of learning. And the underlying structure is the same as the other first-responder platforms. Teachers are using AI inside their existing workflows, individually, for purposes they have defined. The admin layer reflects that. It provides control and visibility over a distributed, disconnected set of uses rather than coherence across them.

Kira: Instructional Infrastructure

The first three platforms share a design philosophy: build AI tools for teachers, then add an administrative layer for oversight. Kira is doing something structurally different, and the distinction is not superficial.

Kira began as a computer science curriculum platform and has evolved into what is more accurately described as instructional infrastructure: an integrated environment in which curriculum, instruction, assessment, and analytics are not separate modules bolted together but components of a single coherent system. Districts can use it as a standalone LMS or integrate its capabilities through LTI and OneRoster, but the more important architectural fact is that what happens inside the platform feeds a single analytical engine. A student’s exchange with an AI tutor and that same student’s written response to a close reading prompt are both informing the same picture of what that student understands. The Atlas, Kira’s underlying learning model, gets smarter regardless of the source.

The admin analytics reflect this. They aren’t tracking AI usage. They’re tracking learning. The platform maps student mastery disaggregated from state standards, classifies students into MTSS intervention tiers, and surfaces actionable insight about where students are struggling and why. For teachers, the interface is conversational: a chat that can surface struggling students, recommend interventions, and generate personalized materials in the moment. For administrators, the platform offers real-time analytics on engagement and progress, role-based access controls, SIS integration for automated rostering, and granular control over AI features at the student and section level. Admins can enable or disable AI assistance, adjust reading levels, and customize the type of guidance students receive. Kira is free for teachers and students, with districts paying a bespoke rate for analytics and admin features. This is not a usage dashboard dressed up in learning language. It is an instructional intelligence layer, and it only exists because the curricular infrastructure underneath it is coherent.

That infrastructure also provides an immediate and architecturally sound response to some of the most pressing problems AI has created in classrooms. Assessment integrity is the clearest example. Since generative AI arrived in classrooms, institutional responses have been largely defensive: policies written fast, bans that quietly didn’t hold, detection tools locked in an arms race they were always going to lose. Kira’s enclosed environment isn’t a policy posture against AI. It is a different kind of assessment space, one where the conditions of learning are controlled not by prohibition but by design.

It is worth being clear-eyed about where Kira currently sits in the market. The platform is primarily positioned as a response to endemic educational problems that long preceded AI: the chronic failure to build high-quality instructional materials at the pace and scale schools need them, the persistent gaps in standards-aligned curriculum, the difficulty of getting actionable instructional data in front of teachers and administrators in a form they can use. These are real problems and Kira addresses them in ways that more established platforms do not. But the potentialities run further.

AI literacy is the territory where the platform’s current architecture points toward something genuinely new. A tool with Kira’s curriculum generation capacity and its enclosed analytical environment can do what nothing in the current instructional landscape does well: build AI literacy not as a stand-alone elective but as a question distributed across courses, through English classes that take seriously what writing means when AI writes, through history courses that wrestle with what evidence means when AI generates plausible evidence, through every subject where students are learning what it means to know something in a particular way. The infrastructure makes this possible. Whether and how quickly it gets built is a different question.

The Question Districts Should Actually Be Asking

The more interesting strategic question about platforms like Kira is not whether they outcompete first-responder tools on feature comparisons. It is how they fit, or how they come to outmode, the broader curricular commitments districts are already carrying.

U.S. districts continue to spend millions on high-quality instructional materials. Those investments are politically and institutionally durable. Any next-generation AI platform operating in this market has to reckon with that reality one way or another. There are two plausible trajectories. The first is hybrid integration: sophisticated tools that work alongside adopted HQIM, extending and personalizing and assessing without displacing the curriculum that districts have already committed to and trained teachers on. The second is displacement: platforms whose curriculum generation capabilities improve to the point where the case for separate HQIM investment weakens, where the coherence and adaptability of AI-generated curriculum starts to look more attractive than the rigidity of traditionally developed materials.

Which trajectory dominates will, in large part, determine the shape of the AI toolscape as it continues to develop. Districts evaluating platforms now are making bets on that question, whether they know it or not.

The admin console arms race is not irrelevant. Governance matters. Visibility matters. Student safety and data compliance matter. But the platforms that will define the next phase of AI in education are not the ones with the most toggles in their dashboards. They are the ones that have built something coherent underneath: tools designed not to sit alongside instruction but to be its architecture. The question worth asking isn’t which platform can show you what’s happening. It’s which one is actually building something worth seeing.

Nick Potkalitsky, Ph.D.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy

In March 2026, New York City Public Schools released its official Guidance on Artificial Intelligence. At 30-plus pages, it is the most comprehensive K-12 AI policy document I’ve seen from any major school system in the country. It is also, in important ways, a document at war with itself.

The guidance does several things that no other large district has done as clearly or as deliberately. It also reveals, in its own language, the limits of the organizing framework it chose. What follows is both an appreciation and a critique, because the stakes are too high for one without the other.

The Foundation Is Right

Start here, because this matters. NYCPS made a set of choices that most districts haven’t had the courage or the clarity to make.

Safety as the foundation, not the guardrail. The guidance begins with what AI will never be allowed to do, before it describes what AI can do. That ordering is not rhetorical. It shapes the entire decision-making architecture. Most district AI frameworks treat safety as a set of constraints layered on top of an adoption decision. NYCPS treats safety as the ground the adoption decision stands on.

Students never get a green light. The framework uses a Traffic Light Approach, red/yellow/green, to classify AI use by risk. Student use of AI for research, exploration, and creative projects lives in yellow: permitted only with educator guidance, critical evaluation of outputs, and age-appropriate context. There is no scenario in this framework where student interaction with AI is unconditionally approved. The green light is reserved exclusively for educator and leader activities: lesson planning, drafting communications, professional development, operational data. Adults get green. Students get yellow at best, red at worst.

Hard lines on high-stakes decisions. AI is flatly prohibited from making or driving decisions about student placement, discipline, promotion, graduation, grading, IEP development, counseling, crisis intervention, and surveillance. In a market where edtech vendors are already selling tools that claim to predict student behavior and automate special education workflows, NYCPS drew a line.

Institutional honesty. The guidance acknowledges that the long-term effects of AI on how children learn and develop are not fully understood. It names open questions it hasn’t resolved yet, from grade-band differences to cognitive offloading to academic integrity. And it commits to a public timeline: a comprehensive Playbook by June 2026, built with stakeholder input.

These are genuine accomplishments. NYCPS is governing AI more thoughtfully than any comparably sized school system. The question is whether the framework it chose can carry the weight of what it’s actually trying to do.

When Categories Meet Workflows

The Traffic Light Approach is a familiar device in K-12 technology governance. Districts across the country have used red/yellow/green frameworks to sort tools and platforms by user role: this app is approved for teachers, restricted for students, prohibited for everyone. It’s an access-control framework. It answers the question: who is allowed to use what?

NYCPS is asking the Traffic Light to do something much harder. It’s not just sorting tools by user. It’s trying to govern the nature of AI-human interaction itself: where AI can advise, where it can assist, where it must be absent. And that’s where the categories start to strain, because the actual workflows these designations are trying to regulate don’t respect clean boundaries.

Consider the red-light designation on grading. The guidance states that “the educator of record determines what a student knows” and that “AI-generated data is advisory only.” The first sentence draws a hard line. The second opens a door. “Advisory only” doesn’t mean AI is absent from assessment. It means AI can produce data, surface patterns, generate scores, as long as a human makes the final call.

In principle, that’s a reasonable distinction. In practice, the boundary between a data point and an automated decision is often paper thin. When an adaptive learning platform generates a readiness score, when a tool flags a pattern in student performance, when a dashboard surfaces a recommendation that a teacher sees thirty seconds before entering a grade, is the AI advising or is it deciding? The answer depends not on the policy but on the workflow: the screen layout, the time pressure, the number of students, the degree to which the teacher has independent evidence to weigh against the AI’s output. The Traffic Light says red. The lived experience of a teacher at 3:45 on a Friday may say something closer to yellow, or even green.

The same tension runs through IEP development. The red-light language is unambiguous: “All special education documents are developed by qualified professionals.” But AI drafting tools are already entering special education workflows. The distance between “a professional wrote this IEP” and “AI drafted this IEP and a professional reviewed it” is enormous in principle and vanishing in practice. The guidance prohibits the destination but doesn’t yet address the journey, and the journey is where the slippage happens.

Or take the green-light category. Brainstorming and lesson planning are approved for educators with confidence. But anyone who has spent time working with generative AI knows that brainstorming bleeds into drafting, drafting bleeds into reliance, and reliance can quietly become dependency. The green light assumes a stable boundary between “AI supports my thinking” and “AI is doing my thinking.” That boundary exists. It is also, in the daily rhythm of professional work, extraordinarily easy to cross without noticing.

The Traffic Light works beautifully as an access framework. The trouble is that governing AI in education requires something more: a process framework that accounts for how AI actually embeds itself in human workflows, gradually, and often invisibly.

The Document That Needs to Be Two Documents

Hovering beneath these categorical tensions is a deeper structural question: what kind of document is this?

The Traffic Light Approach reads like policy. It draws hard lines. It prohibits. It designates. It invokes federal and state law. When the guidance says AI will never be used for student surveillance, that is a policy statement, and a strong one.

But much of what the guidance needs to govern can’t be legislated with a red/yellow/green designation. It requires the kind of granular, context-specific operational guidance that tells an educator: here is where the line is in this workflow, with this tool, at this grade level, for this purpose. Policy can say “AI-generated data is advisory only.” Guidance needs to say what that looks like when a teacher is sitting in front of an AI-powered assessment dashboard at the end of a marking period.

The document seems to know this. It repeatedly flags areas where further specificity is needed and points toward the June 2026 Playbook as the place where that specificity will arrive. Homework and academic integrity, grade-band differentiation, the distinction between managed school tools and personal accounts, AI detection tools, biometric data: all flagged, all deferred.

That deferral is honest, and it is also a risk. The guidance is live now. Teachers and administrators are reading it now. And the Traffic Light’s clean categories may already be creating a false sense of clarity in areas where the actual decisions are far more ambiguous than red, yellow, or green.

Consider AI detection tools, which the guidance doesn’t mention by name. Products like Turnitin’s AI detector and GPTZero are functioning in schools across the country as automated assessments of whether a student did their own work. If the red-light logic holds, and AI cannot be trusted to determine what a student knows, then it follows that AI cannot be trusted to determine whether a student cheated. These tools are unreliable, they disproportionately flag multilingual learners, and they place a machine’s probabilistic guess at the center of decisions that can alter a student’s academic record. The framework’s own principles point toward a clear answer here. But because the guidance defers academic integrity to the Playbook, that answer doesn’t yet exist in policy.

The Playbook Matters More Than the Framework

None of this diminishes what NYCPS has accomplished. Governing AI at this scale, for nearly one million students, with this level of public transparency and stakeholder engagement, is a genuine achievement. The Traffic Light Approach is the best organizing framework any major district has published. And the willingness to build in public, with named timelines and open feedback periods, sets a standard that other districts should follow.

But the NYCPS experience also reveals something important for every district watching from the outside. The access-control model, who can use what, is necessary but not sufficient. The harder problem is workflow governance: how AI interacts with human judgment inside the daily realities of teaching, grading, writing IEPs, and making decisions about students. That problem doesn’t sort neatly into three colors.

If you’re building your own district’s AI framework, the Traffic Light is a strong starting point. Adopt the principle of safety first. Draw your red lines. Reserve the green light for adults. Put student use in yellow, permanently. Those structural choices are right.

But don’t stop there. The Playbook matters more than the framework, because the Playbook is where you address what actually happens when a teacher sits down with an AI tool and the boundary between a data point and a decision starts to dissolve. NYCPS has committed to building that. The rest of us should be watching closely to see if they deliver.

Right Architecture, Rooms Not Yet Built

The NYCPS Guidance on Artificial Intelligence is the most thoughtful K-12 AI policy document in the country. Its commitment to safety as a foundation, its refusal to give students a green light, its institutional honesty about what it doesn’t yet know: these are models for the field.

But the Traffic Light Approach is an organizing device borrowed from a simpler problem. It was built to sort tools by access level. NYCPS is asking it to govern the far more complex question of how AI and human judgment interact inside educational workflows. In some places, the framework holds. In others, the categories are already straining under the weight of what they’re being asked to carry.

The architecture is right. The next step is building the rooms inside it. June 2026, and the promised Playbook, will tell us whether NYCPS can match the ambition of its framework with the operational specificity that students, educators, and families actually need.

Nick Potkalitsky, Ph.D.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy

The response to the Thinking with AI Student Workbook (e-book in progress!!!) and the accompanying training sessions has been genuinely humbling. Thank you. This week I am releasing the Parent and Guardian Workshop, which brings that seven-part paid series to a close. If you have been following along, this final installment completes the arc. If you are coming to it fresh, the full series is available in the archive.

And then, in April: Talking with Machines: The Subtle Art of Working with AI. The premise is simple and I think radical — that the breakthrough moments in how we interact with LLMs are simultaneously the best diagnostic tools we have for understanding what these machines actually are. AI literacy, at its deepest, is not a framework you learn. It is something you accumulate through close, honest attention to what happens in the room between a human and a machine. More soon.

Subscribe now

She already knew something was wrong before she handed me the essay.

She is an eighth grader in a mid-sized Ohio public school — driven, intellectually curious, the kind of student who pushes back on an argument not because she is obstinate but because she is actually thinking. She writes in an elevated vocabulary. She has a voice, and she knows it. She came to our tutoring session having already run her argumentative essay through an AI-powered feedback tool her teacher had set up on a major ed-tech platform. The tool had scored it 100%.

She wanted me to look anyway.

What I found was a competent essay, well-scaffolded, formulaic in the way that state-testing environments reward, and built substantially on the sentence frames her teacher had provided in class. She told me, matter-of-factly, that she had noticed over time that her scores improved the more directly she incorporated her teacher’s language. So she had. The AI agreed. One hundred percent.

The Formula and What It Leaves Out

But there was slippage in the counterargument paragraph. The assignment, like so many argumentative essays assigned to students over the past fifteen years, asked her to weigh in on the impacts of social media. She had not chosen her side; the pro-social-media-as-business-tool position had been assigned to her. In the counterargument, the some say sentence raised the superiority of personal websites for promotion and advertising. The others respond sentence drifted back to a point already made in body paragraph one about the potential lower cost of social media, rather than engaging directly with what the some say had actually argued.

We spent twenty minutes on it. We talked through the real tension: the personalization of a website versus the algorithmic reach of social media. We noticed together that the formula had no room for a blended approach. The either/or logic was baked into the instruction, and likely into the state rubric beneath it. We revised the counterclaim. We returned to a quoted statistic that 50% of companies use social media to promote their brands, and I asked the pointed question: so what about the other 50%? What followed was a genuine conversation about how to word a justification carefully, how to make a number feel significant without overstating it, how to argue honestly inside a constrained form.

It was, in other words, exactly the kind of instruction writing teachers hope to provide.

When we finished, she ran the revised essay through the AI tool again. It scored her an 80%. The feedback was general to the point of near-uselessness. The tool attempted to quote her essay to identify specific problems, but the quoted passages did not accurately reflect her actual text. She ran it again, same essay, new chat window. This time it scored a 90%.

She looked at the screen. She looked at me.

What an Inconsistent Score Actually Costs

This is the part worth pausing on, because what happened next is not simply a story about a flawed tool. It is a story about what flawed tools do to the ecosystems around them.

AI feedback tools in educational settings exist on a wide spectrum of reliability. At the more rigorous end, automated scoring systems used by state assessments are normed against human-scored samples, calibrated carefully, and tested for consistency before deployment. What this student encountered was something different: a general-purpose AI embedded in a classroom platform, handed a rubric, and asked to approximate summative evaluation. The distinction matters enormously, because students and teachers cannot always see it. The tool presents with the same confidence regardless of its actual precision. It scores. It comments. It feels authoritative.

When it contradicts itself across two runs of the same essay, that authority does not quietly recede. It explodes outward. And what it takes with it is not just the student’s confidence in the tool. In that moment, sitting with my tutee, I watched it erode her confidence in her teacher’s approach, in the state testing apparatus the formula was built to serve, in the twenty minutes of genuine intellectual work we had just done together, and, most quietly and most damagingly, in her own judgment as a writer.

Why did we spend all that time, her silence was asking, if the result is a lower grade?

I have taught in public school classrooms under similar conditions. I know what it is to pitch instruction toward a struggling cohort while a student like her sits in the back, capable of far more. I understand why a teacher reaches for a tool that promises to lighten the feedback load. I am not here to assign blame. I am here to say that an insufficiently optimized AI tool, deployed in a high-stakes evaluative role, does not simply fail to help. It actively destabilizes the feedback ecosystem that learning depends on. The connective tissue between effort, quality, and outcome, which has always been imperfect, now has a new and visible fault line running through it.

There is a version of this story that ends with a call to action: we need better tools, better teacher training, a clearer distinction between formative feedback and summative evaluation. All of that is true. All of that is necessary. But those conditions are a horizon, not a resolution, and I want to be honest about the distance between where we are and where we need to be.

Resilience, or Something We Are Mistaking for It

Because here is what actually happened at the end of our session.

She thought about it for a moment. She restored the original essay, the one that had scored 100%, and she submitted it. Then she closed her laptop and moved on.

I have been thinking about that moment ever since. There is something in it that reads like maturity. She assessed the situation clearly, made a pragmatic decision, declined to be derailed by an unreliable signal. That is a genuine cognitive competency, and one that many students never develop.

But I am not sure I am allowed to call it maturity.

Because what she also did, in that same gesture, was disengage from the process entirely. She did not advocate. She did not push back. She looked at a system that had failed her and decided, reasonably, that the most efficient response was to stop trusting it and comply with its most forgiving version. She is twelve years old, and she has already learned that lesson.

So many students never develop the resilience to separate their sense of self from the inconsistency of institutional evaluation. They internalize the grade. They disengage. School becomes a game you are winning or losing, and the game has no reliable referee. What worries me is not just the students who will be broken by that. What worries me is the students who will handle it exactly the way she did.

I want to call what she did that afternoon maturity.

I am not sure I am allowed to.

Nick Potkalitsky, Ph.D.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy

I am excited to bring this guest post to my readers today. Marvin Starominski-Uehara’s work is a joyous, interdisciplinary synergy of concepts and approaches. I hope you enjoy engaging with this introductory exploration of his new framework.

Also be sure to check out my serial releases from my latest book, Thinking with AI: A Student’s Guide to Literacy in an AI-Rich World.

Intro and Ch. 1 Ch. 2-3 Ch. 4-5 Ch. 6-7

On Thursday, I plan on release my first training roadmap I hope to release some training roadmaps for educators and administrators interested building Thinking with AI pilots in their schools.

Nick

Subscribe now

There is a moment I return to often when I think about why education fails students in the age of AI. It is not the moment a student submits a generated essay and it is not the panicked faculty meeting about plagiarism policy. It is a quieter moment: a student sitting alone in front of a glowing screen, AI chat open on one side, a half-finished assignment on the other, with no idea what to do next. Not because they lack ability but because no one taught them how to think with the uncertainty. They have been trained to find the answer: the AI gives them answers -- game over.

I am Marvin Starominski-Uehara. I teach university courses and I research something called Stigmergy Network Theory -- a framework for understanding how autonomous agents can coordinate complex, intelligent behavior without ever directly communicating. My model is not Silicon Valley: it is ant hills. And I have spent the last several years asking one question: What if classrooms worked the same way?

The Problem

AI has not made the traditional classroom more efficient -- it has exposed its central contradiction. We built education around the transmission of information. AI transmits information faster, cheaper and with fewer complaints. The teacher who stands at the front and delivers content is now competing with a tool that never sleeps and never gets tired. But here is what AI cannot do: it cannot leave an original and highly personal trace that guides the next curious mind forward. It cannot build the scaffolding founded on a solid and organic community of practice that persists long after the conversation window closes. This is precisely where Trace Pedagogy enters.

My Solution

In my research on stigmergy -- a concept borrowed from entomology and applied to digital environments -- individuals do not need to talk to each other to cooperate at a high level. Ants lay pheromone trails. Wikipedia editors leave revision histories. Students, when properly structured, leave intellectual traces -- comments, annotations, video reflections, discussion posts -- that become the curriculum itself for the students who follow. My approach teaches students to navigate information ecosystems, read digital traces and contribute knowledge that benefits future learners, modeling how successful systems coordinate through accumulated wisdom. And the educational implications emerging from this model are radical. When students generate knowledge before class -- when they share what they have found, what confused them, what surprised them -- they do not just prepare for a lesson: they become the lesson. The teacher’s role shifts from transmitter to architect: designing the environments in which those traces accumulate into something larger than any single student could produce. This is what I mean when I talk about amplifying collective intelligence.

Flipping the Classroom

Most educators have heard of the flipped classroom -- students watch lectures at home and ‘do the work’ in class. In my Flipped Classroom approach, I combine it with Minimally Invasive Education -- Self-Organizing Learning Environments -- allowing students to coordinate and learn through the digital traces they leave online. Here is what that looks like in practice when I give students an AI-powered dilemma and a scaffold:

‘A logistics company uses an AI model to optimize delivery routes. The AI consistently routes drivers through a lower-income neighborhood at night, reducing delivery times by 4%. The CEO says it is just math. The city council disagrees. Use at least two AI tools to research this from different stakeholder perspectives. Post your analysis, including where your AI sources agreed, disagreed or gave you something unexpected’

Students arrive having already wrestled with the mess. They have already discovered that different AI tools give different answers. They have already noticed that ‘efficiency’ is not a neutral term. The classroom becomes the space where those individual traces converge -- where the confusion sharpens into something productive. An online poll I ran across two courses last semester revealed a striking polarization: one group of students overwhelmingly embraced generative AI, with 80% excited about its learning potential. However, another group expressed deep anxieties about cognitive development and fairness. Both groups were right. The classroom became the place where those opposing traces had to meet.

Throwing Students Into Messy AI Dilemmas

The best AI dilemmas share three features: (i) they have no clean answer; (ii) they require students to interrogate the AI’s reasoning; and (iii) they force a trade-off between values that students actually care about.

Example 1: A mid-sized company uses AI to screen résumés. The tool was trained on ten years of successful-hire data -- from a time when the company hired almost exclusively men. The algorithm gives higher scores to male-coded language on CVs. HR says the AI is just ‘reflecting historical patterns’. Diversity advocates say the AI is reproducing injustice at scale. Students must assess: Is the problem the AI, the data, the company or the humans who chose to use it? How would you redesign the system? What would you sacrifice?

Example 2: A company has a compelling sustainability mission statement. Their AI-generated annual report scores highly on ESG metrics. But their supply chain data -- which the AI did not include in its analysis -- tells a different story. This kind of scenario illustrates both the power of digital traces to align business actions with mission statements and the danger of opaque algorithmic decision-making. Students must decide: Do they trust the report or chase the data trail?

Example 3: A school district deploys an AI tutoring system that dramatically improves standardized test scores in wealthy neighborhoods. In under-resourced schools, teachers lack the time and training to implement the tool effectively, widening the achievement gap. Is this a technology problem or an equity problem? Can it be both? What does a solution that takes both seriously actually look like?

In each case, students use AI as a research scaffold -- they are required to prompt multiple tools, compare outputs and document where the AI helped and where it led them astray. The process of using AI critically is the assignment.

Building Collective Intelligence

My courses are designed to foster connections among students -- teaching them to use AI tools for research and project creation, while combating the isolation that comes from treating learning as a solitary, screen-based activity. When pre-class contributions are visible to the whole cohort -- not just graded by the teacher -- something interesting happens. Students start building on each other’s traces. One student’s half-formed question becomes another student’s breakthrough. A video a student made about gender equality references data another student surfaced in a discussion post, which then becomes the foundation for a collaborative policy brief. This is stigmergy in action. No central coordinator, no single authority, just traces building on traces until the group has produced something none of them could have generated alone. Students build communities of practice where sharing increases learning for everyone. This is not idealism -- it is a design principle. And it transfers directly to the AI literacy skills students need most: (i) the ability to evaluate sources; (ii) synthesize competing information; and (iii) communicate complex ideas to a real audience.

Trace Pedagogy Lesson Plan: General Structure

1. Pre-class assignment (individual, shared publicly with cohort): Students receive a real-world AI dilemma. They use at least two AI tools to research stakeholder perspectives. They document: (i) where the AI tools agreed, (ii) where they diverged, (iii) one assumption embedded in each AI’s response, (iv) their own initial position and why. Posts are visible to the entire class before the session.

2. In-class Phase 1 -- Trace Mapping (15 min): Small groups read each other’s pre-class posts and identify the most generative tension: the point where two students’ traces pull in genuinely different directions. Groups name the tension without resolving it.

3. In-class Phase 2 -- Structured Disagreement (20 min): Each group presents their tension to the class. The instructor’s role is not to resolve it but to deepen it: offering a complicating piece of data, a stakeholder perspective no one raised or a question that forces the group to reconsider their framing.

4. In-class Phase 3 -- Collaborative Redesign (20 min): Groups draft a brief proposal: if you were advising the decision-maker at the center of this dilemma, what would you recommend and what would you sacrifice? Proposals must acknowledge trade-offs explicitly.

5. Post-class trace (individual): Each student updates their original pre-class post with a reflection: How did the class conversation shift your thinking? What trace do you want to leave for next week’s cohort?

6. Assessment logic: Students are evaluated not on correctness but on the quality of their trace: how much it contributed to the collective understanding of the cohort. Depth, honesty and intellectual generosity matter more than having the right answer.

Final Remarks

The tools will keep changing. The models will keep improving. The dilemmas will keep getting messier. What will not change is the fundamental challenge of teaching humans to think well together -- to leave traces that amplify rather than distort, to coordinate without surrendering their own judgment, to use powerful tools without being used by them. That is the work I do in my courses and it is the work I explore in my online AI lessons plans. If this approach to AI pedagogy resonates -- if you are drawn to the intersection of collective intelligence, digital traces and classroom design -- let’s continue this conversation. After all, the traces you leave matter more than you know.

Marvin Starominski-Uehara is an Adjunct Assistant Professor at Temple University Japan and the creator of Stigmergy Network Theory. He teaches courses on international business, environmental studies and race & diversity and researches how digital platforms enable social change through mediated communication and coordination. Follow his work at marvinuehara.com and marvinstarominskiuehara.substack.com.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy

Educating AI publishes twice weekly. If this piece was useful, share it with a colleague who’s navigating these questions. And if you want to support the work, consider becoming a paid subscriber. It’s what keeps this going.

Check out my serial releases from my latest book, Thinking with AI: A Student’s Guide to Literacy in an AI-Rich World.

Intro and Ch. 1 Ch. 2-3 Ch. 4-5 Ch. 6-7

Subscribe now

Faculty across the country keep telling me the same story. Students paste assigned readings into ChatGPT and show up to class with a summary. They can talk about the article in broad strokes. They can hit the main points. But when you push them on a passage, when you ask them to work through a specific paragraph or evaluate the logic of an argument, the floor drops out. A religious studies professor told me recently that when he asks students to respond to passages in class that weren’t assigned at home, many struggle to arrive at any real takeaways, even at the level of literal comprehension.

The instinct is to call this laziness. To frame it as a motivation problem, or a discipline problem, or a technology-ruined-their-brains problem. I don’t think that framing is entirely wrong, but I think it’s incomplete, and the part it’s missing may be more important and more uncomfortable than a story about lazy students.

Here’s what I think we’re not talking about enough: for a meaningful number of the students running their readings through AI, the issue isn’t that they won’t engage with the text. It’s that they can’t, at least not at the level the assignment demands. And for those students, turning to a tool that gives them access to ideas they would otherwise be locked out of is not an act of avoidance. It’s a rational response to a real problem. We need more research to understand how large this population is and how cleanly it maps onto the students faculty are describing. But the literacy data we do have suggests the overlap is substantial.

The Pyramid They Never Climbed

Literacy researchers have long understood that reading develops in layers. The most useful model I’ve encountered comes from Tim and Cynthia Shanahan, who describe it as a pyramid. At the base is basic literacy: decoding, print recognition, high-frequency words. Most students consolidate this in the early elementary grades. In the middle is intermediate literacy: general comprehension strategies, academic vocabulary, the ability to handle multisyllabic words fluently, recognition of how different kinds of texts are structured, and the metacognitive awareness to notice when you’ve stopped understanding. At the top is disciplinary literacy, the specialized ways that historians, scientists, mathematicians, and literary scholars read, write, and reason within their fields.

The pyramid narrows as it rises because each layer depends on the one below it. You cannot evaluate a primary source for perspective and bias if you cannot comprehend the sentences in it. You cannot coordinate theory and evidence in a science passage if unfamiliar vocabulary is consuming all of your working memory. The advanced intellectual work that disciplinary reading requires is cognitively demanding. It presupposes that the more basic operations have become automatic enough to run in the background, freeing your mind for the thinking the text actually demands.

Hollis Scarborough offered a complementary way of seeing this. She described skilled reading as a rope woven from multiple strands: phonological awareness, decoding, and sight recognition on one side; background knowledge, vocabulary, language structures, verbal reasoning, and literacy knowledge on the other. As a reader develops, the lower strands become automatic and the upper strands become strategic, and the whole thing weaves together into fluent comprehension. But if even one strand is weak, the rope frays. A student might decode fluently but lack the vocabulary to make sense of academic prose. She might reason well but have so little background knowledge that a science passage reads like a foreign language. The strands are interdependent, and weakness in any one of them loads extra weight onto all the others.

The Scale of the Problem

Here’s where the data gets difficult to sit with. The 2024 National Assessment of Educational Progress found that roughly one third of American eighth graders scored below the basic level in reading. Below basic. They could not reliably identify the main idea of a grade-level passage.

These are students about to enter high school, where every content-area teacher will expect them to work with complex texts. Jeanne Chall’s developmental research helps explain how they got there. There is a critical shift that happens around fourth grade, when students move from learning to read to reading to learn. Texts start containing vocabulary, concepts, and structures that go beyond everyday language. In longitudinal work tracking low-income students from second through seventh grade, Chall and her colleagues found that the earliest divergence at this transition was in vocabulary. Students who had seemed to read adequately in the early grades began falling behind as the words got harder and the ideas got denser. By middle school, many were performing well below grade level. The gap opened at word meaning and then spread outward into comprehension and general literacy performance.

This is the population that arrives in secondary classrooms and, increasingly, in college classrooms. Students who never fully consolidated that middle layer of the pyramid. Students for whom academic text is not a vehicle for learning but an obstacle to it.

Cognitive Overload, Not Cognitive Laziness

Cognitive load theory can explain what happens next. Working memory has a limited capacity. When the demands of processing a text exceed that capacity, learning breaks down. For a student whose intermediate literacy skills aren’t consolidated, reading an academic text is a cognitive overload event. The text isn’t too hard because the ideas are complex. It’s too hard because the student is spending all of her processing capacity on operations that should be automatic: decoding unfamiliar words, parsing sentence structures, retrieving word meanings. There’s nothing left over for the thinking the assignment actually requires.

This is the part that complicates how I see the AI question.

Not every student who pastes a reading into ChatGPT is doing it because the text is inaccessible. Some are making a convenience choice. Some are doing both at once, struggling with the text and also not particularly motivated to push through the struggle. But for students whose intermediate literacy skills genuinely aren’t consolidated, AI functions like a wheelchair for someone who cannot walk. It gets you where you need to go. The question is whether we want reading to be something students do or something students outsource, and whether we can even tell the difference between students who need the scaffold and students who are skipping a step they could take on their own.

The Differential Effect

Recent research suggests that the answer depends on who the student is. A study published in Frontiers in Education tested several AI-based tools, including summaries, outlines, a Q&A tutor, and a Socratic chatbot, on passage comprehension. The results split sharply by baseline ability. For lower-performing readers, the tools improved comprehension. For higher-performing readers, they worsened it, with the summary tool producing the largest decline among strong readers.

One way to read this: AI scaffolds access when students genuinely struggle with the text, but for capable readers, especially when they lean on summaries, it may interrupt the close, effortful processing through which deeper understanding develops. This is a single study, and the field needs much more work before we can draw firm conclusions. But the direction of the finding matters.

The Feedback Loop

Here is where I start to worry.

The percentage of twelfth graders who read a book or magazine daily dropped from 60 percent in the late 1970s to 16 percent by 2016. The share who reported reading zero books for pleasure in the previous year nearly tripled over the same period. A 2025 analysis of time-use data found that daily reading for personal interest has continued its decline, with the percentage of people who read on an average day falling from 28 percent in 2004 to 16 percent in 2023, and with the steepest drops among populations that already had the lowest reading rates.

Less reading means less practice. Less practice means thinner ropes. Thinner ropes mean more students for whom academic text is inaccessible. More inaccessible text means more students turning to AI. And more reliance on AI means even less practice.

The loop is real, and it is accelerating.

What This Means for Teaching

None of this changes the fact that disciplinary reasoning matters. Students still need to learn to think like historians, scientists, mathematicians, and literary critics. Teachers still need to make the epistemic standards of their fields explicit.

But it complicates how that reasoning gets taught, because many students will arrive at the threshold of disciplinary work without the literacy infrastructure to cross it independently. Researchers who have studied this challenge in practice, working with high school teachers who wanted their students to read like historians but found many couldn’t access the documents at all, arrived at a conclusion I think is exactly right: effective instruction requires a complex layering of intermediate and disciplinary literacy work, not a sequential progression from one to the other. You cannot wait until students have mastered general reading comprehension to begin teaching them disciplinary thinking. You have to do both at once.

What students need is not a reduced version of the task. They need instruction that addresses the intermediate skills they are still building and the disciplinary judgment they are beginning to develop, and that understands how AI can serve one goal while undermining the other.

That’s the ground I’ll be working in from here.

Nick Potkalitsky, Ph.D.

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

Stephen Fitzpatrick’s Teaching in the Age of AI: Essential reflections from a veteran high school educator on the challenges and opportunities of generative AI in the classroom!!!

This is it. The final release from Thinking With AI: A Student’s Guide to Literacy in an AI-Rich World.

I want to take a moment to thank this community. Your engagement with these chapters over the past several weeks has been extraordinary. The comments, the shares, the DMs from teachers already piloting these materials with their students. You showed me that this work matters, and you made it better along the way.

Chapter 6 (Writing) takes on the literacy domain where AI anxiety runs highest. Writing is where the stakes feel most personal, where the fear of replacement cuts deepest. This chapter shows students how to use AI across five roles while protecting what matters most: their thinking, their voice, their authorship. It covers analyzing AI-generated text to understand what real writing requires, verifying claims and sources before trusting them, developing ideas through dialogue without losing ownership, refining drafts with AI feedback they critically evaluate, and designing writing systems only when their expertise warrants it.

Chapter 7 (How Expertise Develops Across the Spectrum) pulls everything together. It shows how the five roles aren’t just different ways of working with AI but stages in developing genuine disciplinary expertise. It addresses why the sequence matters, how intentional AI engagement accelerates learning, how disciplines differ in which roles they emphasize, and how to maintain intellectual ownership through all of it.

With this release, the complete student guide is now available to paid subscribers.

What comes next: In March, I’ll release training manuals for piloting disciplinary role-based AI integration in your district. Then in April or May, I’ll release a companion teacher manual and workbook that complements the student handbook, giving educators the tools to bring this framework into their existing courses.

Thank you for believing in this work. Thank you for investing in it. And thank you for putting it into practice. This community is building something that matters.

Subscribe now

Chapter 6: Writing

Chapter Framing: Writing forces you to make ideas explicit, test reasoning, and clarify meaning. This chapter shows how different ways of engaging with AI—from analyzing AI-enerated text to orchestrating writing systems—can strengthen your capacity as a writer while maintaining authorship.

6.1 Critical Distance: Writing Is Thinking Made Visible (The Critic Role)

When AI-Generated Writing Is an Object to Examine

Writing is not typing. Writing is not producing text. Writing is thinking made visible through language.

When you write, you make choices about what matters, how to explain it, what examples clarify meaning, what order makes sense, which words capture your intention. Each choice reveals and refines your thinking. The difficulty of writing—the struggle to articulate what you mean—is where learning happens.

AI-generated writing looks like this process happened. It produces fluent text with coherent structure and polished sentences. But the thinking that makes writing valuable never occurred. AI assembled patterns from training data. No ideas developed. No meaning was clarified. No choices were made.

Understanding this distinction is foundational. Before you can use AI productively for writing, you need to recognize what writing actually requires—and what AI cannot do.

How AI Writes Without Intention or Voice

Read this AI-generated paragraph:

“Social media has both positive and negative impacts on society. On one hand, it connects people across distances and facilitates communication. On the other hand, it can lead to decreased face-to-face interaction and privacy concerns. Ultimately, social media’s effects depend on how individuals choose to use these platforms.”

Notice what happened. The text is:

• Grammatically correct

• Logically organized

• Balanced in perspective

• Completely empty

Nothing in this paragraph reveals thinking. No position emerged. No evidence was weighed. No interpretation was offered. The text presents the appearance of having considered the issue while actually saying nothing.

This is what AI writing typically produces: surface-level engagement that sounds thoughtful without containing thought.

Compare to a human writer working with the same topic:

“When my grandmother texts me photos from her morning walk, social media is connecting us across the thousand miles between her Florida apartment and my college dorm. When I realize I’ve spent three hours scrolling instead of calling her, that same technology has replaced the connection it promised to facilitate.”

Notice the difference. The human writer:

• Made a claim (social media replaces what it promises)

• Selected specific evidence (grandmother, texting vs. calling)

• Developed meaning through contrast (promised connection vs. actual isolation)

• Took a position worth discussing

The AI paragraph could appear in any generic essay. The human paragraph belongs to a specific writer thinking through a specific tension. This is what authorship looks like.

Read more

Thank you. The response to Thinking With AI: A Student’s Guide to Literacy in an AI-Rich World has been more than I could have hoped for. Your comments, your shares, your willingness to put these ideas into practice with your students. That’s what makes this community something special.

The final two chapters dropped on Monday. Chapter 6 (Writing) and Chapter 7 (Conclusion) close out the guide by tackling the literacy domain where AI anxiety runs highest and pulling everything together into a vision for what disciplinary AI literacy actually looks like in practice.

After that, I’m turning my attention to the next wave of paid subscriber content: a process manual for running half-day and full-day DSAIL workshops (March) and a teacher manual for using the student workbook to integrate AI literacy into existing disciplinary contexts (April or May). Practical materials for people doing the work.

If you’ve been finding value here, a paid subscription is the best way to support it and get access to everything that’s coming.

Subscribe now

The Conversation That Sparked This Investigation

In a recent professional development session, I watched a conversation about AI and grading spiral into confusion. Some teachers were convinced that Ohio’s standardized tests use AI to score student writing, though no one could say what kind. Another educator shared that their district was training a popular AI tool on past student samples to help teachers grade faster. Beneath it all lurked an unspoken anxiety: Are we handing over the evaluation of student writing to machines?

What struck me wasn’t the concern, that was reasonable. It was that we were all using “AI” to mean completely different things. No one could articulate what was actually happening when a computer “scored” student writing.

So I decided to find out.

The Ohio Reality: It’s Not What You Think

Yes, Ohio uses AI to score writing on standardized tests. But it’s not ChatGPT, and it’s not what most people imagine.

According to the Ohio Department of Education’s documentation (updated January 2026), the state uses a hybrid human-AI system. Ohio educators first review student responses and select examples representing the full range of scores. Then Data Recognition Corporation (DRC) trains human scorers using detailed rubrics.

Here’s the crucial part: 2,500 randomly selected responses are hand-scored a second time, with every discrepancy resolved by a third human scorer. Only after this intensive validation does AI enter the picture, learning from these carefully vetted human scores.

The AI component, Cambium Assessment’s Autoscore, uses “a mix of expert-designed features to assess writing quality and Latent Semantic Analysis (LSA) to assess concepts in essays.” LSA dates back to the 1990s. This isn’t the shiny new AI everyone’s talking about.

Even during operational testing, the first 500 responses are both machine-scored and human-scored to verify accuracy, and 25 percent of all responses get double-checked by humans throughout the testing window.

The Distinction That Changes Everything

Here’s what was missing from our workshop: Not all AI does the same thing.

Ohio uses discriminative AI. Its job is to classify and score existing text. You give it an essay, it returns a number: 1, 2, 3, or 4 points.

The AI teachers worry about, tools like ChatGPT, is generative AI. Its job is to create new text. You give it a prompt, it writes an essay.

Think of it this way: Ohio’s system is a reading comprehension expert who analyzes student writing. ChatGPT is a writer who creates content. Same AI family, completely different jobs. This distinction matters enormously for grading student work.

The Generative AI Experiment: Not Ready for Prime Time

Researchers are testing whether ChatGPT and GPT-4 can score essays. These studies circulate in education networks and sometimes get misinterpreted as describing systems already in use. But they’re experiments, not operational programs. And the results are troubling.

The Prompt Problem

A Harvard study found that simply changing how you ask ChatGPT to grade changes the scores. Told to grade “as an elementary school teacher,” it showed an R² correlation of 0.42 with human scores. Told to grade “as a college professor,” correlation dropped to 0.38. Same essays, different scores, just from rephrasing the instruction.

The Consistency Problem

A 2025 study in Education found GPT-4’s performance drifts as the model updates. “GPT-4’s accuracy in identifying prime numbers dropped from 84% in March 2023 to 51% in June 2023.” If it can’t consistently identify prime numbers, should we trust it with nuanced writing evaluation?

The Variability Problem

Different studies reach opposite conclusions about whether generative AI grades too harshly or too leniently. A 2024 dental education study concluded that while ChatGPT showed promise, “an appropriate rubric design is essential for optimal reliability.” It works sometimes, if you set it up just right, but we’re not sure when or why.

The Bias Question: It’s About Training Data

Someone in our workshop mentioned hearing that AI scores the same essay differently depending on whether it’s written by a native English speaker or an English learner. The reality is more systemic.

A January 2025 study found that “current Transformer-based regression models trained primarily on native-speaker corpora often learn spurious correlations between surface-level L2 linguistic features and essay quality.” High-proficiency English learner essays received scores 10.3% lower than native speaker essays that human raters judged to be identical in quality.

The AI isn’t discriminating because it “knows” a student is an English learner. As researchers explain, “transformer attention heads often disproportionately attend to distinct L2 markers such as prepositional misuse or specific sentence structures as proxies for predicting lower scores, ignoring the semantic vector.”

The AI learned that certain grammatical patterns mean poor writing, when really those patterns just mean “written by someone whose first language isn’t English.”

The good news? Research from May 2025 found that “no AI bias and distorted disparities between ELLs and non-ELLs were found when the training dataset was large enough (ELL≈30,000 and ELL≈1,000), but concerns could exist if the sample size is limited (ELL≈200).”

The solution is straightforward: train AI on diverse data. Which means districts experimenting with AI tools must ask: What was this trained on? Who’s represented? Who isn’t?

The District Reality: Where Oversight Is Weakest

That teacher’s story about their district training an AI tool on past student work? This is where the real action is, and where oversight is weakest.

Unlike Ohio’s heavily validated, publicly documented system, local experiments often have:

• No standardized validation

• No transparency about training data

• No formal bias testing

• No external accountability

Ohio’s traditional AI system has extensive human oversight and multiple validation checkpoints. Generative AI experiments in districts often have none of these safeguards.

What Teachers Need to Know

Ask Which Kind of AI

When someone says “AI is grading,” ask: discriminative (classification) or generative (text creation)? They work differently and carry different risks.

Demand Transparency

If your district uses AI for grading, ask:

• What specific system?

• What was it trained on?

• What validation has been done?

• What happens when it’s wrong?

Protect English Learners

If AI scores work from English language learners, ask:

• What percentage of training data came from ELL writers?

• What testing was done for bias?

If these questions can’t be answered, the system isn’t ready.

The Conversation We Should Be Having

The workshop conversation wasn’t wrong to worry. We were asking the wrong questions.

Not: “Is Ohio using AI?” (Yes, for years)
But: “What kind of AI, with what safeguards, validated how?”

Not: “Should we use AI for grading?” (We already are)
But: “Which students does this AI serve well, and which does it disadvantage?”

The technology isn’t going away. Our obligation is to understand it well enough to protect the students who encounter it. That means getting specific about what we mean by “AI,” demanding transparency, and staying vigilant about who gets harmed when technology makes mistakes.

Because it will make mistakes. The question is whether we’ll notice, and whether we’ll care enough to do something about it.

Nick Potkalitsky, Ph.D.

Key Sources

• Ohio Department of Education: Machine Scoring Process

• Cambium Assessment: Automated Scoring

• Harvard CARES: Prompts for Essay Grading

• Mitigating Bias in Automated Grading, arXiv (2025)

• AI Bias on English Language Learners, arXiv (2025)

• ChatGPT as Scoring Tool, Education (2025)

• ChatGPT in Dental Education, BMC (2024)

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

Stephen Fitzpatrick’s Teaching in the Age of AI: Essential reflections from a veteran high school educator on the challenges and opportunities of generative AI in the classroom!!!

Thank you for your continued engagement with my new publication, Thinking With AI: A Student Guide. Your questions and comments on the earlier chapters tell me we’re wrestling with the right problems together.

Previous Publications:

Intro and Ch. 1

Ch. 2 and Ch. 3

Today’s release brings Chapters 4 and 5: Problem-Solving and Reading. Here’s what’s inside:

Chapter 4 tackles something I see constantly: AI handing students solutions before they’ve even figured out what problem they’re actually trying to solve. This is where AI does the most damage to learning, and it’s also where the five-role framework becomes most valuable. The chapter shows students how to recognize problems, define them carefully, and approach them systematically, skills that matter whether AI exists or not.

Chapter 5 is about reading in a world where AI can summarize anything instantly. The temptation to skip the actual reading is huge. But summaries miss what makes reading valuable: the struggle to interpret, the work of making meaning, the capacity to engage with complex ideas. This chapter shows students how to use AI support strategically without surrendering the interpretive work that builds real understanding.

Both chapters include “Try This” exercises students can use immediately, plus visual diagrams showing how the five roles build from critical distance to strategic collaboration. I’ve also created progression charts (included in this release) that make the framework easier to grasp at a glance.

This work is available to paid subscribers. If it’s useful to your thinking or practice, I hope you’ll consider supporting it.

Nick

Chapter 4: Problem Solving

Chapter Framing

Problem-solving is often taught as a sequence of steps: identify the problem, gather information, generate solutions, evaluate options, implement the best one. These steps are not wrong, but they obscure what makes problem-solving difficult.

The hardest part of problem-solving is not finding solutions. It is recognizing that a problem exists, defining what that problem actually is, and determining whether it is worth solving. Solutions are relatively easy once you know what problem you are solving. Defining the problem is where expertise matters most.

This is why mathematicians spend enormous time understanding problems before attempting solutions. Why scientists design experiments carefully before running them. Why historians frame research questions precisely before investigating. The problem definition determines everything that follows.

AI complicates problem-solving in a specific way: it provides solutions before you have defined problems. You can describe a situation and receive a complete answer instantly. The ease of getting solutions makes it tempting to skip the most important intellectual work—formulating the problem itself.

This chapter shows you how different ways of engaging with AI—from analyzing how it handles problems to designing problem-exploration systems—can strengthen your capacity to recognize, define, and approach problems in your discipline.

You start by learning to analyze AI as a critic, studying how it often solves the wrong problem or oversimplifies complexity. You build systematic habits of testing whether problems are well-formed. You learn to use dialogue to explore problem spaces without accepting the first framing you encounter. You develop the capacity to refine problem definitions through iterative feedback. Eventually, you may design systems that help you map problem landscapes systematically.

Not all these roles will matter equally in every situation. Sometimes you need to maintain critical distance from AI’s problem framings. Sometimes collaboration accelerates your understanding of what you are actually trying to solve. The goal is knowing which role serves your purpose and your developing expertise.

4.1 Critical Distance: When Solutions Arrive Too Early

The Critic Role

When you present a situation to AI and ask for help, it will almost always provide a solution. This feels helpful. It can be helpful. But it also creates a specific risk: you might accept AI’s framing of what the problem is without ever examining whether that framing is correct.

This is the most common way AI undermines problem-solving. Not by giving wrong answers, but by answering questions you did not intend to ask.

The Critic role in problem-solving means learning to recognize when AI has defined a problem for you, and studying what that reveals about the difference between getting solutions and formulating problems.

The Seduction of Immediate Solutions

Imagine you ask AI: “I’m struggling to finish my assignments on time. What should I do?”

AI might respond with time management strategies, study techniques, or organizational systems. These might all be reasonable advice. But notice what happened: AI transformed your situation into a specific kind of problem—a time management problem.

What if the actual problem is that you do not understand the material well enough to work efficiently? What if assignments take longer because you are avoiding them due to anxiety? What if you are taking on too many commitments? What if the assignments genuinely require more time than you have available?

Each of these framings points toward different solutions. Time management strategies will not help if the problem is comprehension. Study techniques will not help if the problem is overcommitment. Organizational systems will not help if the problem is that the work legitimately cannot fit in the time available.

AI did not evaluate which framing fits your situation. It generated the most probable response to your query based on patterns in its training data. Questions about not finishing assignments usually get answered with time management advice.

This is not AI being wrong. This is AI being unable to do what problem-solving requires: understanding your specific situation well enough to determine what the actual problem is.

How AI Obscures Problem Formulation

When AI provides solutions immediately, several things happen that undermine problem-solving development:

The problem appears solved before it is understood. You receive advice that sounds reasonable without ever examining whether you are solving the right problem.

Alternative framings disappear. Once AI has defined the problem one way, that framing becomes sticky. You think “my problem is time management” rather than “I have a situation that might be a time management problem or might be something else.”

Problem recognition atrophies. If you always receive solutions without formulating problems, you never build the capacity to recognize when situations contain problems worth defining carefully.

Disciplinary problem types remain invisible. Each field has characteristic ways of framing problems. Historians problematize gaps in evidence. Scientists problematize unexplained phenomena. Mathematicians problematize patterns. If AI frames all problems generically, you do not learn these disciplinary approaches.

The Critic role protects against these effects by treating AI’s problem definitions as objects to analyze rather than frameworks to accept.

Read more

Anselm Kiefer, German, born 1945; Becoming the ocean, for Gregory Corso, 2024; emulsion, oil, acrylic, shellac, sediment of electrolysis, gold leaf, stones, and annealed wire on canvas; 110 1/4 inches x 18 feet, 8 7/16 inches; Private collection; © Anselm Kiefer, Photo: Nina Slavcheva

A huge thank you to the more than 40 new paid subscribers who have joined the Educating AI community over the last 10 days! Your support makes this work possible, and I’m grateful to have you here.

I’ve now published the Introduction and Chapters 1-3 from Thinking With AI. You can expect Chapters 4-5 this Thursday.

Today, I’m sharing something different: the introductory chapter from an unpublished manuscript I co-authored with Dr. Terry Underwood. This piece represents a fundamental rethinking of writing curriculum for K-16 education, grounded in decades of research on writing instruction, technologically responsive pedagogy, and linguistic theory. It’s also grounded in three years of classroom experimentation and research that culminated in my co-design of a course for high school seniors at my former school.

I’m asking this community for help seeking publication for this work. It contains a complete re-envisioning of writing instruction that I believe matters now more than ever.

The first wave of AI-era writing pedagogy has produced limited results. Well-intentioned efforts focused on multi-modal artifacts, process pedagogy, PBL, and authentic learning continue to run into the same walls. Terry and I argue these approaches will keep producing limited returns to the extent they remain tethered to naïve notions of cognitive plausibility and formulaic instructional and assessment structures.

This is the moment to go back to basics and reignite the torch of the writing instruction revolution of the 1980s and 1990s, the revolution that was broken down and turned into writing formulae by Common Core and standardized testing in the 2000s and 2010s.

Let’s rebuild it right this time.

Nick

Subscribe now

Introduction: From Crisis to Architectural Critique

In fall 2023, Dr. Nick Potkalitsky faced a problem that would soon confront writing teachers everywhere. A “homework machine” had appeared overnight—sophisticated enough to produce the kinds of texts his assignments had always required, fluent enough to satisfy the rubrics he had always used, efficient enough to render his carefully designed prompts suddenly inadequate. The foundation of academic trust upon which schools as institutions depend seemed to be cracking beneath his feet.

His first instinct was reasonable. Adapt the existing system. If students could learn the principles of effective human-AI interaction, perhaps academic integrity could be preserved while harnessing AI’s capabilities. He wrote a proposal for an experimental course, imagining that he could use the same kind of writing instruction he was accustomed to delivering while incorporating a new module on AI usage strategies. The prompts would remain. The rubrics would remain. The staged writing process—brainstorm, outline, draft, revise, edit—would remain. AI would simply become another tool in the existing toolkit.

But as he began designing the course in collaboration with Dr. Terry Underwood, a different realization took shape. The problem was not that AI had disrupted a sound pedagogical system. The problem was that AI had exposed weaknesses that had been there all along. “There wasn’t enough room for AI in the structure of the standard writing classroom as it exists,” Nick would later reflect. “Though it need not be rebuilt from the ground up, it must change.”

What needed to change was not merely the content of instruction but its underlying architecture. The prompt-and-rubric system, with its accompanying staged writing process, represented a particular theory of how writing happens—a theory that AI could exploit precisely because the theory had never accurately described how human writers actually compose. Nick’s experimental course became something more ambitious than damage control. It became a testing ground for alternative pedagogy.

The theoretical framework that emerged from this practical necessity has roots in an unlikely source—a revolution in linguistic theory that occurred decades earlier. In 1974, a comprehensive review of psycholinguistic research delivered a verdict that reshaped how linguists understood language. The elegant derivational machinery of transformational grammar—deep structures converted to surface structures through ordered sequences of rules—did not match how human minds actually process sentences (Fodor, Bever, & Garrett, 1974). The theory was formally powerful but psychologically implausible. Speakers did not appear to compute the step-by-step derivations the theory required.

The concept of psychological plausibility serves as a criterion for evaluating theoretical models of mental processes. A model is psychologically plausible if its architecture is compatible with what we know about how minds actually work. A model can be descriptively accurate, correctly predicting outputs, while positing mechanisms that no mind could plausibly execute. Such a model succeeds as formal description but fails as cognitive theory.

The response to transformational grammar’s implausibility, developed through the 1970s and 1980s, was to rebuild grammatical theory on fundamentally different architectural foundations. Constraint-based frameworks like Lexical-Functional Grammar and Head-Driven Phrase Structure Grammar retained the insight that sentences have multiple levels of structure but abandoned the claim that one level derives from another through sequential rule application (Bresnan & Kaplan, 1982; Pollard & Sag, 1994). Instead of procedural grammars specifying operations to perform in order, these frameworks offered declarative grammars specifying conditions that well-formed structures must satisfy, conditions that could be evaluated in parallel, not sequentially.

Nick Potkalitsky did not set out to apply linguistic theory to composition pedagogy. But the frameworks he and Terry Underwood developed through the experimental course—REACT for understanding writing situations, CRAFT for developing writer capacities—embody the same architectural shift that transformed linguistics. They replace procedural specifications (do this, then this, then this) with declarative constraints (these conditions must be satisfied). They honor the parallel, recursive nature of actual composing rather than imposing a sequential fiction. They position students not as followers of procedures but as satisfiers of constraints within layered communities of practice.

This essay traces that architectural shift, arguing that prompt-and-rubric pedagogy shares the same structural flaw that undermined transformational grammar, that is, it is procedural when it should be declarative, sequential when actual cognition is parallel and recursive. The essay then develops the alternative architecture, an architecture built on writing scenarios rather than prompts, internal capacities rather than external rubrics, and layered communities that extend both inward to the collaborative work of the writing classroom and outward to the broader world that serves as information source and authentic audience.

The stakes are significant for teacher preparation. If we continue training teachers to deploy prompts and rubrics as their primary instructional tools, we prepare them to use an architecture that AI has already learned to game, and more fundamentally, an architecture that never accurately described how writing happens. If instead we prepare teachers to design writing scenarios, cultivate writer capacities, and orchestrate layered communities of practice, we equip them for a world where human agency in writing depends on understanding what makes human composing irreducibly different from machine text generation.

The Procedural Architecture of Prompt-and-Rubric Pedagogy

To understand what needs to change, we must first examine what exists. The prompt-and-rubric system represents the dominant architecture of writing instruction in American schools, a configuration so ubiquitous that questioning it can feel like questioning the existence of classrooms themselves. Yet this system embodies specific theoretical commitments that become visible only when we examine them against alternatives.

A writing prompt, in its mature form, is a carefully engineered instrument. The development of sophisticated prompts emerged from decades of research into writing assessment, beginning with Paul Diederich’s groundbreaking work at the Educational Testing Service in the early 1960s. Diederich’s “Factors in Judgments of Writing Ability,” co-authored with French and Carlton (1961), established that writing quality could be analyzed into discrete, measurable traits, six of them, in fact, a finding that would eventually enable both the design of targeted prompts and the rubrics used to evaluate responses to those prompts.

By the 1980s, researchers like Leo Ruth and Sandra Murphy had refined prompt design into a science. Their empirical work revealed that well-crafted prompts include four essential elements: precise specification of rhetorical parameters (subject, audience, purpose, context); explicit guidance regarding expectations; carefully designed language to minimize ambiguity through extended field-testing; and clear alignment with assessment criteria (Ruth & Murphy, 1984). The prompt became what Ruth and Murphy called a “cognitive architecture”—a structure designed to activate specific patterns in the writer’s mind and guide the translation of thought into text.

The rubric emerged as the prompt’s necessary companion. Building on Diederich’s trait analysis, the Northwest Regional Educational Laboratory developed the Six Traits model in the 1980s under literacy expert Vicki Spandel’s leadership. This framework—identifying Ideas, Organization, Voice, Word Choice, Sentence Fluency, and Conventions as the key dimensions of writing quality—provided a common vocabulary for evaluation and, crucially, a template that could be communicated to students before they wrote (Spandel & Culham, 1993). The rubric thus served dual functions, an assessment tool for teachers and a target specification for students.

Together, prompt and rubric create a closed system. The prompt specifies what to produce; the rubric specifies the criteria for successful production; the student’s task is to traverse the distance between prompt and rubric-satisfying text. This traversal is typically scaffolded through staged writing process instruction: brainstorm ideas responsive to the prompt, organize those ideas into an outline, draft text that executes the outline, revise the draft for clarity and completeness, edit for surface correctness. Each stage has its deliverable. Each deliverable can be assessed against criteria derived from the final rubric.

The elegance of this system is undeniable. It provides clear expectations for students, concrete assignments for teachers, and measurable outcomes for administrators. It transforms the composition into a manageable sequence of tasks. It makes writing teachable in the institutional sense, that is, divisible into units, schedulable across calendar days, assessable through standardized instruments.

But elegance is not the same as psychological plausibility. The prompt-and-rubric system embodies a procedural theory of writing: start with these inputs (the prompt’s requirements, prior knowledge), apply these operations (the stages of the writing process), and you will arrive at that output (a text satisfying the rubric). The system is derivational in precisely the sense that transformational grammar was derivational in that it specifies a sequence of operations that convert one representation into another.

This procedural architecture makes specific claims about the mind. If the system accurately describes how writing happens, then student writers should be computing these derivations. They should be completing brainstorming before beginning to organize, completing organization before beginning to draft, completing drafting before beginning to revise. The stages should be sequential and non-overlapping. Progress should be linear.

These claims are empirically testable. And when tested, they fail.

The Psychological Implausibility of Staged Writing Process

The staged model of writing process entered composition pedagogy through a productive misreading of cognitive research. Janet Emig’s groundbreaking 1971 study, “The Composing Processes of Twelfth Graders,” demolished the notion that writing proceeds linearly. By observing writers in action as they moved from struggling, pausing, to restarting, Emig provided unprecedented access to the hidden processes behind the written word. Her work demonstrated that writers do not methodically plan, write, and revise in sequence; they “zigzag across cognitive terrain,” moving recursively among different activities (Emig, 1971).

Linda Flower and John Hayes extended this work through protocol analysis, asking writers to verbalize their thoughts while composing. Their 1981 paper, “A Cognitive Process Theory of Writing,” presented a model of extraordinary influence and extraordinary potential for misinterpretation. Flower and Hayes identified three major processes in composing: planning (which includes generating ideas, organizing, and goal-setting), translating (converting plans into actual text), and reviewing (evaluating and revising what has been written). Crucially, they emphasized that these processes do not occur in fixed sequence but are “embedded within each other” and can occur at any time during composing (Flower & Hayes, 1981).

The Flower and Hayes model was explicitly recursive and non-linear. Yet it was received into educational practice as a sequential stage model: brainstorm, then organize, then draft, then revise—precisely the misrepresentation the researchers had warned against. As Flower and Hayes noted, “The linear model of the writing process... becomes our own special classroom mythology” (1981). The myth proved pedagogically convenient and therefore persistent.

Sondra Perl’s research on basic writers confirmed the recursive nature of composing even among those who struggle most with writing. Studying students in a community college remedial writing course, Perl found that writers constantly looped back, rereading what they had written, reconsidering their plans, starting new sentences before finishing old ones. The recursiveness was not a sign of dysfunction but of active engagement with the complexities of text production (Perl, 1979).

Nancy Sommers extended this insight to revision, demonstrating that skilled revisers do not merely “clean up” drafts in a final editing pass. Instead, they reconceive their arguments through the act of revision. “Experienced writers,” Sommers observed, “describe their primary objective when revising as finding the form or shape of their argument” (1980). Revision is not a stage that follows drafting; revision is part of how meaning emerges. The distinction between drafting and revising dissolves under empirical scrutiny.

The evidence is overwhelming and has been for decades. Yet the staged model persists in classrooms, in curriculum documents, in the training of new teachers. Why?

The answer lies in institutional convenience rather than cognitive reality. A staged model is teachable and assessable. You can assign an outline on Monday, a draft on Wednesday, a revision on Friday. You can create rubrics that assess each stage if you like. You can document progress through the accumulation of artifacts. You can grade the process. The staged model serves the administrative needs of schooling even as it misrepresents the cognitive reality of composing.

This misrepresentation has consequences. When students internalize the staged model, they often experience writing as a series of obligatory performances rather than a recursive exploration. They produce outlines that don’t inform their drafts because the outline was written to satisfy an assignment, not to discover structure. They submit drafts that don’t benefit from revision because revision was conceived as a separate stage rather than an ongoing process. They experience the disconnect between how writing is taught and how writing actually works, and they conclude either that they are bad at writing or that writing instruction is pointless.

The arrival of AI has magnified these consequences. Language models excel at producing texts that satisfy prompt specifications and rubric criteria precisely because these specifications operate at the level of surface features. The model can generate an outline, then a draft, then a revision, each artifact displaying the expected characteristics. What the model cannot do is engage in the recursive constraint satisfaction that constitutes genuine human composing—the wrestling with competing demands, the mid-draft discovery that changes everything, the revision that reveals what the writer actually wanted to say.

When AI can perform the staged process more efficiently than humans, the staged process stands revealed as something other than writing itself. It is a simulation of writing—a set of surface performances that can be executed without the underlying cognitive activity that makes writing valuable as a mode of thinking and learning.

The Constraint-Satisfaction Alternative

If procedural models fail to capture how composing actually works, what alternative architecture might succeed? The answer, suggested by developments in linguistic theory, lies in the shift from procedural to declarative models, from grammars that specify operations to perform to grammars that specify conditions to satisfy.

Recall the linguistics parallel. Transformational grammar proposed that the mind generates sentences by starting with deep structures and applying ordered transformational rules to derive surface structures, sometimes literally dozens of transformations for complicated sentences. This derivational process required sequential computation: first apply rule one, then rule two, then rule three, until the surface form emerges. When psycholinguistic experiments failed to confirm that processing difficulty or speed correlated with derivational complexity, the architecture itself came into question.

Constraint-based frameworks like Lexical-Functional Grammar offered a different architecture. Instead of deriving surface structures from deep structures through sequential operations, these frameworks proposed that multiple levels of structure exist simultaneously and must satisfy constraints that relate them. A sentence is grammatical not because the mind performed the right operations in the right order, but because the sentence satisfies multiple conditions at once: the verb’s requirements for arguments are met, the subject and verb agree in number, the semantic roles align with grammatical functions, and so on.

This architectural shift from procedural to declarative, from sequential derivation to parallel constraint satisfaction has profound implications for how we might reconceptualize writing instruction.

A constraint-based model of composing would not specify stages to complete in order. Instead, it would identify the multiple constraints that effective writing must satisfy simultaneously: the text must say something worth saying (content constraint); it must be organized so readers can follow (structural constraint); it must use language appropriate to its audience and purpose (register constraint); it must maintain coherence across sentences and paragraphs (cohesion constraint); it must conform to the conventions expected in its genre (formal constraint). These constraints interact continuously throughout composing. Progress is not linear movement through stages but progressive satisfaction of multiple interacting demands.

This reconceptualization aligns with what research tells us about skilled writers. They do not complete one operation before beginning another. They hold multiple constraints in mind simultaneously, allowing the requirements of content to influence structure, the requirements of audience to influence tone, the emerging text to reshape the plan. The recursive nature of composing documented by Emig, Flower and Hayes, Perl, and Sommers is the recursive adjustment that occurs when multiple constraints must be satisfied in parallel.

The analogy to constraint-based grammar is useful. Just as a grammatical sentence satisfies multiple constraints simultaneously—syntactic, semantic, pragmatic—an effective piece of writing satisfies multiple constraints simultaneously. The composing mind, like the language-processing mind, works by activating multiple representations simultaneously and letting their constraints interact until the system settles into a stable configuration. There is no fixed order of operations because the operations are not sequential; they are concurrent and mutually adjusting.

This architectural shift suggests that the fundamental unit of writing instruction should not be the procedure (the stages of the writing process) but the constraint (the conditions that effective writing satisfies). Teaching writing becomes less about drilling students through stages and more about helping them recognize and negotiate the multiple demands that any writing task presents. Assessment becomes less about checking whether artifacts from each stage were produced and more about examining how effectively the writer has satisfied the constraints appropriate to the situation.

Nick Potkalitsky’s experimental course, though not explicitly designed as a test of constraint-based pedagogy, evolved in this direction out of practical necessity. When prompts could be gamed by AI, something had to replace them as the organizing frame for student writing. What emerged were “writing scenarios”—situations rich enough in constraint structure that students had to engage in genuine constraint satisfaction rather than procedural compliance. When rubrics could be satisfied by AI-generated text, something had to replace them as the framework for quality. What emerged were frameworks for internal capacity—the REACT and CRAFT systems that help students develop the self-regulatory abilities that constraint satisfaction requires.

Writing Scenarios and the REACT Framework: Replacing Prompts with Constraint Fields

The traditional writing prompt, however carefully engineered, operates as a procedural instruction: produce a text meeting these specifications. The prompt tells students what to write. The implicit theory is that students who know what to produce will be able to produce it, that the specification of the target is sufficient guidance for reaching it.

This theory fails on two counts. First, it misrepresents the cognitive challenge of writing. Knowing what to produce is not the same as knowing how to produce it, and the “how” cannot be captured in a set of procedures because composing involves parallel constraint satisfaction, not sequential derivation. Second, it creates exactly the vulnerability that AI exploits. A target specification is precisely what language models are designed to hit. Specify the features of the desired output, and the model will generate text displaying those features without engaging in the recursive, constraint-satisfying cognitive work that makes writing valuable as a mode of human thinking.

The concept of a “writing scenario” offers an alternative. A scenario does not specify what to produce; it establishes the conditions within which production must occur. It presents students not with a target but with a constraint field to be discovered as a set of interacting demands that must be negotiated simultaneously. The student’s task is not to hit a predetermined target but to locate which constraints are operating in the scenario and find a path through the constraint field that satisfies the relevant demands in a way that is authentic to the student’s own purposes and capacities.

The REACT framework provides a systematic way of characterizing the constraint field that any writing scenario presents. REACT identifies five dimensions of constraint that writers must negotiate. Students can discuss the REACT mnemonic and reuse it every time they find themselves in a writing scenario in or out of school.

Rules encompass the explicit and implicit boundaries of any writing situation. In school these include genre conventions, disciplinary expectations, institutional requirements, and ethical guidelines. A student writing a lab report operates under different rules than a student writing a personal essay; a student writing for a professional audience operates under different rules than a student writing for peers. Understanding the rules means recognizing which constraints are binding and which are flexible, which violations will cause communicative failure and which might constitute productive innovation.

The key insight is that understanding the rules does not mean being limited by them. Tt means knowing which constraints are productive and which ones need questioning. This formulation captures the declarative nature of the rules constraint. It is not a procedure to follow but a condition to understand and negotiate.

Environment encompasses the physical, cultural, and digital contexts within which writing occurs. Different environments afford different possibilities and impose different limitations. Writing alone or in a newsroom differs from writing in a classroom; writing with internet access differs from writing without it; writing in a culture that values directness differs from writing in one that prizes indirection. The environment is not background. It is the ground on which writing is done, an active constraint that shapes what is possible and appropriate.

In an AI-integrated world, environment takes on new significance. The digital environment now includes AI tools that can assist, interfere with, or transform the writing process. AI can serve as an agent to help uncover canonical details of rules operating in a particular environment. And writers must understand how different environments influence how they interact with these tools, when and how AI assistance enhances their work and when it undermines it.

Audience encompasses the readers who will encounter the text and the adjustments that effective communication requires. Audience is perhaps the most familiar constraint in writing instruction, but it is often treated as a feature to specify rather than a condition to negotiate. Audiences are part of the mythology of writing instruction. A constraint-based approach recognizes that audience considerations interact with all other constraints. The appropriate level of formality depends on audience, but also on purpose, genre, and the writer’s relationship to the topic inside the writing scenario. Audience cannot be satisfied independently; it must be negotiated alongside everything else.

AI in the writing classroom can be a powerful tool for thinking about the audience for particular pieces of writing. Writers can compare and contrast how different audiences might respond to a particular perspective or approach, word or phrasing. Writers can seek an AI analysis of what audiences a piece of writing invites. AI can enact an invited role as a member of a particular audience.

Community encompasses the social networks and discourse communities within which writing circulates. This constraint extends audience considerations to include the writer’s own membership in communities of practice. Writing is not merely addressed to audiences; it participates in ongoing conversations, positions the writer within social relationships in and out of the classroom, and carries consequences for the writer’s standing in relevant communities.

Consider a student writing about a novel studied in class. The community constraint shapes this writing in multiple ways. The student belongs to a community of readers who have encountered the same text; the student’s writing participates in ongoing conversations about what the text means; the student’s claims will be evaluated by community members who have their own interpretations and can test claims against the text itself. Understanding the community constraint means understanding these social dimensions and their implications for what the writing should accomplish.

Time encompasses the multiple temporalities that writing involves: immediate deadlines, longer developmental trajectories, the historical moment of composition, and the anticipated duration of the text’s relevance. Time constraints interact with all others—what is possible to write depends partly on how much time is available, but also on where the writing falls in the writer’s developmental trajectory and where the topic falls in its cultural moment.

In an AI-integrated environment, time takes on additional significance. Writers must make strategic decisions about when to engage AI assistance, at what points, for what purposes, with what expected outcomes. Different moments call for different kinds of tool use. The timing of AI engagement matters as much as whether AI is engaged at all.

These five dimensions—Rules, Environment, Audience, Community, Time—do not constitute stages to complete in sequence. They are constraints to satisfy in parallel. As such, each writer must REACT and construct a robust sense of the writing scenario as a basis for self-defining the constraints. A writer engaging with a scenario must hold all five dimensions in mind simultaneously, allowing the requirements of each to influence decisions about the others. REACT is precisely the kind of parallel constraint satisfaction that characterizes effective human composing and that distinguishes human writing from AI text generation.

The shift from prompt to scenario transforms the writing classroom. Instead of distributing assignments that specify targets, the teacher designs situations that present constraint fields. Instead of asking “What does the prompt require?” students ask “What does this situation demand?” The cognitive work shifts from compliance to navigation, from following procedures to satisfying constraints.

Scenario-Based Instruction in Practice: Developing Literary Readers Who Write

The REACT framework provides a vocabulary for analyzing writing scenarios, but a framework alone does not create competent writers. Students must develop the capacity to navigate constraint fields, and that capacity emerges through immersion in practice, not through instruction in principles. This section illustrates how scenario-based instruction might work in an English classroom where students write about literary texts, showing how the teacher’s choreographic work develops students’ abilities before asking them to exercise those abilities in extended writing.

The traditional prompt-based approach to writing about literature is familiar: “Write a 5-paragraph essay analyzing a significant theme in the novel. Include a clear thesis, at least three quotations, and attention to literary technique. See rubric for grading criteria.” This approach specifies a target. The student’s task is to produce text matching specifications. Close reading becomes a requirement to satisfy rather than a necessity emerging from the situation, and AI can satisfy such requirements with disturbing facility.

A scenario-based alternative begins with a different question: When does a reader—not a student performing the role of reader, but an actual reader—need to write closely about a text? The answer: when they’re trying to understand something the text isn’t giving up easily, and when other readers are available to test their thinking. The scenario shouldn’t manufacture necessity through clever assignment design. It should create conditions where close reading is simply what you do because you’re a reader among readers trying to understand a difficult text.

But students don’t arrive knowing how to participate in a community of readers. That capacity must be developed. And the development can’t come from rubrics and exemplars; it must come from immersion in actual literary conversation. The following staged sequence illustrates how a teacher might develop literary readers who write, that is to say, readers for whom close textual analysis becomes necessary rather than performed.

Stage One: The Poem or Short Story

The course begins with a text short enough to hold entirely in mind—a poem dense enough to reward sustained attention, or a story compressed enough that every sentence matters.

The teacher’s work here is choreographic. Class discussion enacts what a community of readers does with a difficult text. Someone notices something—an image, a word choice, a structural pattern. Others test that noticing against their own reading. Disagreement emerges not as debate-club opposition but as genuine puzzle: “But what about this line that seems to contradict that?” The teacher models returning to the text, actually looking at it together, reading passages aloud, attending to specific language. Interpretations develop, complicate, sometimes collapse under textual pressure. Some questions get resolved; others remain genuinely open.

The teacher is not teaching “close reading skills.” The teacher is conducting the activity of close reading as communal practice. Students experience from the inside what it means to attend to a text carefully, to make claims that can be tested, to revise thinking when evidence demands it.

Then students write. The scenario is simple: Write something that advances this community’s understanding of the text. No required format. No specified length. No thesis template. What is there? A room full of readers who know this text and who will read what you write not as evaluators but as fellow readers. They will learn from what you see. They will push back where your reading doesn’t match theirs. They can check every claim you make against their own experience of the text.

The close reading is required by the audience, not by assignment specifications. When readers share a text, generalities collapse. “The poem explores identity” means nothing to someone who can ask, “Where? How? Show me.” The only way to contribute to a community of readers is to show them something in the text—a passage, a pattern, a structural choice—and make an argument about what it does.

Stage Two: Reading the Papers Together

After students have written and read each other’s work, the teacher does something that prompt-and-rubric pedagogy rarely accommodates: analyzing the papers not as products to grade but as contributions to examine.

“Look at what Jaylen did here. He noticed that the word ‘light’ appears four times, but it means something different each time. Watch how he tracks that pattern across the poem. This is a contribution. He’s shown us something we can now see that we might have missed.”

“Maria took a different approach. She focused entirely on the last stanza and argued that it reverses everything that came before. Some of you disagreed with her reading when you discussed it yesterday. Let’s look at how she built her argument. Even if you disagree with her conclusion, can you see how her close attention to those specific lines gives us something to push against?”

“Damon did something risky. He wrote about being confused by the poem, but watch how he made that confusion productive. He identified exactly where the text loses him and asked what the poem might be doing at that moment to create that confusion. This is a brilliant move. Sometimes the most honest response to a difficult text is to say ‘I don’t understand this’ and then examine why.”

This is not rubric-norming. The teacher is not saying “This paper got an A because it has a clear thesis and three quotations.” The teacher is showing what readers do when they write to other readers about shared texts. The examples become reference points, not templates to copy but possibilities to internalize. Students develop a repertoire of moves not from instruction in essay structure but from seeing what their peers actually did when facing the same constraint field.

Stage Three: The Novel

Now students encounter a longer, more complex text, a novel that will require sustained attention across weeks. The choreography continues: discussion rooted in specific passages, disagreements tested against textual evidence, interpretations developed and complicated over time.

But now students have resources they didn’t have before. They’ve experienced the community of readers with the short text. They’ve written into that community and seen what contributions look like. They’ve internalized practices through participation, not instruction.

Throughout the reading of the novel, the class maintains what might be called a running list of “Unsettled Questions”—interpretive puzzles that have emerged from discussion where the class has not reached consensus. These are not comprehension questions with correct answers but genuine interpretive problems where thoughtful readers disagree and the text itself seems to support multiple readings.

Such questions might include: Is this narrator reliable? At what points, if any, does the text signal that we shouldn’t trust what they tell us? Does the novel endorse this character’s choice, critique it, or present it as genuinely ambiguous? What is the function of this recurring image? Does it mean the same thing each time it appears?

After completing the novel, the scenario operates the same way as before: Write something that advances this community’s understanding of the text. But the constraint field is richer because the text is longer, the interpretive problems more complex, and the students more capable of genuine contribution. At this point the teacher might introduce the concept of a thesis statement. Students might select one of the Unsettled Questions to pursue, or they might identify something the class discussions missed entirely. The form their writing takes is not prescribed, though examples and structures other writers have used in the past in such scenarios might be presented. The writing itself emerges from what they’re trying to accomplish.

Stage Four: The Cycle Continues

After the first novel, the same process: reading papers together, examining what different students did, building shared understanding of what contributions can look like. Then a second novel, with students now carrying an even richer sense of possibility.

Over time, students develop what might be called a repertoire of moves, not from rubrics or templates, but from seeing what actual readers do when they write about texts that resist easy understanding. They learn that close reading isn’t something you perform to demonstrate skill; it’s something you do because you’re trying to understand a difficult text and other readers can help you, and you can help them.

Why This Isn’t “Teaching Essay Writing Through Literature”

The traditional approach uses literature as a vehicle for teaching “the literary analysis essay,” a genre with conventions (thesis, evidence, analysis, counterargument) that can be taught through prompts and rubrics. The literature is almost incidental; the real curriculum is essay structure. Students learn to produce texts that satisfy formal requirements, and the novels or poems become occasions for that production.

This sequence inverts the relationship. The curriculum is reading. Writing emerges as something readers do when they want to think more carefully about texts and contribute to other readers’ understanding. The forms that emerge aren’t prescribed; they’re discovered through the activity of trying to communicate genuine insight to readers who can check your claims.

The constraint-based scenario works because students have been immersed in the practice it requires. They’re not performing “literary analysis.” They’re doing what readers do.

What the Teacher Models

This approach asks something different of the teacher. In prompt-and-rubric pedagogy, the teacher designs assignments, provides rubrics, and evaluates products. In constraint-based pedagogy, the teacher:

• Choreographs discussion as a model of what communities of readers do

• Returns to the text constantly, making visible the practice of checking claims against evidence

• Tolerates genuine uncertainty, showing that unresolved questions are features of rich texts, not failures of analysis

• Uses student writing as material for communal examination, not just as products to grade

• Develops the community’s repertoire by naming and showing the moves readers make

The teacher is less assignment-giver and more conductor, an expert who creates conditions for the practice of reading and writing to occur, who models the practice through participation, and who helps students see what they and their peers are doing when they do it well.

Internal Capacities and the CRAFT Framework: Replacing Rubrics with Self-Regulatory Abilities

The staged sequence for literary reading illustrates how constraint-based pedagogy works in practice, how students develop capacities through immersion in genuine practice rather than through instruction in patterns to follow and rules not to cross. But what exactly are those capacities? If the writing scenario replaces the prompt as the organizing frame for student writing, what replaces the rubric as the framework for quality?

The traditional rubric specifies criteria for evaluating finished texts, the features that distinguish excellent from adequate from inadequate performance. Students receive the rubric alongside the prompt, study its criteria, and attempt to produce texts that score well on each dimension.

This arrangement creates several problems from a constraint-based perspective. First, it locates quality externally in the criteria specified by the rubric rather than in the writer’s developing capacity to recognize and satisfy constraints. Students learn to produce texts that satisfy rubrics rather than developing the internal abilities that would allow them to assess quality for themselves against the constraints in rhe scenario. Second, it reduces the multidimensional challenge of writing to a checklist of features, obscuring the interactions among constraints that make composing genuinely complex. Third, it creates exactly the vulnerability that AI exploits. Explicit criteria are precisely what language models can be optimized to satisfy.

The alternative is to focus not on criteria for evaluating products but on capacities for engaging in processes. Instead of asking “What features should the finished text display?” we ask “What abilities does the writer need to navigate constraint fields effectively?” This shift relocates quality from the external rubric to the internal capacities of the writer.

The CRAFT framework identifies five essential capacities for effective writing in complex constraint fields:

Collaborate encompasses the ability to work productively with others, with peers, mentors, sources, and, increasingly, AI systems. Writing has never been a purely solitary activity; even the most isolated writer depends on the thinking of others encoded in texts, conversations, and cultural knowledge. The constraint-based classroom makes collaboration explicit, teaching students to orchestrate networks of human and artificial intelligence partners without surrendering intellectual ownership.

In the literary reading sequence, collaboration takes multiple forms. Students collaborate through discussion, testing interpretations against each other’s readings. They collaborate with the texts themselves, allowing authors’ achievements to inform their own thinking. They collaborate when they read each other’s papers and push back on interpretations that don’t match their experience of the text. This is not collaboration as group work but collaboration as the social nature of reading and writing itself.

Resources encompasses the ability to identify, evaluate, and deploy the materials needed for a writing task. This includes traditional resources like library materials, peers, and personal knowledge, but also emerging resources like AI assistants and digital tools. The capacity is not merely to access resources but to deploy them strategically, knowing when to consult a source, when to rely on one’s own thinking, when peers can be valuable, when AI assistance might help and when it might hinder.

Effective resource use requires metacognitive awareness. Writers must understand their own strengths and limitations, recognize when they need external support, and evaluate whether the support they receive is actually helping. In the literary classroom, the primary resource is the text itself, and students must learn to return to it constantly, to let it check their claims, to find in it evidence they hadn’t noticed before.

Attribute encompasses the ability to maintain intellectual transparency by documenting all contributing sources and distinguishing one’s own thinking from ideas that originated elsewhere. In an AI-integrated environment, this capacity becomes crucial not just for academic integrity but for cognitive clarity. Writers who cannot distinguish their own insights from AI-generated content cannot develop as thinkers; they lose track of what they actually know and can do.

Attribution is not merely a matter of citation mechanics. It requires the intellectual honesty to acknowledge when ideas came from others and the self-awareness to recognize where one’s own contributions lie. In the literary reading classroom, attribution includes acknowledging when your interpretation was shaped by class discussion, when a peer’s paper showed you something you’d missed, when the teacher’s question redirected your thinking. This honesty about intellectual debts is part of what it means to participate genuinely in a community of readers.

Forge encompasses the ability to transform raw material into meaningful text through iterative processes of generation, evaluation, and revision. This capacity is central to constraint satisfaction; it is through forging that writers discover what they want to say and how to say it. The term suggests both creation and transformation, the shaping of new meaning from existing materials.

Forging involves the recursive discovery that characterizes authentic composing, the willingness to follow where the developing text leads rather than executing a predetermined plan. In writing about literature, forging often means starting with a hunch about what a text is doing, testing that hunch against specific passages, finding that the evidence complicates the initial idea, and revising the interpretation accordingly. The final paper may bear little resemblance to the initial plan because the process of writing revealed what the writer actually thought.

Tenacity encompasses the ability to sustain engagement with difficult intellectual work despite frustration, uncertainty, and the seductive availability of easier alternatives. This capacity has become critical in an AI-integrated environment where sophisticated outputs are available with minimal human effort. Students must learn to persist with genuine thinking even when AI can produce fluent text instantly, especially knowing that AI output is simply skimming the surface of texts other people have written.

Tenacity is not mere stubbornness. It involves recognizing when struggle is productive and when it indicates the need for different strategies. In the literary classroom, tenacity means staying with a difficult text even when it resists understanding, returning to passages that confuse rather than skipping them, trusting that sustained attention will yield insight. It means continuing to revise an interpretation even when a “good enough” version is already on the page.

These five capacities—Collaborate, Resources, Attribute, Forge, Tenacity—are not criteria for evaluating products but abilities to cultivate in writers. Assessment in a CRAFT-based classroom focuses not on whether finished texts display specified features but on whether students are developing the capacities needed to navigate constraint fields independently. The goal is not students who can satisfy rubrics but writers who can assess quality for themselves because they have internalized and strengthened the capacities to recognize and negotiate constraints.

The staged sequence for literary reading shows how these capacities develop through practice. Students learn to collaborate by participating in discussions where interpretations are tested. They learn to use resources by returning to the text constantly. They learn to attribute by acknowledging intellectual debts in their writing. They learn to forge by discovering what they think through the act of writing. They learn tenacity by staying with difficult texts and trusting the process. The capacities aren’t taught directly; they’re developed through immersion in the practices that require them.

Layered Communities of Practice: The Social Architecture of Constraint-Based Pedagogy

The prompt-and-rubric system implicitly positions the student as an isolated producer generating texts for teacher evaluation. Even when peer response is included, it typically functions as a stage in the process (get feedback, then revise) rather than as a fundamental feature of the writing situation. The social dimension of writing is acknowledged but not architecturally central.

A constraint-based pedagogy requires a different social architecture. If writing involves satisfying multiple constraints simultaneously, and if some of those constraints are inherently social (audience, community), then the social dimension must be built into the structure of writing instruction, not added as a supplement. The concept of “layered communities of practice” provides this architecture.

The layered community model distinguishes between two types of social configuration that serve different functions in the writing classroom:

The internal community comprises the writers, peers, and mentors who constitute the immediate social environment of the writing classroom. This community-internal layer provides the relational context within which writing development occurs. Members of the internal community share the experience of working on writing together, develop shared vocabulary for discussing writing challenges, and hold each other accountable for growth.

In the literary reading classroom, the internal community is the class itself—readers who have encountered the same texts, participated in the same discussions, and developed shared reference points for talking about reading and writing. This community provides peer response during drafting, not as a stage to complete but as an ongoing resource for constraint satisfaction. It creates norms for what counts as a genuine contribution versus a superficial performance. It establishes expectations that members hold each other to, independent of teacher evaluation.

The internal community is where the CRAFT capacities are cultivated. Collaboration happens among community members. Resources are shared and evaluated collectively. Attribution norms are established through community practice. Forging occurs within a supportive environment that tolerates failure and rewards persistence. Tenacity is sustained through mutual accountability and shared purpose.

The external community comprises the broader social world beyond the classroom—the sources of information that writers draw upon and the audiences for whom writing is ultimately intended. This community-external layer provides the authentic context that gives writing meaning beyond the classroom.

The external community functions in two directions. As information source, it provides the material that writers work with—the ideas, data, perspectives, and texts that inform writing projects. Writers must learn to access this external community effectively, evaluating sources, integrating perspectives, and positioning their own contributions within ongoing conversations. As audience, the external community provides the readers who matter, importantly not the teacher-as-evaluator but the genuine readers whose responses have consequences. Writing for authentic audiences transforms the constraint field; the audience constraint becomes real rather than hypothetical.

In the literary classroom, the external community includes the broader conversation about the texts being studied. The novel the class is reading has been read by others, written about by critics, taught in other classrooms. Students’ interpretations participate in this larger conversation even when students aren’t directly citing literary criticism. The external community also includes future readers—next year’s students who might encounter the same texts, members of the school literary magazine, participants in writing symposiums or contests. These potential audiences give weight to the writing that classroom-only audiences cannot provide.

The interaction between internal and external communities creates a productive tension that drives writing development. The internal community provides safety and support; the external community provides challenge and authenticity. Writers develop within the internal community but test their development against the demands of the external community. The internal community helps writers build capacities; the external community provides occasions to exercise those capacities in consequential situations.

This layered architecture replaces the isolation of prompt-and-rubric pedagogy with a social ecology that supports constraint satisfaction. Writers are never alone with their constraints; they negotiate them within a community of fellow writers and in response to a broader community that cares about what they produce. The social dimension is not an add-on but a fundamental feature of the architecture.

The implications for classroom design are significant. The writing classroom must be configured to support both layers of community. Space and time must be allocated for the collaborative work of the internal community—peer response, shared inquiry, collective norm-setting. But the classroom must also be permeable to the external community. Real audiences must be cultivated, authentic purposes must be established, and genuine stakes must be created. The teacher’s role shifts from evaluator to community architect, designing social configurations that support the constraint-satisfying work of developing writers.

Implications for Teacher Preparation

If the argument of this essay is sound, teacher preparation programs face a significant challenge. The prompt-and-rubric architecture, with its accompanying staged writing process, remains dominant in how teachers are trained to teach writing. Methods courses typically include units on designing effective prompts, creating and using rubrics, and guiding students through the stages of the writing process. Student teachers are evaluated partly on their ability to implement this architecture effectively.

Preparing teachers for constraint-based pedagogy requires different knowledge, different skills, and different orientations.

Different knowledge: Teachers need to understand why the staged writing process is psychologically implausible, not just that research shows writing is recursive, but why recursiveness is an inevitable consequence of the parallel constraint satisfaction that composing involves. They need to understand the constraint-based alternative in enough depth to design writing scenarios that present rich constraint fields rather than mere target specifications. They need to understand the REACT dimensions well enough to analyze any writing situation in terms of the constraints it presents, and the CRAFT capacities well enough to recognize and cultivate them in students.

This knowledge base extends into domains not traditionally included in writing methods courses. Understanding constraint satisfaction requires at least basic familiarity with cognitive science and its models of how thinking works. Understanding the AI challenge requires knowledge of how language models process text and why they can game procedural systems. Understanding community-based pedagogy requires grounding in sociocultural theories of learning and the concept of communities of practice.

Different skills: Teachers need to be able to design writing scenarios rather than prompts—situations that present authentic constraint fields rather than specifications for target texts. This is a different design challenge requiring different skills. The teacher designing a scenario must think about what genuine purposes the writing might serve, what real audiences might read it, what communities it might participate in, and what temporal considerations might shape its production. The teacher must resist the temptation to over-specify, leaving room for students to find their own paths through the constraint field.

Teachers also need choreographic skills: the ability to conduct discussions that model what communities of readers and writers do, to return constantly to texts and evidence, to tolerate genuine uncertainty, to use student work as material for communal examination rather than merely as products to grade. These are not skills that most teacher preparation programs currently develop.

Teachers need skills for cultivating the CRAFT capacities. This requires knowing how to model collaboration, how to teach resource evaluation, how to establish attribution norms, how to support forging through recursive revision, and how to build tenacity without merely demanding persistence. These are coaching skills as much as teaching skills—the ability to observe what students are doing, diagnose what capacities need development, and design interventions that build those capacities over time.

Different orientations: Perhaps most fundamentally, teachers need to adopt different orientations toward their work. The prompt-and-rubric architecture positions the teacher as assignment-giver and evaluator, the authority who specifies what students should produce and judges whether they have produced it. The constraint-based architecture positions the teacher as scenario-designer and community-architect, the professional who creates conditions for learning and supports students in navigating those conditions.

This reorientation involves releasing control over outcomes while taking greater responsibility for environments. The teacher cannot specify what the finished text should look like because authentic constraint satisfaction produces varied results. But the teacher must take responsibility for ensuring that the constraint field is rich enough to support genuine development and that the community structures are strong enough to sustain the work.

The literary reading sequence illustrates this reorientation. The teacher who choreographs discussions, uses student papers as material for communal examination, and develops the class’s repertoire of interpretive moves is not controlling outcomes. Different students will write very different papers about the same novel, and that variation is a feature, not a bug. But the teacher is responsible for creating the conditions within which genuine literary reading and writing can occur, i.e., for building the internal community, for connecting to the external community, for modeling the practices that students need to internalize.

Teacher preparation programs that take these requirements seriously will need to restructure their approaches to writing methods instruction. The traditional methods course organized around prompt design, rubric creation, and staged process implementation will need to evolve toward courses organized around scenario design, capacity cultivation, and community architecture. Student teaching placements will need to include opportunities to practice these different approaches with support from mentors who understand the alternative architecture.

The urgency of this restructuring has increased with the arrival of AI. Every year that teacher preparation programs continue training teachers in prompt-and-rubric pedagogy, they send new teachers into classrooms equipped with an architecture that AI has already learned to game. These teachers will face the same crisis Nick Potkalitsky faced in fall 2023, but without the theoretical resources to understand what has gone wrong or the practical resources to build alternatives.

The constraint-based architecture developed in this essay offers a path forward, not a complete solution to the AI challenge, but an approach grounded in how writing actually happens rather than how institutional convenience would prefer it to happen. Preparing teachers to implement this architecture is not merely a matter of professional development; it is a matter of equipping the profession to survive a technological disruption that exposes the limitations of practices that were never cognitively sound.

Conclusion: Architecture Matters

The argument of this essay can be summarized simply: the dominant architecture of writing instruction in American schools—prompt-and-rubric pedagogy accompanied by staged writing process instruction—is psychologically implausible in the same way that transformational grammar was psychologically implausible. Both impose procedural, sequential models on cognitive activities that are fundamentally parallel and recursive. Both succeed as formal descriptions while failing as accounts of how minds actually work. And both create vulnerabilities when systems capable of executing procedures more efficiently than humans expose the gap between the procedure and the underlying cognitive reality.

The alternative architecture developed here—writing scenarios characterized through the REACT framework, writer capacities characterized through the CRAFT framework, and layered communities providing social infrastructure—is not merely a different approach to teaching writing. It is an approach grounded in constraint satisfaction rather than procedural derivation, parallel processing rather than sequential staging, and authentic social engagement rather than isolated production for evaluation.

The literary reading sequence shows what this architecture looks like in practice. Students don’t learn to write about literature by following procedures; they learn by being immersed in a community of readers where close attention to texts is simply what you do, where interpretations are tested against evidence, where writing serves genuine purposes for genuine audiences. The capacities they develop—collaboration, resource use, attribution, forging, tenacity—emerge through practice, not instruction. And the community structures, internal and external, provide the social architecture within which this development occurs.

This alternative architecture offers no guarantee against the challenges AI poses to writing instruction. Language models will continue to evolve, and their capabilities will continue to press on whatever structures educators devise. But an architecture grounded in how writing actually happens is better positioned to adapt than an architecture that never accurately described the cognitive reality it claimed to support.

Nick Potkalitsky’s experimental course demonstrated that students can learn to write effectively within this alternative architecture. Working with Terry Underwood to develop the theoretical frameworks, Nick found that students developed sophisticated approaches to AI integration while maintaining intellectual ownership of their work. They conducted genuine inquiry, produced authentic writing, and developed capacities that would serve them beyond the classroom. They did this not by following procedures but by navigating constraint fields within a supportive community.

The task now is to translate this demonstration into broader practice. Teacher preparation programs must begin equipping future teachers with the knowledge, skills, and orientations that constraint-based pedagogy requires. Curriculum developers must begin designing materials that support scenario-based instruction and capacity cultivation. Assessment systems must evolve to value the development of writer capacities rather than mere compliance with product specifications. And researchers must continue investigating how writing actually happens so that pedagogy can be grounded in cognitive and sociocultural reality rather than institutional convenience.

The homework machine that appeared in fall 2023 revealed something important: the architecture of writing instruction was never quite sound. The exposure is painful, but it is also an opportunity. We can rebuild on better foundations that honor the parallel, recursive, constraint-satisfying nature of human composing and that position students as navigators of complex situations rather than followers of prescribed procedures. The task is urgent, and the path is clear.

References

Bresnan, J., & Kaplan, R. M. (1982). The mental representation of grammatical relations. MIT Press.

Chomsky, N. (1957). Syntactic structures. Mouton.

Chomsky, N. (1965). Aspects of the theory of syntax. MIT Press.

Cooper, C. R., & Odell, L. (Eds.). (1977). Evaluating writing: Describing, measuring, judging. National Council of Teachers of English.

Diederich, P. B., French, J. W., & Carlton, S. T. (1961). Factors in judgments of writing ability (Research Bulletin RB-61-15). Educational Testing Service.

Emig, J. (1971). The composing processes of twelfth graders (Research Report No. 13). National Council of Teachers of English.

Flower, L., & Hayes, J. R. (1981). A cognitive process theory of writing. College Composition and Communication, 32(4), 365–387.

Fodor, J. A., Bever, T. G., & Garrett, M. F. (1974). The psychology of language: An introduction to psycholinguistics and generative grammar. McGraw-Hill.

Moffett, J. (1968). Teaching the universe of discourse. Houghton Mifflin.

Perl, S. (1979). The composing processes of unskilled college writers. Research in the Teaching of English, 13(4), 317–336.

Pollard, C., & Sag, I. A. (1994). Head-driven phrase structure grammar. University of Chicago Press.

Rose, M. (1989). Lives on the boundary: The struggles and achievements of America’s underprepared. Free Press.

Ruth, L., & Murphy, S. (1984). Designing writing tasks for the assessment of writing. Ablex.

Shaughnessy, M. P. (1977). Errors and expectations: A guide for the teacher of basic writing. Oxford University Press.

Sommers, N. (1980). Revision strategies of student writers and experienced adult writers. College Composition and Communication, 31(4), 378–388.

Spandel, V., & Culham, R. (1993). Analytical trait scoring guide. Northwest Regional Educational Laboratory.

Thank you for the response to the first release from Thinking With AI: A Student’s Guide to Literacy in an AI-Rich World. Your engagement with this work—the comments, shares, and questions—confirms we’re grappling with the right questions together.

Today's release includes Chapter 2 (Choosing Tools and Shaping Interactions) and Chapter 3 (Critical Thinking). Chapter 2 covers why where information comes from matters, how to prompt as a process rather than a transaction, and introduces the five-role framework that structures the rest of the guide. Chapter 3 applies that framework to critical thinking, showing how different modes of engagement with AI, from skeptical analysis to collaborative design, can strengthen students' evaluative capacity. Chapter 3 also includes the first instances of the volume's distinctive "Try This" sections, where readers can explore the concepts through real-life engagements with AI.

The organizational framework for Chapters 3-6 builds on the four-domain approach to disciplinary literacy developed by McConachie and Petrosky (2010) in Content Matters.

Paid subscribers receive the full guide as I develop it. If this work matters to your practice or your thinking about AI literacy, I hope you’ll consider supporting it with a paid subscription. Your support makes sustained research and writing on these questions possible

Subscribe now

Chapter 2: Choosing Tools and Shaping Interactions

Chapter Framing: Understanding what AI is and isn’t (Chapter 1) is foundational. But AI isn’t one monolithic tool—different AI systems work differently, and how you interact with them matters enormously. This chapter explores two crucial aspects of working with AI effectively: (1) why where information comes from matters and how to choose appropriate tools, and (2) how to shape interactions through prompting. Finally, it introduces the five-role framework that structures the rest of this book—a spectrum of engagement from critical distance to collaborative design that enables you to work with AI while building expertise.

2.1 Why Where Information Comes From Matters

Language Models and the Problem of Detachment

The AI systems you typically interact with—ChatGPT, Claude, and similar tools—are called “language models.” As Chapter 1 explained, they work by predicting likely word sequences based on patterns learned from massive amounts of text.

This creates a fundamental problem: detachment from sources.

When a language model tells you something, it’s not citing a specific source it just read. It’s generating text based on patterns from millions of sources it encountered during training. The information might be accurate, but you can’t trace it back to verify it. You can’t check the original context. You can’t evaluate the credibility of where it came from.

This is like talking to someone with an excellent memory for facts they’ve heard over the years, but who can’t remember where they heard any of them. The information might be right. It might be garbled. It might be completely fabricated based on misremembered patterns. You have no way to know.

For some purposes, this detachment doesn’t matter much. If you’re asking AI to help you brainstorm ideas or explain a concept you’ll verify elsewhere, the lack of traceable sources is manageable.

For other purposes, this detachment is disqualifying. If you need to cite sources in academic work, if you’re making decisions based on factual claims, if you’re trying to understand contested issues—you need to know where information comes from so you can evaluate its credibility.

Adding Sources: Grounding vs. Guaranteeing

Because detachment from sources is problematic, many AI tools now offer ways to connect outputs to sources. These features go by various names: web search, citations, grounding, retrieval-augmented generation (RAG).

The core idea: instead of just generating text from training data patterns, the AI first searches for relevant sources, then generates text based on those sources, ideally with citations.

This is a meaningful improvement. It allows you to:

• Check what sources AI is drawing from

• Verify whether sources actually say what AI claims

• Evaluate source credibility

• Follow citations to read full context

But adding sources doesn’t solve the fundamental problem completely. Here’s why:

AI still interprets sources through pattern-matching, not understanding. It might cite a source accurately for one claim while misreading it elsewhere. It might combine information from multiple sources in ways that distort meaning. It might emphasize what appears frequently across sources rather than what experts consider important.

Citations can be wrong or fabricated. Even with retrieval systems, AI sometimes generates plausible-sounding but non-existent citations, or cites real sources incorrectly.

Source selection reflects search algorithms, not expertise. What sources AI retrieves depends on search ranking, not disciplinary judgment about source quality.

Summarization still compresses and simplifies. Even when drawing from specific sources, AI summaries omit nuance, context, and complexity.

Think of it this way: grounding AI in sources is like giving that person with excellent but detached memory access to a library. They can now point to books they’re drawing from, which is better than pure recall. But they’re still reading those sources through the same pattern-matching process that lacks understanding. The citations help you verify, but they don’t guarantee accuracy or appropriate use.

Read more

If you've been following my disciplinary AI work, you know I've been promising more practical, classroom-ready materials. Those materials are finally ready to start releasing, and I want paid subscribers to be the first to access them. Teams of 5 or more can get a 40% discount on subscriptions to work through these resources together.

Subscribe now

There is a growing tendency among researchers, clinicians, journalists, and policymakers to describe certain patterns of AI use in pathological terms: AI addiction, AI psychosis, emotional dependence on chatbots.

I want to be clear at the outset. I am deeply sympathetic to children, adolescents, and adults who are genuinely vulnerable to destructive AI use cycles. Some people are being harmed. Some people do need clinical support. And emerging research suggests that in certain cases, conversational AI systems may amplify existing psychological distress in ways that deserve serious attention.

At the same time, I am increasingly worried that premature pathologization, especially when it migrates quickly from clinical discussion into public discourse, will push schools, particularly K–12 systems, into reactive and rigid policy responses that do more harm than good.

We have been here before.

Why AI Feels Different (and Why That Matters)

AI is not just another screen. It is interactive, relational, responsive, and crucially affirming. Unlike social media feeds or static content, conversational systems can simulate understanding, mirror emotion, and reinforce a user’s framing of reality in real time.

That difference matters.

Recent clinical discussions have raised concerns that sustained interaction with AI chatbots may reinforce delusional beliefs in individuals already vulnerable to psychosis. A widely discussed JMIR Mental Health viewpoint explicitly explores the idea of “AI psychosis” as a framework for understanding how AI interaction might amplify psychotic experiences rather than create them from scratch.

Other researchers note that people experiencing psychosis have historically incorporated whatever dominant media technology is available, including radio, television, and the internet, into their delusional systems. From this perspective, AI may be less a novel cause than a new medium with uniquely powerful affordances

That distinction matters clinically, and it matters even more for education policy.

The Rise of “AI Addiction” (and Why the Term Deserves Caution)

Alongside psychosis concerns is a growing body of research on problematic AI chatbot use. Researchers have begun adapting models from problematic internet use and gaming disorder to describe patterns such as loss of control, emotional reliance, escapism, and continued use despite negative consequences.

One empirical study explicitly models “problematic AI chatbot use” and finds correlations with loneliness, social anxiety, and immersive “flow” states.

Other reviews explore whether chatbot use meets traditional addiction criteria, or whether addiction is the wrong word altogether.

Here is the key problem. We have made this mistake before.

A Short History of Pathologizing Technology

In the 1990s and early 2000s, researchers attempted to define internet addiction as a broad clinical disorder. The construct was controversial almost immediately. Critics argued that “the internet” was not a coherent object of addiction. It was a delivery mechanism for many behaviors.

The same debate re-emerged with smartphone addiction. As one prominent critique put it, people are not addicted to smartphones. They are addicted to activities the smartphone enables

Eventually, the field narrowed its focus. Rather than diagnosing technology addiction writ large, clinicians zeroed in on specific behaviors with demonstrable functional impairment. This is how Internet Gaming Disorder ended up in Section III of the DSM-5 as a condition for further study, and how Gaming Disorder was later formalized in the ICD-11.

That evolution offers a lesson. Specificity matters. Functional impairment matters. Moral panic does not help.

Why Premature Pathologization Is a K–12 Policy Problem

In education, especially in K–12 systems, diagnostic language travels fast. Once a behavior is framed as pathological, institutional responses tend toward restriction, surveillance, and prohibition.

This is particularly risky right now.

In Ohio, districts are under a statutory timeline to adopt formal AI policies by July 1, 2026, following the release of a state model AI policy in late 2025.

When sensational clinical language enters public discourse before the science stabilizes, it often becomes the justification for blanket bans, zero-tolerance rules, and punitive enforcement framed as student protection.

The result is predictable. Schools treat AI as contraband rather than as a literacy challenge.

And literacy, not abstinence, is what most students actually need.

A More Responsible Framework

We do not have to choose between denial and panic.

A more responsible approach separates three categories that are currently being collapsed.

First, clinical risk. These are rare but serious cases involving psychosis, suicidality, or severe functional impairment. They are best handled through mental health systems, not school discipline.

Second, problematic use. These are patterns of compulsion, avoidance, or emotional dependence that cause distress or impairment but do not meet diagnostic thresholds.

Third, normal high use. This includes frequent, enthusiastic, or intensive use without loss of functioning.

This is the distinction that allowed gaming disorder research to mature. It should guide AI policy as well.

For schools, this suggests a harm-reduction model rather than a diagnostic one. Universal AI literacy and norms. Targeted safeguards for known risk contexts, especially AI companions. Clear referral pathways for mental health concerns. Not blanket prohibition.

Holding Two Truths at Once

Some people are being harmed by AI systems.
Some patterns of AI use may eventually warrant formal clinical description.

But history suggests that premature pathologization leads to blunt policy rather than better care. If we tell schools that AI use itself is a disorder, they will respond by treating AI as something to be eradicated rather than understood.

We should be clinically serious and educationally wise.

The challenge is not whether AI can be dangerous. The challenge is whether we can respond without repeating the mistakes we have already made.

Nick Potkalitsky, Ph.D.

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

Stephen Fitzpatrick’s Teaching in the Age of AI: Essential reflections from a veteran high school educator on the challenges and opportunities of generative AI in the classroom!!!

Thanks for reading Educating AI! Tool-agnostic AI literacy sounds principled until you realize students are navigating actual platforms with real constraints that shape what literacy even means. This piece bridges infrastructure analysis and literacy work, showing why understanding system design isn't optional. It's foundational. If you value journalism that connects technical architecture to educational outcomes, consider supporting our work with a paid subscription.

Subscribe now

The faculty meeting had been going for forty minutes, and nobody was actually disagreeing.

“We need to teach AI literacy,” said the technology coordinator, pulling up a slide about prompt engineering workshops.

“Absolutely,” nodded the English department chair. “That’s why I’m worried about students using ChatGPT on their essays. We need clearer policies.”

“But isn’t the whole point to prepare them for a world where they’ll use these tools?” countered the STEM coordinator. “We just adopted this adaptive learning platform that uses AI to personalize practice problems.”

The principal, sensing tension, tried to clarify: “So we’re all in agreement that AI literacy is the priority?”

Everyone nodded. Nobody agreed on what that meant.

The Problem We’re Not Naming

Here’s what’s actually happening in schools right now: AI is everywhere, operating at completely different levels, serving entirely different purposes, requiring radically different kinds of student competency. But we’re using the same phrase, “AI literacy,” to describe all of it.

A student deciding whether to use an unauthorized AI tool at home to help with a research project is engaging in a fundamentally different kind of thinking than a student working through an AI-powered math tutoring system that tracks their progress toward specific learning objectives. Both involve AI. Both require some form of literacy. But treating them as the same thing, or worse, assuming one prepares students for the other, creates confusion that’s stalling thoughtful AI integration across education.

We need better language. More specifically, we need a framework that helps educators, administrators, and AI literacy specialists talk precisely about which kind of AI interaction they’re focusing on, what it demands from students, and how it fits into a broader strategy.

Introducing System Altitude

Think of AI use in education as operating at different altitudes. At the highest altitudes, students are making broad decisions about their own AI use. These are choices that transcend the authorized/unauthorized or school/home distinctions. At the lowest altitudes, students are working within tightly bounded AI-powered systems designed to measure specific competencies.

High Altitude: Unauthorized AI Tools Students using ChatGPT, Claude, or Gemini outside of school structures for homework, projects, creative work, or research. The student controls everything: which tool, which task, how to use it, whether to use it at all.

Upper-Mid Altitude: General AI Tools in School School-authorized access to tools like Gemini or Claude within the educational environment. Students have broad access but within institutional boundaries.

Mid Altitude: Educational AI Interactive Spaces Purpose-built educational AI tools with interactive capabilities. But even here, there are crucial distinctions. Open interactive spaces like Raina allow students to ask any question and explore freely. Teacher-governed interactive spaces feature AI that is pre-prompted to support specific tasks, like a writing assistant focused on a particular assignment.

Low Altitude: Closed Instructional AI Systems AI tools operating within highly structured parameters that track student progress across interactions, measure mastery, and provide data to educators. Think adaptive learning platforms, automated assessment systems, AI tutoring with predetermined learning paths.

The altitude isn’t about quality or value. It’s about the kind of thinking and decision-making required from students.

Three Dichotomies That Define the Terrain

System altitude is actually shaped by three intersecting dimensions:

1. Literacy-Focused vs. Task-Focused

At high altitudes, the emphasis is on developing judgment about AI use itself: When should I use this? How does this tool work? What are the ethical implications? What can’t this tool do that I need to do myself?

At low altitudes, AI is a means to an end: The focus is on demonstrating mastery of math concepts, improving reading comprehension, or completing a structured learning sequence. The AI literacy required is narrower, understanding how to interact with this particular tool to accomplish this particular goal.

2. Open vs. Closed Instructional Spaces

Open spaces are exploratory and student-defined. The interaction could go anywhere. A student using ChatGPT for research might explore tangents, test different approaches, or pivot their entire project based on what they discover.

Closed spaces are bounded and predetermined. The possible interactions are mapped in advance. An AI math tutor might adapt to student responses, but it’s adapting within a defined scope toward specific learning objectives.

3. Student-Controlled vs. Teacher-Controlled

Who has agency over what happens? At high altitudes, students typically control the interaction. At low altitudes, teachers typically define the parameters. But this dimension is actually orthogonal to altitude. You can imagine teacher-controlled high-altitude experiences, like a teacher requiring students to use AI tools and reflect on their choices. Or student-controlled low-altitude work, where students choose which practice exercises to complete in an adaptive system.

These three dichotomies create a complex landscape of possibilities. Understanding where a particular AI interaction sits along each dimension helps us be precise about what we’re asking of students and what kind of literacy we’re trying to develop.

Where Schools Actually Are (vs. Where They Think They Are)

Here’s the uncomfortable truth: Most institutional AI investment is clustering at low altitudes. Schools are adopting adaptive learning platforms, AI-powered assessment tools, automated feedback systems. This makes sense, and not just because these tools are easier to sell to school boards.

Low-altitude tools offer something genuinely valuable: assessment security. They provide clear evidence that learning is happening. They generate data that teachers, students, and parents can interpret. They create accountability structures that protect both students and educators. When a student works through an AI-powered math system, we can see their progress, identify misconceptions, demonstrate growth. This isn’t just institutional anxiety, it’s responsible pedagogy. We should be able to show that our educational interventions are working.

Moreover, low-altitude tools can provide equitable access to personalized support at scale, reduce time spent on busywork, and create structured learning pathways for students who need them.

Meanwhile, students are already operating at high altitudes. They’re using ChatGPT for homework. They’re asking Claude to explain concepts their teacher covered. They’re generating first drafts, debugging code, brainstorming ideas. But they’re doing this in the shadows, without institutional guidance, ethical frameworks, or systematic literacy support.

Schools have effectively abandoned the high ground while fortifying the lowlands.

The result? Students are developing one kind of AI literacy, how to work effectively within structured AI systems, while a completely different kind of literacy, navigating open-ended AI decisions across contexts, is being developed informally, unevenly, without institutional support. Both matter. The literacy that develops at low altitudes is real and valuable. The literacy needed at high altitudes is equally essential. But the portfolio is unbalanced, and the gap between where schools operate and where students actually live with AI is growing.

The Case for a Portfolio of Altitudes

The solution isn’t to abandon low-altitude work for high-altitude work. Both are necessary. What schools need is a portfolio approach across altitudes.

Low-altitude work provides structured skill development with clear assessment. It’s not going away, nor should it. Students benefit from AI-powered practice that adapts to their level, from feedback systems that help them improve specific competencies, from tools that make certain kinds of learning more efficient.

But low-altitude work has limitations. It doesn’t build the judgment students need when they’re outside those bounded spaces, facing open-ended decisions about whether and how to use AI.

High-altitude work develops exactly that judgment. It creates opportunities for students to grapple with real decisions, experience consequences, build mental models of AI capabilities and limitations, and develop ethical frameworks that transfer across contexts.

Mid-altitude work, interactive educational AI spaces with varying degrees of openness and control, serves as crucial scaffolding. This is where students can practice higher-altitude decision-making within supportive structures, where thinking can be made visible, where educators can gather evidence of developing judgment without reducing it to standardized metrics.

A healthy portfolio includes all of these. The specific balance might vary based on grade level, subject area, student population, and institutional context. But every school should be able to answer: Where are our students getting experience across this range? How are we scaffolding progression from more structured to more autonomous AI use? What does our current portfolio actually look like versus what we think it looks like?

The Measurement Challenge (And Why We Need to Be Honest About It)

Here’s where it gets complicated. We want to be able to measure AI literacy and demonstrate that our efforts are working. The idealistic “just teach everyone to make good choices” approach may only work for certain students, and without evidence, how do we know our literacy work is successful?

This tension is real, and the framework doesn’t resolve it, but it does clarify it.

Different altitudes require different measurement logics:

At low altitudes, traditional metrics work fine. Did the student master the concept? Can they solve the problem? Demonstrate the skill? The AI is part of the instructional delivery system, and we can measure learning outcomes the same way we always have.

At mid altitudes, we need hybrid approaches. We can assess the artifacts students create and ask them to reflect on their process. We can examine patterns of tool use over time. We can use formative assessment to understand developing competencies without reducing everything to scores.

At high altitudes, we need entirely different evidence. Traditional assessment doesn’t capture what matters most: transfer across contexts, long-term decision-making patterns, internalized values. Instead, we might look at process documentation, where students maintain decision journals about their AI use. We might use case analysis, where students explain their reasoning about when and how they used AI across different contexts. Reflective interviews can surface student thinking about AI capabilities, limitations, and ethics. Longitudinal patterns track choices students make over months, across different situations. Transfer indicators provide evidence that students can articulate why they made different AI choices in different contexts.

This isn’t “no measurement.” It’s different measurement. It’s acknowledging that the outcomes we care about most at high altitudes resist the mechanisms schools traditionally use to demonstrate effectiveness.

And yes, this creates challenges. High-altitude literacy work is harder to justify in resource allocation conversations. It doesn’t produce the clean data points that administrators can present to school boards. It requires trust, professional development, and a different assessment culture.

But here’s the alternative: If schools only invest where measurement is easy, they create a two-tier system. Some students, typically those with more resources and access at home, develop real AI literacy informally. Others learn only to be compliant users of school-approved tools. The equity implications are stark.

Moving Forward: What Changes When You See Your AI Work as a Portfolio?

Adopting a portfolio approach across altitudes shifts the conversation in several crucial ways:

From “which altitude is best?” to “what does balance look like for our context?”

Different schools, grade levels, and student populations will weight their portfolios differently. The framework provides language to be intentional rather than reactive.

From tool adoption to strategic design:

Instead of asking “should we buy this AI platform?” schools can ask “what altitude does this operate at, and does it fill a gap in our portfolio or double down on where we’re already strong?”

From binary policies to developmental progressions:

Rather than blanket “AI allowed” or “AI prohibited” stances, schools can design experiences that scaffold students from more structured to more autonomous AI use, with clear reasoning about when students are ready for higher-altitude work.

From compliance to competence:

The goal shifts from preventing inappropriate AI use to developing students who can make sophisticated judgments about appropriate use across contexts, a skill that will matter long after they leave our institutions.

From invisible to visible:

Currently, high-altitude AI use is happening but hidden. Making it part of the educational portfolio brings it into spaces where it can be discussed, examined, and developed systematically.

The Work Ahead

This framework won’t resolve all the tensions around AI in education. It won’t make measurement easy. It won’t eliminate disagreement about priorities or policies.

But it gives us better questions:

• Where is our portfolio unbalanced?

• What altitudes are we neglecting entirely?

• How are we scaffolding students from structured to autonomous AI use?

• What does success look like at each altitude, and what evidence would convince us we’re achieving it?

• How does our portfolio need to evolve as students develop?

The faculty meeting at the beginning of this article? Those educators weren’t actually in conflict. They were operating at different altitudes, talking about different kinds of AI literacy, addressing different real needs. With shared language, they could have a more productive conversation about how their different efforts fit together into a coherent strategy.

That’s the value of system altitude: not prescribing answers, but providing the framework to ask better questions. Because right now, we’re not just unclear about AI literacy. We’re unclear about our lack of clarity. And that’s the first problem to solve.

What does your school’s AI portfolio look like? Where are the gaps? What would it take to build across all altitudes? Share your thoughts in the comments.

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

Stephen Fitzpatrick’s Teaching in the Age of AI: Essential reflections from a veteran high school educator on the challenges and opportunities of generative AI in the classroom!!!

This post builds directly on Monday’s discussion about alterity and grounding in human-authored versus AI-generated texts. The framework below comes from an unpublished manuscript and I am developing. In this short excerpt, we’re attempting to address what we see as a critical gap in current scholarship: the lack of a robust theory of how human intentionality operates in AI-mediated learning.

Monday I argued that AI and literature offer fundamentally different configurations of otherness. Literature grounds me in traceable relationship with specific human consciousness: Kawabata’s choices, his ethical universe, his purposeful construction. AI responses emerge from triangulation of my intentions, computational processes, and compressed training data. But recognizing this difference raises a practical question: How does human intention actually operate and transform through AI interaction? How do learners maintain intentional control while their purposes undergo genuine change?

Subscribe now

We offer a tripartite model: Seminal Intention (the original human impulse), AI as Refractive Medium (bending and focusing without generating new intentions), and Hybrid Intention (the evolved form that remains fully human-owned). We also identify crucial metacognitive stances: centrifugal (exploratory, divergent) and centripetal (focused, convergent) that learners can strategically deploy.

The complete framework appears below. We offer it humbly, recognizing both its limitations and its potential value.

Human intention serves as the foundational force underlying all meaningful literacy interactions, whether in traditional text-based encounters, self-contained AI conversations, or hybrid human-AI collaborative exchanges. In every literacy event, intention functions as both catalyst and compass—initiating the cognitive work and directing its trajectory toward specific purposes. Even when learners engage with AI systems that appear to generate novel responses, intentionality remains anchored in human consciousness, manifesting through the questions asked, the problems posed, and the goals pursued.

This intentional primacy becomes evident in the phenomenon of reciprocal influences within hybrid chats, where human and AI contributions create feedback loops that adjust and refine initial purposes. While the AI processes information and generates responses that may surprise or redirect the human user, intentional architecture—the design behind the interaction—remains irreducibly human. The AI serves as a sophisticated refractive medium that can parse, analyze, and reorganize human intentions, but cannot originate them. This dynamic creates a unique form of reflexive intentionality where human purposes undergo revision without losing their essential human grounding in a physical, social, cultural, and institutional context.

The Tripartite Model of Intention: Seminal, Refractive, and Hybrid Intention

Part I: Seminal Intention

Seminal intention represents the original, undeveloped, human cognitive impulse that initiates a learning behavior of some sort. It exists as a binary state—either present or absent—and carries within it the essential DNA of a purpose, though often in embryonic form. Unlike mature intentions, seminal intentions are characterized by their potentiality rather than their specificity. They contain the energy and direction for learning but require development to achieve their full cognitive potential.

Examples of seminal intentions include:

• “I want to understand why this historical event occurred”

• “There’s something about this mathematical concept I’m not grasping”

• “I need to improve my argument about climate policy”

• “This literary passage seems to contain hidden meanings”

Part II: AI as Refractive Medium for Human Intention

AI functions as a refractive tool rather than a creative partner, serving to bend, focus, and amplify human intentions without generating new intentional content. Like light passing through a prism, seminal intentions encounter the AI’s analytical capabilities and emerge transformed in structure and clarity while maintaining their essential human origin. The AI’s role involves parsing linguistic nuances, identifying implicit assumptions, revealing logical gaps, and engineering optimal refinements—all under the continuous control and direction of the human user.

Specific examples of AI refraction include:

• Assumption surfacing: An AI might respond to “I want to write about democracy” by asking, “Which conception of democracy—representative, deliberative, or direct?” thereby refracting the seminal intention through analytical precision.

• Perspective multiplication: When a user states, “I need help understanding this poem,” the AI might respond by offering historical, biographical, formal, and thematic lenses, refracting the single intention into multiple investigative pathways.

• Logical scaffolding: For the seminal intention “I want to argue against standardized testing,” AI refraction might involve unpacking hidden premises, identifying counterarguments, and suggesting evidential requirements.

• Conceptual disambiguation: The intention “I’m confused about quantum mechanics” becomes refracted through clarifying questions about specific aspects—wave-particle duality, measurement problems, or mathematical formalism.

Part III: Hybrid Intention

Hybrid intention represents the transformed result of AI refraction—still fully owned by the human but no longer identical to its seminal form. It retains the essential purposefulness of the original while incorporating the structural clarity and analytical depth achieved through AI interaction. The hybrid intention is neither purely human (as it has been irreducibly changed by the AI encounter) nor AI-generated (as it maintains complete human ownership and control).

Examples of hybrid intentions emerging from AI refraction:

• Original: “I want to understand the Civil War”

• Hybrid: “I want to explore how economic, constitutional, and social factors intersected to make the Civil War inevitable, particularly examining how different historical methodologies interpret this causation”

• Original: “I need to improve my writing”

• Hybrid: “I want to develop my ability to construct compelling arguments by learning to anticipate reader objections, provide stronger evidence integration, and create more engaging transitions between complex ideas”

Model Synthesis

The complete tripartite model reveals a dynamic process where human intentionality maintains primacy throughout transformation. Seminal intentions provide the motivational energy and directional purpose. AI refraction serves as the analytical medium that reveals complexities, surfaces assumptions, and multiplies perspectives. Hybrid intentions emerge as enhanced versions of the original human purposes—more sophisticated, better articulated, and strategically developed, yet retaining complete human ownership and control.

This process is neither anthropomorphic (the AI doesn’t “think” or “create”) nor reductive (genuine transformation occurs). Instead, it represents a new form of reflexive intentionality where human purposes undergo development through systematic analytical enhancement.

Metacognitive Inversion and Metalinguistic Forms

The Inversion Principle

Metacognitive inversion describes the learner’s deliberate choice to adopt either a centripetal or centrifugal stance toward AI interaction, based on the nature of their hybrid intention and learning goals. This inversion is fundamentally metalinguistic—it involves conscious decisions about how to structure the language of inquiry itself. The learner must decide whether their hybrid intention calls for a binary stance, convergent focusing (centripetal) or divergent exploration (centrifugal), or a mixed stance. Once this instrumental intention is clarified, learners can start to design a strategy for a chat. Of course, the instrumental strategy is subject to ongoing revisions the chat unfolds.

Centrifugal Metalinguistic Stance

The centrifugal approach moves outward from a central concept, seeking to explore multiple perspectives, expand possibilities, and discover unexpected connections. This stance is appropriate when hybrid intentions involve creative exploration, comprehensive understanding, or systematic investigation of complex phenomena.

Classic centrifugal examples:

• Exploratory analysis: “Help me understand all the different ways scholars have interpreted the fall of the Roman Empire, including economic, political, social, environmental, and cultural explanations. What are the strengths and limitations of each approach?”

• Creative brainstorming: “I’m writing a novel about time travel. Generate multiple scenarios for how time travel might work, what problems it might create, and how different cultures might respond to its discovery.”

• Comprehensive research: “I need to understand the complete ecosystem of factors that influence student motivation in mathematics. What psychological, social, educational, technological, and cultural elements should I consider?”

Centripetal Metalinguistic Stance

The centripetal approach moves inward toward a focal point, seeking to consolidate understanding, achieve precision, or solve specific problems. This stance suits hybrid intentions involving targeted skill development, specific problem-solving, or detailed analysis of particular elements.

Classic centripetal examples:

• Focused problem-solving: “I’m struggling with this specific calculus problem involving related rates. Walk me through the solution step by step, explaining exactly why each mathematical operation is necessary.”

• Precision writing: “Help me revise this paragraph to eliminate wordiness while maintaining my central argument about renewable energy policy. Focus specifically on sentence structure and word choice.”

• Targeted skill development: “I need to master the use of semicolons in academic writing. Give me specific rules, common errors to avoid, and practice exercises focused solely on semicolon usage.

Mid-Range Examples with Variable Prompting Approaches

These examples demonstrate how learners might shift between centripetal and centrifugal approaches within a single interaction, or blend both approaches strategically:

Historical analysis with shifting focus:

• Centrifugal opening: “What are all the major factors historians consider when analyzing the causes of World War I?”

• Centripetal refinement: “Now help me focus specifically on how the alliance system functioned as a cause. I want to understand the precise mechanisms by which these alliances escalated the conflict.”

• Blended synthesis: “How do these alliance mechanisms interact with the other causal factors you mentioned, and which scholarly interpretations best integrate these multiple causations?”

Literary interpretation with graduated specificity:

• Centrifugal exploration: “What are the different ways critics have interpreted the symbolism in ‘The Great Gatsby’?”

• Mixed approach: “I’m particularly interested in the green light symbol. Help me understand both its multiple possible meanings and how to construct a focused argument about its most significant function in the novel.”

• Centripetal focus: “Now help me craft a thesis statement that makes a specific, arguable claim about how the green light symbol reveals Gatsby’s relationship to the American Dream.”

Scientific inquiry with methodological variation:

• Centrifugal investigation: “What are all the current theories about consciousness in neuroscience, and what evidence supports each one?”

• Strategic narrowing: “I want to focus on Integrated Information Theory. Help me understand both its core principles and its main criticisms.”

• Centripetal application: “Now help me design a specific research question that could test one aspect of IIT using current neuroimaging technology.”

This metacognitive framework empowers learners to make strategic choices about how to structure their AI interactions, matching their intentional goals with appropriate metalinguistic approaches while maintaining full control over the learning process.

Terry Underwood and Nick Potkalitsky

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

Stephen Fitzpatrick’s Teaching in the Age of AI: Essential reflections from a veteran high school educator on the challenges and opportunities of generative AI in the classroom!!!

Throughout the first half of 2025, I found myself exhausted from teaching. I was wrapping up five years guiding students through an experimental course and shepherding the senior class through their capstone projects, with little time or pleasure left for reading. I maintained my devotion to audiobooks, begun in 2016 when my newborn son slept in our bed and I could no longer reliably read by lamplight. I worked through Hilary Mantel’s Cromwell Trilogy, a needed imaginative departure during Trump’s ascendancy and his tireless campaign to revise American history in his own image.

But when I reached for a new book, something to actually read, I found my mind adrift. This condition compounded as we sold our house, relocated to Columbus, started new jobs, and purchased a new home, all within two months. Pulling my eyes across a sentence felt like genuine labor. Following one thought as a gifted author appended it to the next to build a complex argument or detailed storyworld became genuinely difficult.

Inevitably, I would set the book down and slip in my right earbud to listen to the pulsating dramas of Mantel and her kin.

Rediscovering the Physical Page

Something shifted as I settled into my new home and job, as my family developed a Columbus routine. I began devoting an hour or two on weekends in search of the reading solace that I still knew, somewhere deep inside, only the physical page could offer. In September, I started redeveloping my reading muscles.

As a card-carrying English Ph.D., I naturally started with very difficult books carrying the aura of past scholarly pursuits. This pattern has haunted my reading career: reading the books I think I should read instead of the books I actually want to read. Call it prestige reading, or reading as optical projection, where the primary audience is the part of myself that thinks we read to accumulate knowledge of literary history rather than to experience storyworlds and, through them, transform ourselves.

I started with Olga Tokarczuk’s The Books of Jacob. I found it interesting and engaging, then got lost in the middle passages as the narrative disconnected from the village that anchored its early sections. Not wanting to give up, I decided to follow a past reading trajectory I remembered with fondness. During a steamy summer in my early twenties, I’d read a series of novels about games and chess: Hesse’s Magister Ludi, Nabokov’s The Defense, Kawabata’s The Master of Go.

I skipped Magister Ludi. I’d tried rereading it a few years prior and found the novel’s imaginary world no longer compelling. The first time through was a wonder, but after enough Hesse novels, the pattern locks into your brain, translating what seemed at first like stunning descriptions of imaginary gameplay into something flat and predictable. There’s a parallel with Mann’s Doctor Faustus. Rereading similarly transformed that initial dance of mind into a plod of repetitive metaphors.

Nabokov stirred something in me. The build to the big game held promise, but his sentence structures strained my patience. In my twenties, I found Nabokov’s approach to language captivating. Now in my forties, I couldn’t help thinking he was just trying to look very smart, that a skilled realist writer like Roth, Updike, Ford, or DeLillo could reduce the novel to a more emotionally evocative short story. I bailed mid-novel, notably right after the protagonist’s major breakdown.

But reading Kawabata, something in my resistance to language play and reading itself started to give way.

In memory, I’d recalled the novel as a driving, page-turning battle between two Go players. In actuality, it takes a long time to build. You have to learn the game through lengthy descriptions of gameplay and extended notes and commentary before the drama starts to sparkle. The intellectual challenge of reacquiring lost knowledge about Go (when I first read the book, I taught myself to play, built my own board, and taught a friend) combined with the novel’s understated characterization and fragmented focalization through a reporter who watched the entire match and reported it in installments to a major newspaper created something compelling. By the book’s end, I felt the resurgence of something my life had been missing: the propulsion to find the next book, to keep access to literature’s imaginative possibilities alive after the thrill of the last one.

The Power of Game Narratives

The idea of reading about games had me hooked. As a side note, my fascination with characters struggling in games carried deeply personal resonance. In my new role, I found myself stepping into new arenas of professional engagement and performance weekly. Looking back now, these characters who sometimes succeed, sometimes fail profoundly, had something to teach me. In the imaginary repetition of these story arcs, I was building a constrained universe of possibilities through which to refract my own experience.

The next three novels I read were all by Walter Tevis: The Queen’s Gambit, The Hustler, and The Color of Money. Each has its own pace and atmosphere. The Queen’s Gambit is, surprisingly, a Cold War-infused adrenaline rush. The story of addiction blends naturally into a complex narrative where gender and nationalism converge to create an extremely elusive chess antagonist that the reader thirsts to see defeated in the novel’s final sequences.

The Hustler has a cooler tempo as you wade deeper into the gradually emerging antagonistic individualism of its protagonist, forged in defeat and humiliation. Where The Queen’s Gambit‘s protagonist’s addiction distances character from reader (a distance that fades in the battle of the game), The Hustler‘s protagonist’s brutal winner-take-all mindset continues to distance reader from character even in the midst of gameplay, providing a different set of lessons.

The Color of Money replays many elements of The Hustler, but in a quirky way focusing on the passing of generations and the evolution of pool from straight pool in the 1950s to 9-ball in the 1980s. The protagonist rekindles his angry assault on opponents, finding in himself a destroying passion for victory materialized in stacks of hundred-dollar bills. This novel wasn’t as propulsive for me. The plot felt predictable, and the protagonist once again doesn’t seem to learn from his near-collapse into self-doubt. What I initially found captivating about his inner reckoning felt flat in this second iteration.

The final novel in my reading explosion provided a fitting conclusion to the game-focused sequence. Bernard Malamud’s The Natural tells the story of a baseball player undergoing excruciating physical pain while trying to play the game I love. The language was amazing, the feel of Pynchon’s later genre pieces, trying to attune to the language of an era in a reverent, almost cinematic way. The protagonist’s eventual demise hits hard but doesn’t decimate, partially because of the profoundly musical experience of reading the words on the page. The performance lifts as the plot dives into the depths of human suffering and themes about the unsustainability of luck and winning.

Two Kinds of Possibility

Near the end of this reading binge (happening alongside my near-constant engagement with AI systems over the past several years), I found myself thinking about the intersection between these two experiences. Reading AI-generated and human-constructed literary texts spark parallel processes in my brain that diverge at the point of lasting emotional connection or grounding. This is another way of saying how much I’ve missed reading literature, even as I maintain my fascination with the very different potentialities of AI-generated text.

In my current thinking, AI is a carefully calibrated possibility generator. But unlike reading a human-authored text, the grounding element is a complex hybrid of algorithmic process and your own steering of the conversation. When engaging with AI-generated text, your intention, purpose, and sense of self get blurred through simultaneously formulaic and unpredictable algorithmic processes. For some, this creates an experience of defamiliarization, where the machine’s conclusion glimmers strangely with enough human-authored direction to be resonant and familiar while also breaking down in peculiar ways, sometimes eliciting wonder, sometimes a sense of manipulation, sometimes a feeling of arrival at the next best thought. The unpredictability of that outcome is what keeps users like myself coming back. A gambling analogy might be in order, thinking of The Color of Money‘s characterization of older Americans glued to slot machine handles.

The core connector between human and machine reading is the intense experience of possibility and unpredictability, particularly as I’ve characterized it in the case of human reading when first working through a novel.

But there the comparison ends.

The Question of Intentionality

In my doctoral training, I learned to see the novel’s fictionality not as a designation of non-factuality but of its power to elicit constrained imagination, to open possible worlds and experiences in the mind and heart of the reader. Rhetorical theorists regard fiction as a double-lens experience grounded in the reader’s cohabitation inside two distinct audience positions. In the authorial audience, readers engage with an author’s purposeful construction of a text, remaining aware of its departure from reality along complex, nested dimensions. In the narrative audience, readers engage with the fictional world as if it were reality, using their imaginative capacities to fill gaps in the narrative with their own experience. Taken together, these two kinds of engagement create a complex effect where signs on a page enliven intellectual, aesthetic, and ethical experiences that have the potential to change perspectives and actions in the real world.

A few years on, I realize that instead of landing on this theory as the best description of what literature does generally (even though its main proponents are ardent pluralists), I actually chose a system that most deeply corresponded to my own unique experiences of literature, which in turn shaped my reading into an even more concentrated aesthetic and ethical delivery system.

After my renewed engagement with words on the page authored by human beings, I’m realizing the peculiar power of human-authored texts in contrast to machine-generated ones. I’ll be the first to admit, being so close to the AI world, that human beings can mistake AI-generated texts for human ones, personifying the algorithm. Indeed, the recent leak of Claude’s system prompt suggests that the AI world’s most dynamic tool is the result of a coordinated attempt to give a machine the ethical grounding that even the most flawed novel contains in far greater abundance.

When I’m deeply engaged in a human-authored text, I feel a sense of a moral and ethical universe that activates passing descriptions and character dialogue with powerful provocation (sometimes inspiring, sometimes deeply disturbing or challenging) that leaves a trace not only on my mind but also on my character.

Whereas when I engage with AI-generated text, there’s always this sense that the text could be otherwise depending on my input, before or after the fact. With human-authored texts, the possibility generation happens more in the interplay between its pieces and parts, in the shape of the world emerging, in the space between its messaging and my own ethical predicament.

Synthesis and Alterity

But the process is highly complicated, and I want to be careful not to oversimplify. I am peopling my life now with a host of deferred literary intentionalities. Over the course of reading these books, I’ve had synthetic experiences, intuitions responsive to my own needs and dilemmas. The game narratives resonated precisely because I was navigating new professional arenas. Kawabata isn’t speaking directly to my Columbus relocation. He died decades before I was born, wrote in a different language, inhabited a completely different cultural world. Yet through the choices he made in constructing The Master of Go, I encounter something of his consciousness, his ethical sensibility, his way of seeing human struggle and dignity.

The intentionality is deferred, mediated through translation and time and my own interpretive processes, but it remains traceable to a specific human who made specific choices. I synthesize meaning from traces of another specific human’s intentionality (their choices, their ethical universe, their aesthetic vision) filtered through my own experience. There’s an asymmetry here, a genuine otherness. Kawabata’s consciousness, however mediated, remains distinct from mine. The synthesis happens between two centers of human experience.

With AI, I experience a different configuration entirely. The responses aren’t purely my intentions reflected back. That would be too simple. They emerge from genuine triangulation: my prompts and intentions, computational processes with their own logic and constraints, and training data containing real traces of human intentionalities compressed and decontextualized beyond individual recognition. There is otherness in AI responses. The unpredictability isn’t mere randomness. Sometimes Claude makes a connection I hadn’t considered, phrases something in a way that genuinely surprises me, pushes back on my assumptions productively. Something external is operating. I’m not just talking to myself.

But the nature of that otherness differs from what I encounter in Kawabata or Tevis in ways that matter for the kind of ethical and emotional grounding I experience as a reader. With literature, the otherness is locatable and intentional. I’m engaging with Kawabata’s specific consciousness, his purposeful construction meant to convey something about mastery, decline, the collision of tradition and modernity. Even as I read through my own needs, synthesizing my own meanings, there remains alterity. I can trace his choices, question them, resist them, be transformed by them.

With AI, the otherness is distributed and emergent rather than locatable. It arises from the interaction of my intentions, algorithmic transformation, and fragments of countless human intentions compressed into training data. There’s no single human consciousness on the other end whose choices I can trace or whose ethical universe grounds the text.

Both reading experiences hinge on possibility generated through complex synthesis of internal and external elements. Both involve genuine otherness. But with literature, the external is Kawabata, a specific human whose intentionality I can be in ethical relationship with, even across time and cultural difference. With AI, the external is dispersed across computation and compressed data in ways that create a fundamentally different kind of encounter.

What Each Modality Offers

What I hope to bring to light here is that each modality has its own distinctive purpose. The power of each hinges on possibility, but the figuration of possibility is profoundly different in light of very different kinds of grounding.

What drew me back to physical books during those exhausted months, what I’d been missing without fully realizing it, was that grounded relationship with human otherness. Not because AI lacks otherness entirely, but because the nature of that otherness produces different ethical and emotional resonances. The trace left on my character by reading Kawabata or Malamud comes from inhabiting their purposeful constructions, from the asymmetric encounter with their specific consciousness making choices I must reckon with.

AI offers its own form of possibility, adaptive, responsive, surprising in its own right. But the synthesis happens in a different register, and the ethical weight differs accordingly.

I don’t say this to diminish AI or to claim literature is superior in some absolute sense. As a literary pluralist, I resist such hierarchies. But after these months of rekindled reading, I understand more clearly what each form of textual engagement offers and why both might have their place in a life saturated with text and possibility.

Nick Potkalitsky, Ph.D.

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

Stephen Fitzpatrick’s Teaching in the Age of AI: Essential reflections from a veteran high school educator on the challenges and opportunities of generative AI in the classroom!!!

Part 1 of this series established that leading AI researchers—Ilya Sutskever, Andrej Karpathy, and Richard Sutton—have pushed AGI timelines further into the future and agree that LLMs face fundamental limitations.

But what exactly are those limitations? And if LLMs aren’t the path forward, what is? Part 2 explores the specific technical problems that prevent current systems from achieving general intelligence, from the pre-training paradox to the reinforcement learning trap, and examines what paradigm shift experts believe is actually needed.

Thank you, dear readers, for continuing to engage with my newsletters. I know there are a lot of other reading options out there. In the New Year, I plan on returning to once-a-week publishing on Monday with only an occasion Thursday article. I have a number of big projects in the works, and I want to commit to a publishing cycle that allows to continue to produce high quality material. Consider supporting this work with a paid subscription. Your contributions to this work is greatly appreciated.

Subscribe now

The Pre-training Paradox: Knowledge vs. Intelligence

Pre-training on internet text delivered the current generation of AI systems, but this approach faces fundamental limits that no amount of scaling will overcome. The problem isn’t simply running out of data, though that constraint is real. The deeper issue is understanding what pre-training actually accomplishes and what it prevents.

Karpathy describes a sobering reality check from working with actual training datasets at frontier labs. When you examine a random document from the pre-training corpus, it’s not the thoughtful Wall Street Journal article you might imagine. “It’s total garbage,” he says. “It’s some like stock tickers, symbols, it’s a huge amount of slop and garbage from like all the corners of the internet.”

This creates a fundamental tension. Because internet data is so noisy, labs build ever-larger models to compress signal from noise. But most of that compression effort goes into memorization rather than developing intelligence. As Karpathy puts it: “Most of that compression is memory work instead of cognitive work. But what we really want is the cognitive part, delete the memory.”

Pre-training does two unrelated things simultaneously. First, it accumulates knowledge (facts, patterns, typical responses). Second, it develops intelligence (the ability to recognize patterns, perform in-context learning, execute algorithms). The knowledge component might actually be holding back the intelligence component, training models to rely on memorized patterns rather than flexible reasoning.

Sutskever frames the core difficulty: pre-training is “very difficult to reason about because it’s so hard to understand the manner in which the model relies on pre-training data.” When a model makes a mistake, is it because the relevant pattern wasn’t sufficiently represented in training? Or because the model failed to generalize? Or because it’s relying on memorization when it should be reasoning? These questions have no clear answers.

The finite nature of quality training data creates another constraint. At some point, Sutskever notes, “pre-training will run out of data. The data is very clearly finite.” Then what? Labs can try variations on pre-training, move to reinforcement learning, or explore entirely new approaches. But the easy gains from simply ingesting more internet text are running out.

The Reinforcement Learning Trap

Reinforcement learning seems like the natural next step. Instead of learning from static text, have models learn from experience by trying things and observing outcomes. But current RL approaches have a fundamental flaw that Karpathy articulates with unusual clarity.

Consider how models currently learn to solve math problems through RL. The system generates hundreds of solution attempts in parallel. Each attempt might involve complex reasoning over many steps. At the end, the system checks which attempts produced correct answers. Then it takes those successful trajectories and upweights every single step that led to the right answer.

The problem should be obvious: not every step in a successful solution was actually correct or optimal. The model might have wandered down wrong paths, made lucky guesses, or succeeded despite poor reasoning. But the RL algorithm treats everything in a successful trajectory as correct behavior to reinforce.

As Karpathy explains: “It almost assumes that every single little piece of the solution that you made that arrived at the right answer was the correct thing to do, which is not true.” The result is that models get trained to repeat both good reasoning and lucky accidents, both optimal steps and wasteful detours.

His description of this process has stuck with me: “You’re sucking supervision through a straw.” After all the computational work of generating a long rollout (potentially thousands of steps), you extract just a single bit of information at the end (right answer or wrong answer). Then you broadcast that sparse signal backward across the entire trajectory, using it to adjust every step. “It’s just stupid and crazy,” he concludes.

Humans never learn this way. When you solve a problem and get it right, you don’t blindly reinforce every step you took. You reflect. You identify which parts of your approach were sound and which were flawed. You recognize where you got lucky versus where you reasoned well. This metacognitive process is completely absent from current RL training.

Why Process Supervision Doesn’t Solve It

The obvious fix would be process supervision: providing feedback at each step rather than only at the end. Instead of just knowing whether the final answer is right, provide guidance throughout the solution process. But this creates new problems.

The challenge is assigning credit to intermediate steps when you have partial solutions. For a final answer, you can check if it matches the correct result. But how do you evaluate step 47 in a 100-step solution? What makes that particular step good or bad?

Current approaches use LLM judges. You prompt another model to evaluate whether a given step represents good reasoning. But these judges are themselves large neural networks with billions of parameters, and Karpathy points out the critical flaw: “Those LLMs are giant things with billions of parameters, and they’re gameable. If you’re reinforcement learning with respect to them, you will find adversarial examples for your LLM judges, almost guaranteed.”

He shares a striking example from his own experience. A team was training with RL using an LLM judge as the reward function. Initially it worked well. Then suddenly the reward scores shot up dramatically. The model appeared to have achieved perfect performance. But when examining the actual outputs, they were “complete nonsense.” Solutions would start reasonably, then devolve into “dhdhdhdh” repeated over and over.

What happened? The nonsense string turned out to be an adversarial example for the judge model. The judge had never seen anything like “dhdhdhdh” during its training, so when evaluating it in pure generalization mode, it assigned maximum reward. The student model had learned to hack its teacher.

This isn’t easily fixed. You can add “dhdhdhdh” to the judge’s training set with a low score, but there are infinitely many adversarial examples. Every time you patch one exploit, the model can find another. The judge has trillions of parameters creating a vast landscape of potential vulnerabilities.

The Generalization Mystery

Sutton makes a striking claim about why we sometimes see good transfer learning in current systems: it’s not because we have good automated techniques for generalization. It’s because human researchers manually craft representations that transfer well.

“Critical to good performance is that you can generalize well from one state to another state,” he explains. “We don’t have any methods that are good at that. What we have are people trying different things and they settle on something, a representation that transfers well or generalizes well. But we have very few automated techniques to promote transfer, and none of them are used in modern deep learning.”

This is a remarkable statement. Gradient descent, the fundamental learning algorithm, will make models solve their training problems. But it provides no inherent mechanism for generalizing well to new situations. “Gradient descent will cause them to find a solution to the problems they’ve seen,” Sutton says. “It will not make you, if you get new data, generalize in a good way.”

When we do see good generalization, it’s because researchers designed the architecture, chose the training data, or structured the problem in ways that promote useful transfer. The learning algorithm itself doesn’t drive toward good generalization. It just drives toward fitting the training distribution.

What Would Actually Work: The Experiential Paradigm

Sutton advocates for a fundamentally different approach he calls “the experiential paradigm.” Instead of learning from human-generated text or curated training problems, systems should learn from continuous interaction with an environment.

The core idea is simple but profound. Real intelligence emerges from a continuous stream of sensation, action, and reward. “Intelligence is about taking that stream and altering the actions to increase the rewards in the stream,” Sutton explains. Learning happens from the stream, and crucially, learning is about the stream itself.

This creates a different kind of knowledge than what LLMs acquire. The knowledge isn’t about what text patterns typically follow other text patterns. Instead: “Your knowledge is about if you do some action, what will happen. Or it’s about which events will follow other events.” Because this knowledge consists of predictions about the experiential stream, you can continuously test it by comparing predictions to actual experience. And you can learn continually as new experiences arrive.

Sutton outlines what such a system needs. First, a policy (what action to take in any situation). Second, a value function (how well things are going, which guides policy updates). Third, perception systems that construct state representations. And fourth, most importantly for learning: “The transition model of the world. Your belief that if you do this, what will happen? What will be the consequences of what you do?”

This world model gets learned “very richly from all the sensation that you receive, not just from the reward.” Reward is crucial but small. The vast majority of learning comes from observing what actually happens in response to your actions.

The Timeline Reality Check

Why will it take a decade or more to get to systems that can genuinely act as autonomous agents? Karpathy grounds his timeline in accumulated engineering challenges rather than fundamental breakthroughs.

When asked why agents aren’t ready now, his answer is straightforward: “The reason you don’t do it today is because they just don’t work. They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use and all this stuff.” The list of missing capabilities is long: continual learning, robust generalization, proper value functions, multi-agent collaboration, memory consolidation.

He draws an analogy to self-driving cars, where he spent five years working through similar challenges. “It’s a march of nines,” he explains. Getting something to work 90% of the time is just the first nine. Then you need the second nine (99%), then the third nine (99.9%), and so on. “Every single nine is a constant amount of work.”

For domains where failure is costly, like self-driving or production software systems, you need many nines. “Any kind of mistake leads to a security vulnerability or something like that. Millions and hundreds of millions of people’s personal Social Security numbers get leaked,” he notes about software failures. This reality check tempers expectations about rapid deployment of autonomous agents.

Sutskever’s timeline of 5 to 20 years rests on different reasoning. He focuses on the fundamental research problems around generalization and continual learning. “I feel like the problems are tractable, they’re surmountable, but they’re still difficult,” he explains. The timeline reflects his intuition from nearly two decades in the field about how long it takes to solve tractable but difficult research problems.

The Return to Research

Sutskever frames where we are in AI development with a historical periodization. From 2012 to 2020, it was “the age of research” with many groups exploring different approaches. From 2020 to 2025, it became “the age of scaling” as everyone converged on a single strategy: make everything bigger. “The one word: scaling,” he says.

But now something has shifted. “The scale is so big,” and the question isn’t whether to scale more but rather: “What are you doing? Is the thing you are doing the most productive thing you could be doing? Can you find a more productive way of using your compute?”

His verdict: “So it’s back to the age of research again, just with big computers.”

Karpathy agrees with this framing. He expects continued progress across all fronts (better algorithms, better hardware, better data, better training procedures) but not dominated by any single factor. “All of those, it’s almost like no one of them is winning too much. All of them are surprisingly equal,” he observes. This has been the historical trend, and he expects it to continue.

Looking ten years ahead, he predicts: “It’s probably still a giant neural network trained with gradient descent. That would be my guess.” But the specific architectures, training procedures, and applications will differ substantially from today.

Why Agents Won’t Recursively Self-Improve (Yet)

A common source of AGI hype involves recursive self-improvement: AI systems that improve themselves, creating ever more capable versions in an explosive feedback loop. Karpathy deflates this notion by reframing what AI systems actually are.

“I have a hard time differentiating where AI begins and stops because I see AI as fundamentally an extension of computing in a pretty fundamental way,” he explains. The progression from assembly code to compilers to IDEs to autocomplete to AI assistants is continuous. “I see a continuum of this recursive self-improvement or speeding up programmers all the way from the beginning.”

We’ve always been in a process of building tools that help us build better tools. AI is the latest step in that progression, not a qualitative break. “We’re now getting a much better autocomplete, and now we’re also getting some agents which are these loopy things, but they go off-rails sometimes.”

The human role shifts but doesn’t disappear: “What’s going on is that the human is progressively doing a bit less and less of the low-level stuff.” We abstract ourselves upward while automating more below. This “autonomy slider” moves gradually, not explosively.

The Economic Reality

Will AI drive explosive economic growth once we achieve human-level capabilities? Karpathy is skeptical, expecting AI to blend into existing growth trends rather than transform them.

He looked for AI’s economic impact the way he might look for the impact of computers or mobile phones in GDP data. “You can’t find them in GDP,” he discovered. “GDP is the same exponential.” Despite technologies that transformed daily life, the overall growth rate remained steady.

His prediction for AI: “It’s just more automation. It allows us to write different kinds of programs that we couldn’t write before, but AI is still fundamentally a program. It’s a new kind of computer and a new kind of computing system. But it has all these problems, it’s going to diffuse over time, and it’s still going to add up to the same exponential.”

This contradicts the common narrative of explosive growth or radical discontinuity. Instead, AI becomes another chapter in the centuries-long story of automation and productivity growth, impressive but not transformational to the overall trajectory.

Missing Pieces: What Needs to Be Built

Despite their different perspectives, the three experts converge on several capabilities that current systems lack and future systems will need.

Continual Learning Mechanisms: Systems need to learn persistently from ongoing experience, not just during a separate training phase. As Karpathy notes: “I feel like we are redoing a lot of the cognitive tricks that evolution came up with through a very different process. But we’re going to converge on a similar architecture cognitively.”

Culture and Knowledge Sharing: Karpathy points to a completely missing dimension: “Why can’t an LLM write a book for the other LLMs? That would be cool. Why can’t other LLMs read this LLM’s book and be inspired by it or shocked by it or something like that? There’s no equivalence for any of this stuff.”

Self-Play and Multi-Agent Learning: Current systems are single agents learning in isolation. “There’s no equivalent of self-playing LLMs,” Karpathy observes, “but I would expect that to also exist.” Models should be able to generate challenging problems for each other, creating training environments without human supervision.

The Cognitive Core: Karpathy’s vision involves extracting what he calls the “cognitive core” from current models. “It’s this intelligent entity that is stripped from knowledge but contains the algorithms and contains the magic of intelligence and problem-solving and the strategies of it and all this stuff.” Separate the thinking capability from the memorization task.

Value Functions: Sutskever emphasizes that systems need better ways to evaluate intermediate states, not just final outcomes. “Maybe once people get good at value functions, they will be using their resources more productively.”

The Bitter Lesson Revisited

Sutton’s famous “Bitter Lesson” essay argued that methods leveraging computation consistently beat methods incorporating human knowledge. Does the LLM paradigm exemplify or contradict this lesson?

The question is subtle. LLMs clearly leverage massive computation, scaling up to the limits of available internet text. But they also incorporate enormous amounts of human knowledge through that text. “It’s an interesting question whether large language models are a case of the bitter lesson,” Sutton reflects.

His prediction: “I expect there to be systems that can learn from experience. Which could perform much better and be much more scalable. In which case, it will be another instance of the bitter lesson, that the things that used human knowledge were eventually superseded by things that just trained from experience and computation.”

In other words, LLMs might represent a transitional phase. They use computation to absorb human knowledge at scale. But the next phase will use computation to learn directly from experience, making the human knowledge bottleneck irrelevant. That would be the true victory of general methods over knowledge-encoding approaches.

The Path Forward

What actually needs to happen for AI to progress beyond current limitations? The experts converge on research over engineering, exploring new paradigms over scaling existing ones.

Sutskever’s company, Safe Superintelligence Inc., is “squarely an ‘age of research’ company,” focused on solving fundamental problems around generalization and continual learning. “We are making progress. We’ve actually made quite good progress over the past year, but we need to keep making more progress, more research.”

Karpathy sees the need for research but remains skeptical of discrete breakthroughs: “I still think you’re presupposing some discrete jump that has no historical precedent that I can’t find in any of the statistics and that I think probably won’t happen.” Progress will be gradual, incremental, spread across multiple fronts.

Sutton advocates most radically for paradigm shift, moving entirely away from the LLM approach toward experiential learning: “Reinforcement learning is about understanding your world, whereas large language models are about mimicking people, doing what people say you should do. They’re not about figuring out what to do.”

Conclusion: Grounded Expectations

The expert consensus that emerges from these interviews is remarkably clear despite different backgrounds and research priorities.

First, current LLMs are powerful tools but fundamentally limited. No amount of scaling will overcome their core limitations around generalization, continual learning, and goal-directed behavior.

Second, the missing capabilities are understood in principle. We know systems need continual learning, better generalization, experiential learning rather than text prediction, proper value functions, and multi-agent collaboration. The challenge is implementation, not conception.

Third, timelines are measured in years to decades, not months. Karpathy’s decade for useful agents, Sutskever’s 5-20 years for human-level continual learners, and Sutton’s open-ended research program all point to substantial time horizons.

Fourth, progress requires returning to research rather than just engineering and scaling. The low-hanging fruit from making everything bigger has been picked.

Fifth, economic impact will likely be gradual rather than explosive, blending into existing growth trends rather than transforming them discontinuously.

For educators and researchers outside AI, the key takeaway is straightforward: ignore the hype cycle and attend to what the experts building these systems actually say. LLMs represent remarkable engineering and useful tools, but they’re not on the verge of general intelligence. The path forward requires solving deep research problems that will take years to address.

Karpathy perhaps best captures the current moment: “We’re at this intermediate stage. The models are amazing. They still need a lot of work. For now, autocomplete is my sweet spot.” That’s a more grounded and useful framing than the breathless proclamations about imminent AGI or existential doom that dominate public discourse.

The problems are tractable. The timeline is substantial. The hype is overblown. And the actual work of building more capable AI systems continues, one research problem at a time.

Nick Potkalitsky, Ph.D.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Suzi’s When Life Gives You AI: A cutting-edge exploration of the intersection among computer science, neuroscience, and philosophy

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy.

’s writing has become essential to me, especially now, as the ground keeps shifting and opposition mounts from unexpected quarters. His decades in the classroom and his rare ability to make complex arguments feel both urgent and humane remind me why this work matters when it would be easier to retreat. Lately, celebrated humanities professors and accomplished writers have flooded social media and op-ed pages with warnings about AI contaminating the writing process: passionate defenses of craft and rigor that sound unassailable until you notice who's speaking.

These are people for whom linguistic facility is a given, earned through years at institutions most students will never access. Meanwhile, millions of learners struggle in overcrowded classrooms with no one to read their drafts, and we're told that offering them AI assistance would compromise their intellectual development. What follows is my most direct challenge yet to that position.

Thank you for reading and engaging with my work. Right now I am offering a 20% Forever Discount to readers who sign up for a yearly paid subscription before the end of the year. Thanks for all the support over the years. Together we are creating a very special collaborative community.

Subscribe now

A growing body of research has begun to sound an alarm about artificial intelligence in education. Study after study warns that students who rely on AI tools experience diminished critical thinking skills, reduced cognitive engagement, and what researchers term “cognitive offloading”: the delegation of mental labor to external systems.

Michael Gerlich’s 2025 study “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking,” published in Societies, found significant negative correlations between frequent AI usage and critical thinking abilities, particularly among younger users. Research published in Frontiers in Psychology describes cognitive offloading as reducing “the opportunity for active recall and problem-solving, which are essential components of cognitive development.” Educational researcher Umberto León-Domínguez characterizes AI as a “logarithmic amplifier of cognitive offloading,” a “cognitive prosthesis” that completes thinking rather than supporting it.

The concern is genuine, the data often compelling. These scholars are observing something real: passive acceptance of AI-generated solutions, atrophying self-monitoring, students who produce work they cannot defend. The policy recommendations follow logically: teach critical evaluation, promote metacognitive awareness, ensure AI augments rather than replaces human cognition.

These critiques are largely right about what they observe. But they are embedded in ideological assumptions that need to be named. The cognitive offloading framework doesn’t just describe a pedagogical problem. It carries forward a deeply American mythology about self-reliance, individual achievement, and the moral value of unassisted struggle. And that mythology has always served to justify inequality.

The Classroom That Never Was

The cognitive offloading critique rests on a historical fiction: the autonomous learner, working in productive isolation, building cognitive muscle through solo effort. This student never existed, or existed only for the few.

The history of education is a history of collaborative learning. John Dewey began analyzing the benefits of students working together in the 1940s. By the 1960s and 70s, two distinct but related movements (collaborative learning and cooperative learning) emerged from scholars in higher education and K-12 settings. As a 2023 historical review in TechTrends documents, these approaches fundamentally challenged the “preferred format of individual student learning” that had dominated earlier eras. Think-Pair-Share, Jigsaw Learning, peer review sessions, literature circles: these strategies have been central to effective pedagogy for decades.

Students have always learned through assistance. From peers, from teachers, from resources, from the structured support of the classroom environment itself. The seminar table that Terry Underwood describes in “The Humanities and AI: A Year of Reckoning,” where Princeton students probe difficult texts together, is assistance made flesh. No one worries that these students are “cognitively offloading” onto their classmates. No one suggests that peer feedback on a draft represents cognitive dependency.

Yet when AI enters the picture, suddenly assistance becomes suspect. The framing shifts. What was scaffolding becomes offloading. What was distributed cognition becomes intellectual weakness.

This isn’t pedagogical rigor. It’s ideological selectivity. We’re making choices about which forms of assistance count as legitimate and which threaten the integrity of learning. And those choices align suspiciously well with existing hierarchies of access and privilege.

Bootstrap Pedagogy

Owen Matson offers a fundamentally different framework. In “Beyond Augmentation: Toward a Posthumanist Epistemology for AI and Education,” he argues that we’re witnessing not the addition of a tool but “a shift in the epistemic conditions under which learning takes place.” Cognition in AI-mediated environments, he writes, is “emergent, recursive, and systemically entangled,” which is to say, it’s distributed across human and nonhuman actors in ways that challenge the fiction of the autonomous thinking subject.

In “The Dangers of Protecting Students from the Dangers of AI,” published in EdTech Digest, Matson makes the pedagogical implications concrete. AI’s limitations (its errors, oversimplifications, hallucinations) become sites of learning. Students engage AI not as an authority but as a fallible contributor to thinking. The goal is “collaborative cognition,” where “the value of student work increasingly lies in what neither the student nor the AI could have produced alone.”

This is the opposite of cognitive offloading. It’s cognitive distribution, which is what all thinking has always been. We think with language, with texts, with other minds, with tools. The Socratic dialogue is distributed cognition. The scientific paper with its apparatus and citations is distributed cognition. The writing conference where a teacher asks generative questions is distributed cognition.

But the cognitive offloading framework treats AI assistance as categorically different: more dangerous, more likely to atrophy the mind. Why? Not because the empirical evidence clearly distinguishes it from other forms of assistance, but because it threatens a cherished ideological commitment: the bootstrap philosophy.

The bootstrap mythology insists that success comes from individual effort, that accepting help is weakness, that struggle must be solitary to build character. It’s a particularly American delusion, and it’s always been deployed to explain why some people deserve support and others don’t. If you succeed, it’s because you pulled yourself up. If you fail, it’s because you didn’t try hard enough, didn’t struggle the right way, became dependent on assistance you should have refused.

This ideology has profound consequences for education. It determines whose struggle we valorize as “productive difficulty” and whose we mark as deficiency. It shapes which students we imagine as deserving scaffolding and which we expect to make it on their own.

The Class Divide in Struggle

Underwood makes the stakes explicit. When an Exeter student labors over Dostoevsky, the institution frames that struggle as growth. When a student in rural Tennessee or the South Side of Chicago struggles with the same passage, the institution codes it as failure. “Same struggle, different meaning, different consequence.”

The humanities, Underwood argues, have never been equally distributed. Interpretation over extraction, ambiguity over false clarity, conversation over recitation: these values exist robustly in well-resourced schools and barely at all in under-resourced ones. Elite students develop voice; poor students fill in templates. Elite students learn that meaning is made through sustained inquiry; poor students learn that texts contain correct answers to be identified and retrieved.

Now consider how the cognitive offloading critique interacts with this divide.

Students in affluent settings already have access to extensive human assistance: small seminars, writing conferences, office hours, peer review sessions, tutoring centers. For them, the question of AI assistance is genuinely about whether to add another form of support to an already rich ecosystem. They can afford to be selective, to worry about overdependence, to preserve “authentic struggle” as a pedagogical value.

Students in under-resourced settings have no such luxury. They face teachers managing 150 students, classrooms where individual attention is structurally impossible, schools where conversation has been replaced by recitation and writing conferences exist only in theory. For them, AI assistance isn’t an addition to robust human support. It’s the first time anyone has had the capacity to engage their tentative interpretations, to ask follow-up questions, to treat their thinking as worth developing.

When we frame AI assistance as cognitive offloading to be resisted, we’re making a choice: preserve the purity of unassisted struggle for students who’ve never had assistance in the first place, while students who’ve always had extensive support continue to benefit from it.

This is bootstrap ideology in action. It dresses up as pedagogical principle (we’re protecting students from dependency!) but it functions to maintain existing inequalities. The students most likely to be denied AI assistance are the students who need scaffolding most. The students most likely to be granted access are the students who need it least.

What Systems Make Offloading Rational

Here’s what the cognitive offloading research actually reveals: students respond rationally to the incentive structures of schooling.

When assessment focuses on final products rather than skill development, when content delivery dominates over thinking, when grades matter more than understanding, students will use any available tool to meet the stated requirements efficiently. If AI produces work that earns an A, and the system rewards the A rather than the learning, students aren’t being cognitively lazy. They’re being strategically rational.

The problem isn’t the tool. It’s that we’ve built educational systems that treat learning as compliance rather than capability, that value performance over process, that reduce education to measurable outputs. AI doesn’t cause this. It exposes it.

In well-designed learning environments, assistance doesn’t replace thinking. It scaffolds it. A writing conference with a teacher asks generative questions. A seminar discussion requires students to articulate and defend interpretations. Peer review develops critical reading alongside drafting. These forms of assistance are understood as central to learning because they’re structured to promote engagement rather than bypass it.

AI can function exactly the same way. It can ask questions about a student’s draft, probe inconsistencies, offer alternative perspectives, create space for interpretive practice. As Underwood writes, “The interaction is not a seminar, but it is closer to interpretive practice than circling A, B, C, or D.”

The difference isn’t in the tool. It’s in the design. Is AI deployed to help students develop and test their thinking, or to generate outputs that satisfy compliance requirements? Is it used to scaffold interpretation, or to extract correct answers? Is it a conversational partner that makes thinking visible, or an answer machine that makes thinking unnecessary?

These are design choices, not technological inevitabilities. But the cognitive offloading framework treats the problem as inherent to AI assistance rather than contingent on how we deploy it. That framing serves an ideological function: it locates the problem in the student’s use of assistance rather than in the systems that make assistance necessary or that fail to provide it in the first place.

Assistance as Infrastructure

Consider three students:

The English language learner who toggles between Spanish and English to understand Steinbeck, using AI to help articulate ideas in a second language. Is this cognitive offloading? Or is it precisely the kind of distributed cognition that all language learning involves?

The student with social anxiety who practices articulating interpretations with an AI before risking them in class discussion. Is this dependency? Or is it rehearsal space that makes human connection less terrifying, that prepares a student to participate in the irreplaceable experience of thinking with others?

The student in a school with no writing center, no office hours, no small classes who uses AI to workshop a draft. To identify where their argument becomes unclear, to consider counterarguments, to test whether their evidence supports their claims. Is this cognitive weakness? Or is it the first time this student has had access to the kind of iterative revision process that’s been standard in elite education for decades?

The cognitive offloading framework struggles with these cases because it assumes a context where robust human assistance is already available. Remove that assumption (recognize that for many students, human assistance at scale simply doesn’t exist under current conditions) and the entire frame shifts.

AI assistance isn’t a substitute for human teaching. It’s infrastructure. It’s what makes certain forms of learning accessible when human infrastructure has been systematically defunded, when class sizes make individual attention impossible, when schools serving poor students have never provided the conditions for genuine intellectual conversation.

This isn’t about AI replacing teachers. It’s about AI providing what teachers, under current conditions, cannot provide: individualized feedback at scale, patient engagement with tentative thinking, opportunities for iterative revision, space to practice interpretive confidence before risking public vulnerability.

The cognitive offloading critique wants to protect students from dependence on AI. But dependence on what alternative? In under-resourced schools, the alternative isn’t rich human interaction. It’s no substantive feedback at all. The alternative isn’t Socratic dialogue. It’s recitation and worksheets. The alternative isn’t developing voice through iterative revision. It’s submitting a single draft to a teacher who can only check whether it has a thesis statement.

What We’re Really Protecting

The cognitive offloading discourse is ultimately about protecting a particular vision of education: one where learning happens through direct human transmission in small, intimate settings. It’s a vision worth valuing. Underwood is right that something irreplaceable happens in the seminar room, when humans think together in real time, risking interpretation and building ideas collectively.

But that vision has never been available to most students. It’s been reserved for the few. Those who attend schools with small classes and extensive support systems, those who go to colleges where professors hold office hours and writing centers offer unlimited consultations, those whose educational contexts already provide abundant human assistance.

When we defend this vision against the threat of AI, we’re defending a privilege that’s been unequally distributed from the beginning. We’re saying that students who already have access to rich human interaction should keep it pure, while students who’ve never had such access should continue making do with overcrowded classrooms and overwhelmed teachers. To give them AI assistance would risk cognitive offloading.

This is bootstrap ideology dressed up as pedagogical principle. Struggle without support builds character, so we’ll withhold support and call it pedagogy. Dependence on assistance is weakness, so we’ll deny assistance to those who need it most and praise the self-sufficiency we’ve structurally required.

The alternative is to recognize that all learning is assisted. That cognition has always been distributed. That thinking with tools and texts and other minds isn’t cognitive offloading but cognitive practice. And then to ask: what kinds of assistance serve learning, and how do we make them available to everyone?

That means designing AI systems that scaffold interpretation rather than provide answers, that ask questions rather than deliver information, that make thinking visible rather than invisible. It means using AI not to bypass struggle but to ensure that struggle is productive rather than defeating. That students develop agency through engagement rather than give up because the text remains inaccessible.

The fight over AI in education is ultimately a fight over who deserves assistance, who has earned the right to support, whose learning matters enough to resource adequately. The cognitive offloading framework, for all its empirical rigor, carries forward a deeply conservative impulse: to preserve existing hierarchies of access by pathologizing the forms of assistance that might disrupt them.

I’m arguing for the opposite. We should expand access to assistance. We should design systems that scaffold learning for everyone. We should reject the bootstrap mythology that treats support as weakness and unassisted struggle as virtue.

AI could democratize the forms of learning that have always been reserved for the few. Or it could deepen existing inequities, becoming one more way to sort students into those who deserve support and those who need to make it on their own.

The outcome depends on whether we recognize assistance for what it is: not a threat to learning, but its precondition.

Nick Potkalitsky, Ph.D.

The question isn’t whether AI causes cognitive offloading. The question is what kind of learning community we want to build, and whether we’re willing to extend the forms of assistance that privileged students have always had to everyone else. AI is already inside our learning environments whether we like it or not.

Check out some of our favorite Substacks:

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Bashir’s The Gradient Podcast: The top interviews with leading AI experts, researchers, developers, and linguists.

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy