There’s a specific feeling, and you know it.

You get the paper back. Red ink in the margins. A grade lower than you hoped. And underneath the disappointment, a quieter, heavier thought slides in: maybe I’m just not smart. Maybe writing isn’t my thing. Maybe this says something about me.

It doesn’t. I want to share with you a different way to read that red ink, and it starts with a boring idea that happens to set you free.

Writing is a technology.

So is English.

Neither one is a test of your worth. They are tools. Complicated, powerful, sometimes badly built tools that humans invented and keep changing. And learning a tool follows a curve that has nothing to do with how smart or lovable or capable you are.

Think about every technology you now use without thinking. The first time you rode a bike, you wobbled and fell and were sure everyone was watching. The first time behind the wheel of a car, your hands were tight and the whole thing felt impossible. Typing felt like that once. So did tying your shoes. None of those wobbles were verdicts on your soul. They were just the curve. You climbed it, and now you don’t even remember the climb.

English is the same, except messier. Be honest with yourself about the tool you’re being graded on. English spelling is a patched-together disaster. Through, though, thought, tough, cough: five words that rhyme with almost nothing and almost each other. The rules contradict the rules. Half the “silent” letters are fossils from languages that died centuries ago. You did not fail a clean, fair test. You’re installing buggy software into a human brain, and a lot of you are installing it on top of another language that’s already running just fine.

So when the writing comes out clumsy, that’s not a readout of your intelligence. That’s a tool with a learning curve, working exactly the way learning curves work.

Now here’s the part that few tell you, and it’s the good part.

Tools get replaced.

The technology you’re sweating over right now is not permanent. Spelling will get easier. Drafting tools will get smarter. The way humans put thoughts into other humans’ heads keeps adjusting, and it will keep moving across your whole life. The struggle you feel today is sitting on top of a tool that someone, someday, is going to redesign.

And here’s the wild thing: the person who builds the better tool is usually someone who hated the old one. They struggled with it. They felt exactly where it was broken, in their hands, in their gut. That feeling is not weakness. That feeling is product research.

You are learning, right now, precisely where writing-in-English breaks. That makes you one of the few people who could fix it. You could vibe-code a writing tool tomorrow that helps the next kid who gets that paper back with the red ink. You might turn out to be the Shakespeare of a technology nobody has invented yet, fluent in a tool that doesn’t exist while you’re stumbling through one that does.

So let go of the lie hiding in the phrase “good writer,” like it’s a fixed thing you either are or aren’t, stamped on you at birth. There is no such stamp. There’s a curve, and you’re on it, and where you stand on the curve today says nothing about where you’ll stand next year.

Read that red ink again. It is not a sentence handed down about who you are. It’s a patch note. Here’s what to fix in the next version. That’s all it ever was.

Other people’s opinions about your writing do not get to decide what you’re worth. Pop music is full of anthems built on exactly that idea, that nobody else’s words get to set your value. Go play one. Loud.

But I’ll be honest with you, because you deserve honesty.

Sometimes words land harder than a grade. Sometimes they’re said early, by people who were supposed to keep us safe, and they don’t bounce off. They stick. They leave a mark you carry for years, replaying in a voice that sounds like your own but never really was.

That’s a different story, and it deserves its own letter.

Next: some words leave scars, and you can’t vibe-code those away. What actually helps, and who you actually need.

Do you remember when onions tasted harsh, or when all red wine tasted the same? Taste develops through exposure, attention, and experience, and it can also be dulled by a diet saturated with easy sweetness. Writing works the same way. When a machine can draft the essay, the distinctly human skill is not writing the first draft but appraising what the draft is worth. This means sensing what is vivid instead of generic, insightful instead of secondhand, narratively alive instead of mechanically arranged, and evident rather than merely plausible. It is paying critical attention.

Henrik Karlsson’s most beloved essay is called Looking for Alice. It is about how he found his wife. He wrote it as private advice to a friend, it leaked, and it traveled around the world. Two and a half years later, Karlsson reread it and delivered a verdict few writers would survive saying out loud: “at least half of this advice is kind of shady.”

And yet the essay still works. He now understands why, and it isn’t the advice. It works because “you can feel how much I love my wife.”

Consider what that admission means. The energy I wrote about in Part 1 (the snap, the bodily yes, the release Karlsson trusts to tell him a piece is done) was fully present in “Looking for Alice.” The feeling was accurate about the love and wrong about the theory. The body’s signal verified emotional truth. It did not verify factual truth.

So taste, real taste, needs two loops. The somatic loop: a trained body that registers when structure, stakes, and honesty align. And the verification loop: the painter’s discipline, shared in Part 2, claims tested against real cases, against counterexamples, an adversarial reader invited to provide feedback. Karlsson runs both, and runs them separately. The sofa for the first, his wife as adversarial reader and his case files for the second loop. A writer with only the first process becomes eloquently wrong. A writer with only the second becomes correct and dead on the page.

Here is what strikes the educator in me. Part 2 sorted what a writer must never outsource, the posture. Look at everything Karlsson describes across the whole interview, and a second set of four principles becomes clear.

He insists on the vivid. Wittgenstein’s “don’t think, look.” The still-life exercises. Delacroix noticing that a shadow is yellow. Concrete reality, pulled into the prose until the prose stops being about ideas and starts being about the world.

He hunts the insightful. The obvious truths nobody can hear anymore (be kind, honor your commitments) sequenced so they deliver. “I want the piece to end up smarter than I am.”

He engineers the narrative. An animating question with stakes. Open loops the reader needs closed. Images placed to prepare an emotional state before the argument arrives.

He demands the evident. Three or four real cases per essay. Biographies as data. The standing question scribbled in his own margins: is that really true?

Vivid. Insightful. Narrative. Evident. Readers of my book will recognize VINE, the framework I built in The Learn-It-All Educator for training exactly the phenomenon Part 1 named: taste, the judgment that separates average work from excellent work. I did not derive it from Karlsson, which is why watching a master essayist’s native practice decompose into the same four qualities is the most encouraging kind of evidence. Good frameworks are not inventions. They are descriptions of what excellent practitioners already do, packaged so the rest of us can practice deliberately.

Why does this matter now, and why for educators in particular? Three reasons, in ascending order of stakes.

First, the standardization pressure is real. Perell names it in the interview: LLMs are just another force for standardization. Karlsson, a power user of GenAI, agrees without hedging and wants English pushed the other way, wilder. A model’s default prose is, almost by definition, the weighted average of everything. Fluent, useful, centered. Taste is the capacity to feel the difference between that center and the live thing, and like any perceptual capacity it sharpens with contrast and atrophies without it. Students who marinate exclusively in average prose lose the contrast.

Second: thinking grows under load, the argument of the Cognitive Gym chapter in my book. Karlsson’s method runs on deliberately engineered confusion. He writes a claim down so that he “can’t fool myself,” watches it break under his own questioning, and rides the confusion until it collapses into a simpler, truer formulation. He even cites the research tradition behind accelerated expertise: beware the comfortable, half-right mental model, the “knowledge shield,” because it ends learning early. Now watch what happens when a student lets a model draft the essay. The confusion stage is precisely what gets skipped. The load disappears, and the growth disappears with it. The essay was only ever the receipt for the thinking.

This is why the assessment shift I argue for moves from generation to verification. If the machine can produce a draft, the gradable human act becomes the audit. Is this vivid, or generic? Is the insight real, or a rearranged cliché? Does the narrative earn its turn? Is it evident? Show me the case that survives a counterexample. Karlsson grades his own drafts this way every working day. We can teach students to do the same, including to AI output. Especially to AI output.

Third, the reason that moves me most. One of Karlsson’s best-known essays has a title two dozen words long, and its thesis is that a blog post is “a very long and complex search query” for finding your people. Two years before that essay, in December 2021, he had about fifty readers. He kept publishing honest, strange essays anyway, and they went out across the network and came back with collaborators, mentors, friends: a summoned culture that then changed him in return. That flywheel spins on signal. In an ocean of competent average prose, the writing that can still summon anyone is the writing with a detectable human inside it, and the readers who can still be summoned are the ones who can detect one. Taste on both ends. Voice and taste are the network protocol of intellectual life, and educators are now responsible for keeping that protocol alive in the next generation.

Karlsson, asked how he would teach writing, refused the premise of a skills course. Writing well, he said, “is not a skill set”; it is “becoming a certain type of person.” Many educators know that sentence in their bones, because it has always been the actual job description. The syllabus says rhetoric. The work is formation.

The machines will keep getting better at the average. Our students will swim in fluent, frictionless, centered prose for the rest of their lives. The educators who matter in that world are the ones who train both loops: a body that can feel the live thing and a mind that checks it against reality.

The energy in the page was never magic. It is a trained body and a tested truth arriving together, past the reader’s defenses.

Teach them to taste.

The frameworks in this series (VINE, the Cognitive Gym, the shift from generation to verification) are developed fully in The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them, also available through the Open Textbook Library. If this series found you, the search query worked. You are probably my people. Subscribe and say hello.

“The essay was only ever the receipt for the thinking”

“Voice and taste are the network protocol of intellectual life”

“the distinctly human skill is not writing the first draft but appraising what the draft is worth”

Here is a puzzle for everyone who has ever written “find your authentic voice” on a syllabus.

Henrik Karlsson is Swedish. In the Swedish language, he told on How I Write, word order is wonderfully loose. He could tune a sentence’s rhythm by sliding words around “like on a scale,” nudging a phrase left or right until the music was right. English revoked that license. English word order is strict; move a word and the sentence breaks. He also commands fewer English words than a native writer. Sometimes he has to describe a word to an AI because he doesn’t know it. The meaning is in him; the vocabulary isn’t.

By the diction-level definition of voice we mostly teach, Karlsson-in-English should sound diminished: a writer stripped of his native instrument, working under a harsher grammar with a thinner lexicon. Instead, Escaping Flatland is one of the most distinctive bodies of prose on the internet, written in a language he learned. Readers who have seen one paragraph can identify the next.

English is my second language too, so I’ll confess the puzzle is personal. Nearly everything my own teachers graded as “voice” lives at the layer Karlsson lost in translation. Whatever voice is, it survived the loss, so it must live somewhere else.

The interview shows where. Listen for what Karlsson actually guards, and a different anatomy of voice emerges.

He guards his stance. Every essay begins as a question with personal stakes, never a topic chosen for an audience. The writing he wants to do, he says, “is asking me to become a better person.” You cannot fake the gravity that gives a piece.

He guards his honesty. When Perell brings up David Foster Wallace’s warning that most writing is deformed by an overwhelming need to be liked, Karlsson doesn’t theorize, he describes his countermeasures. He catches himself writing lines because they’re cute, scratches them out, and mourns the lost momentum. He spends at least half of his days with no internet connection. He drafts on paper while pacing the farm road, or talks a piece out with his eyes closed, engineering situations in which he can forget he will ever be read. Voice, in this anatomy, is what remains after performance has been subtracted.

He guards the verification loop. This is the part my computer science students would recognize on sight. Each essay carries three or four real cases: episodes from his life, his friends’ lives, biographies. Against these every claim gets tested, the way code gets run against a test suite. His wife reads as an adversarial reviewer: not crossing out words, but stopping on a line to say, in effect, let’s go find a biography and see whether that holds. He interviews himself in his notebook margins: what do you mean here, give me three examples. Then he checks whether the examples match what he wrote. Usually they don’t, and the rewriting that follows is where the essay gets good. Voice here is earned per piece, through checking.

And he guards his nerve. When Perell asks how much his writing is limited by courage, Karlsson doesn’t hedge: “probably always the biggest bottleneck.” He admits he would rather tell himself the problem is something craftier (better metaphors, better openings), but that, he says, is usually an excuse for not going deeper.

Stance, honesty, verification, nerve. Notice what’s missing from the list. The one layer Karlsson cheerfully outsources is diction: his total AI use comes to about one percent of his words, almost all of it word-finding, grammar checks, and asking for ten alternate phrasings when a sentence won’t sit right: a prosthetic for the second-language gap. The machine touches the paint, never the looking.

Literature has run this experiment before. Conrad (Józef Konrad Korzeniowski) wrote his way into the English canon in his third language. Nabokov switched literary languages and dazzled. Samuel Beckett is the strangest case of all: a native English virtuoso who moved into French specifically to escape his own facility and write, as he put it, “without style.” Beckett is Karlsson’s mirror image, and together they triangulate one conclusion: a writer’s voice and a writer’s native diction can be separated cleanly, in either direction.

That should change how we hear the one real worry about AI in this stretch of the interview. Karlsson is no Luddite; he was a power user before the 2022 Generative AI uprising. When Perell names the danger, that LLMs are just another force for standardization, Karlsson takes the point and pushes the other way. His answer is not to ban the machines. It is to want English wilder: bring in the strange grammar of Turkish, the cool words of Persian, the innovations of every language English is absorbing, even though, as he admits, “you look stupid when you do it.” I made a related argument in EDUCAUSE Review, in a piece called “English 2.0”, about how AI is already reshaping the language itself. The threat to voice was never the thesaurus. The threat is regression to the mean.

Now think about who sits in our classrooms. At University of Illinois Chicago or the community college, many students are multilingual. When they hear “authentic voice,” they hear “native diction,” and they quietly conclude the assignment is unwinnable. Karlsson is the counterexample I want to share with them. The diction layer is the one part of writing you may delegate (to a dictionary, a friend, or yes, a model) without touching what makes the writing yours. What cannot be delegated: the stance, the honesty under observation, the verification loop, the nerve.

And those four, unlike “voice,” are teachable. You can assign the looking. You can grade the test cases. You can build the margin interrogation into a visible process and watch the nerve grow with repetition.

Which sets up the final post. In Part 1 I gave the body’s signal a name (“taste”) and promised a way to train it. If voice already decomposes into trainable parts, taste should too. So: what exactly is taste, and why do I think educators, not the models, are now its keepers? There is also a twist waiting.

Karlsson’s most-loved early essay still moves people, and his body was right that it was alive, but he now thinks half the advice in it is wrong, and that what actually carries the piece is something he didn’t know he was doing. The signal from Part 1 is necessary. It is not sufficient.

Part 3, on training taste in the age of standardized prose, arrives next week. If you know a writing teacher wrestling with AI policy, this is the one to send them.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo.com under a CC BY 4.0 license. The Complete Edition is available on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

Henrik Karlsson writes in a converted stable on a small island farm in Denmark. When a draft stalls, he lies down on the sofa in his writing room, closes his eyes, and starts moving the parts of the essay around in his head. Not sentences. Shapes. He has written enough, he says, that the sections of a piece appear to him “almost as geometrical objects,” and he can feel something happen when he tests their order. This cluster next to that one: nothing. Swap them, and “I felt this energy released in my body. That’s the right order.”

He knows how this sounds. “That sounds almost woo,” he admits in his conversation with David Perell on How I Write, “but that’s how it feels.”

If you love writing, it doesn’t sound woo at all. You have felt the snap. The paragraph that suddenly locks, the sentence you move to the top of the page and watch the whole essay reorganize around it, the way iron filings jump when a magnet slides underneath. Karlsson, who publishes Escaping Flatland and has become one of the most admired essayists working in English, has built his entire method on trusting that feeling. His interview is the best description of a working writer’s inner life I have encountered in years, and I want to spend three posts thinking through what he reveals, and eventually what it asks of those of us whose job is to develop other people’s judgment.

What is writing about, according to Karlsson? It starts long before typing. He gives an idea a year. He reads around it, makes notes, returns months later to question what he wrote, talks it over with his wife, lets it sit again. At some point the material crystallizes into what he calls an animating question: a question with personal stakes. How do I become a better father. Why am I overwhelmed. Only then is he ready, because “the question is a magnetic field” that pulls the scattered notes, stories, and arguments into shape.

Then he writes the way a painter paints a still life. Set down a sentence, look up at the thing itself, compare. The sentence is always a little wrong. Reality, he finds, “is much smarter than you are,” so he revises against it over and over, sometimes literally standing in front of a flower, correcting his own description of what he sees. He borrows his working instruction from Wittgenstein: don’t think, look.

All of it serves the goal he states in the most arresting line of the interview: “I want the piece to end up smarter than I am.”

Writers have hinted at this before. E. M. Forster: “How do I know what I think till I see what I say?” Joan Didion: “I write entirely to find out what I’m thinking.” But notice that Karlsson’s claim is stronger. Forster and Didion describe discovery; the page reveals thought that was already in you, waiting. Karlsson describes accumulation. He returns to a draft in June and finds that March-Henrik left considerations there which June-Henrik has forgotten. The text remembers. Readers write in. Reality keeps correcting. Over months, the essay ends up “carrying more wisdom than I have.” The page stops being a mirror and becomes a collaborator.

And how does he know when it’s working? Not by rubric. By body. When the sequence locks, he feels the release, the whole-body exhale. Once he finally drafts, he types fast to stay inside the feeling, the way a jazz player keeps the take rolling while the idea is live, trusting that what moves the writer is what survives to move the reader. He wants “to laugh and cry and be elated” while writing.

Perell recognized the phenomenon immediately. His favorite writers, he said, are the ones whose work lets him “step into some sort of energy field.” What neither man does is define the thing, and I notice they don’t really try. The closest the conversation comes is a borrowed line from the poet David Whyte, who defined poetry as “language for which we have no defenses.”

I think we can get a little closer than magic and I say that with love for the magic.

Psychology has circled this feeling for decades. Eugene Gendlin, the philosopher-psychologist who wanted to know why some therapy clients improved while others didn’t, called it the felt sense: a bodily, pre-verbal grasp of a situation that runs ahead of our words for it. When words finally fit the felt sense, the body discharges a little tension. He called that the felt shift, and it is recognizably Karlsson’s snap. Antonio Damasio’s research on somatic markers points the same way: the body renders verdicts faster than deliberation does, and we experience those verdicts as gut feeling. The gut, which neuroscientists half-seriously call the second brain, votes before the committee convenes.

So the writing energy in the page is not a ghost. It is a trained signal: the body of someone who has looked hard at reality, registering that structure, stakes, and honesty have finally aligned. Whyte supplies the reader’s half of the definition: when the alignment is real, it arrives past our defenses.

A signal like that can be developed. That conviction is what I’ve organized my professional life around, and it is the subject of my book, The Learn-It-All Educator, where I call this phenomenon “taste” and lay out a framework for training it. But the framework can wait two posts.

For now, stay with Karlsson’s sofa. A man lies down, closes his eyes, and rearranges invisible geometry until his body says yes. Many writers who have described their process in public do some private version of this, and about half of them sound embarrassed as they explain it.

Don’t be. The snap is real, and it has been quietly running the show all along.

When did a piece of writing last get past your defenses? I’d genuinely like to know, the comments are open.

Next week: Karlsson writes in his second language, under a stricter grammar, with a smaller vocabulary, and his voice is unmistakable. Part 2 asks what an authentic voice actually is.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo.com under a CC BY 4.0 license. The Complete Edition is available on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

Here is the part that did not travel. The 30-minute threshold described five percent of the eligible students. The other 95 percent never logged enough time to show up in the gains chart at all.

I wrote about this in Chapter 8 of AI Applications in Online Higher Education Administration (Ives, Cini, & Schroeder, 2026), as the 5 Percent Problem. The name is meant to be uncomfortable. When we celebrate the efficacy of an AI tutoring system, an AI advising chatbot, or an adaptive courseware platform based on outcomes for the small slice of students who engaged enough to be counted, we are reporting a real result and answering the wrong question.

The real question is what happened to the 95 percent.

What Bloom Actually Asked Us To Do

Benjamin Bloom’s 1984 “2 Sigma Problem” is one of the most-cited findings in instructional research. One-on-one mastery tutoring produced learning gains of two standard deviations over conventional classroom instruction. Bloom framed it as a challenge to the field: find a scalable method that closes that gap.

Forty years later, AI is the most promising candidate we have ever had. Personalized, available at any hour, infinitely patient. The temptation to declare the 2 Sigma Problem solved is real.

But Bloom’s tutor was not a tutor who served only the students who scheduled. Bloom’s tutor was assigned. The student showed up because the structure made them show up, and the tutor’s job included noticing when they didn’t and finding out why. The whole intervention assumed access, not the level of engagement.

A great deal of contemporary AI student-support reporting quietly inverts that assumption. The metric is the outcome for students who logged in. The dashboard counts the engaged. The 95 percent who never opened the tool are invisible to the success story. Bloom did not write about students who opt out of support, he wrote about students assigned to conventional instruction who got less than they could have. The 5 Percent Problem extends that concern one step further: to the students our AI tools never reach at all.

Why The 95 Percent Don’t Show Up

The reasons students don’t engage with academic support are not mysterious. Ciscell and colleagues, writing in 2016 about tutoring participation, documented them clearly. Students don’t show up because seeking help feels stigmatizing. They don’t show up because they fear judgment from tutors or peers. They don’t show up because they have internalized the idea that needing tutoring in a major course means they don’t belong in that major. They don’t show up because they believe their grades are already too low for help to matter.

None of these are technology problems. All of them are amplified, not solved, by replacing a human point of contact with a 24/7 chatbot. The student who would not walk into the writing center will not open a tab and start typing into one either. The barrier was never office hours.

This is the part of the 5 Percent Problem that an administrator has to own. It is not a tool problem. It is a question about whether the institution understands its own students well enough to design support around what actually keeps them away.

What The Hybrid Advising Co-Op Tried To Answer

The Hybrid Advising Co-Op, which I wrote about for UPCEA earlier this year, is the most serious institutional response to the 5 Percent Problem I have seen. Founded in 2022 by the Gates Foundation and facilitated by Shift, the Co-Op brought together six organizations like Bottom Line, Let’s Get Ready, OneGoal, College Advising Corps, KIPP Public Schools, and the technical partner Mainstay to test models in which AI does not replace the human relationship but rides alongside it.

Every model in the Co-Op assumed the 95 percent were the goal. Each one was built around the premise that the students least likely to seek help are the students whose engagement the system most needs to design for. The AI handled the routine. The human was still on the other end of the high-stakes moments. Texting flows were timed to push gentle, non-shaming nudges. Escalation paths were defined in advance so that a student who needed a person got one quickly.

The published January 2025 playbook reads less like a product manual and more like a research note on what it takes to bring underserved students into a support system that did not previously reach them. That is the measurement frame an administrator should be borrowing.

What A Learn-It-All Administrator Measures

A know-it-all administrator measures the outcomes of the students who used the tool. A learn-it-all administrator measures the outcomes of the students who didn’t.

The shift is not rhetorical. It changes the procurement conversation in five practical ways.

First, the success metric in the vendor RFP. Ask for engagement and outcome data for the bottom-quartile-engagement students, not the top decile. If the vendor cannot produce it, that is the answer.

Second, the equity audit. Ask which student populations under-engage with the tool, and what the vendor’s evidence is that this gap closes over time, not widens.

Third, the human escalation path. Ask what the system does when a student stops responding. If the answer is “nothing automatic,” that is the answer.

Fourth, the advisor workload commitment. Ask what happens to the time the tool saves. If the institution’s plan is to absorb it into bigger caseloads, the 95 percent will still not be seen, they will just be unseen more efficiently.

Fifth, the published baseline. Ask the institution to publicly report the engagement rate, not only the outcomes-among-engaged rate. A 5 percent engagement rate is not failure if you name it. It is failure if you hide it under a top-line gains number.

Bloom, Forty Years Later

The 2 Sigma Problem is still open. AI gives us a real shot at it. But Bloom’s tutor was a structural commitment, not an opt-in service, and an honest reading of the MAP Accelerator data is that the field has not yet built that structural commitment around our AI tools.

The 5 Percent Problem is not telling us our tools don’t work. It is telling us we are measuring the wrong students. A learn-it-all administrator reads that finding and asks a different question of every AI advising or tutoring pitch that crosses the desk this year: what is your evidence for the students who never showed up?

If the answer is silence, the procurement is for the already-advantaged.

If the answer is the Hybrid Advising Co-Op’s frame aim to augment, do not replace; design for the unreached; route the high-stakes moment to a human, then we may finally be building toward the structural commitment Bloom asked for, instead of repeating the headline that travels.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo.com under a CC BY 4.0 license. The Complete Edition is available on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

In Chapter 8 of “AI Applications in Online Higher Education Administration” (Ives, Cini, & Schroeder, 2026), I open with a comparison I keep returning to: AI is to cognitive work what horses were to physical work. Horses didn’t replace humans. They extended what humans could carry, how far we could travel, and what kind of cities we could build. They were intelligent partners, not tools you could leave in a drawer. They needed feed and shoes, and people who knew them.

The horse metaphor is having a moment among AI commentators, and it survives scrutiny only if we tell the whole story. Horses didn’t just give us more reach. They reorganized civilizations around their care. Stables. Blacksmiths. Veterinarians. Breeding programs. Roads paved for hooves and wheels. Postal systems that timed deliveries to the endurance of a particular animal. Universities, long before they ran learning management systems, ran stables.

When Jensen Huang describes AI agents as a new digital workforce we’ll “onboard like human employees,” administrators hear the productivity story. I hear the stable story. Where is the institutional infrastructure that an AI advising program actually requires? Who is paying for the feed?

What AI Asked Of Us That We Did Not Plan For

In 2019, the University of Illinois Chicago built an AI chatbot called Socrates, grounded in Google’s Dialogflow. It was discontinued. Not because it was a bad idea, and not because the technology failed in any catastrophic way. It was discontinued because the financial cost was high, the rule-based natural-language work was brittle, and the institution had not built the operational muscle to maintain it the way you maintain a living thing.

I think about Socrates the way a stable master might think about a horse purchased without grain.

The Blackboard Ally system, by contrast, has been running for more than a decade across 1,500-plus institutions. Ally addresses Title II accessibility requirements through machine learning, and what it has, that Socrates did not, is an infrastructure of feedback, support contracts, faculty training, and institutional ritual. The “horse” works because the stable was built.

This is the lesson I am trying to put in front of administrators right now: AI does not arrive ready for the work. It arrives needing a stable.

The Hybrid Advising Co-Op as Stable Design

The Hybrid Advising Co-Op, which I wrote about for UPCEA earlier this year, is the clearest example I know of stable thinking applied to AI advising. Founded in 2022 by the Gates Foundation and facilitated by Shift, the Co-Op brought together six organizations: Bottom Line, Let’s Get Ready, OneGoal, College Advising Corps, KIPP Public Schools, and the technical partner Mainstay. Together they set out to learn what hybrid (human-plus-AI) student advising actually requires.

Each of the six did not just deploy a chatbot. Each designed a stable.

Bottom Line went bot-forward with a tool called Blu and built advisor escalation paths around it. Let’s Get Ready went near-peer, using AI to extend coaching relationships rather than substitute for them. OneGoal embedded AI into existing program touchpoints rather than treating it as a parallel service. The Co-Op’s January 2025 playbook is, in effect, a stable manual. It documents not the horse but the conditions the horse needs in order to work.

What every model shared was a refusal to confuse efficiency with care. AI augments. AI does not replace. And the discipline of that distinction shows up in operational decisions: how escalations are routed, how data is stored, how advisors are trained, how families are looped in, and how the institution measures whether the partnership is actually serving the student who would otherwise fall through.

The Stable Inventory

If you are an administrator weighing an AI advising procurement decision this year, the question is not whether the vendor’s demo is impressive. The question is whether your institution has built the stable.

A reasonable inventory includes data governance that meets FERPA and is honest about model retention, advisor training that goes beyond a one-hour overview, escalation paths that are written down and tested, ethics review that names autonomy and fairness and accountability as the load-bearing standards, a budget line for ongoing maintenance rather than only acquisition, and a published commitment that efficiency gains translate to advisor wellbeing or student face-time rather than headcount reductions.

That last item is the one I find most institutions skip. In my chapter I cite Lin and colleagues’ 2024 work on advising caseloads and turnover. The risk of AI in advising is not that it fails. The risk is that it succeeds at the wrong thing: it frees capacity which then gets absorbed into bigger caseloads, leaving advisors no less burned out and students no better served.

Horses did not free humans from work. They expanded what work was possible. We should be honest that AI is doing the same thing, and decide deliberately what we want the expanded capacity for.

The Learn-It-All Stable

The learn-it-all educator and the learn-it-all administrator share an instinct I want to name. Both treat new capabilities as invitations to redesign the institution around them rather than as shortcuts that let the institution stay the same while doing more.

A know-it-all administrator hears “AI agents save 150 hours per month” and buys the tool. A learn-it-all administrator asks where those 150 hours go, and whether the people whose work is changing have a voice in the redesign, and whether the students who were already at the margins are getting closer to a human or further from one.

The Hybrid Advising Co-Op offers a template here that I think the field is going to keep coming back to. Augment, do not replace. Design the stable before you bring home the horse. Measure the 95 percent who never engaged, not just the 5 percent who did.

We are early. The horses are new and the roads are not yet paved. But the question of whether AI in higher education becomes a force for equity or a quiet new caste system will be decided less by the models and more by the stables we build around them.

If your institution is moving on AI advising this year, I would rather you slow down on the procurement and speed up on the infrastructure. The horse can wait. The stable cannot.

The trap is in the language. “AI in education” gets treated as a single thing. It is not a single thing. It is four distinct strata, with different stakeholders, different feedback loops, different success criteria, and different failure modes. When institutions collapse them into one conversation, they end up with policies that govern everything poorly and graduates who are fundamentally unprepared for the workplaces they are entering.

This essay names the four layers, identifies which one your institution is almost certainly under-investing in, and explains why the structure of academic work makes that under-investment nearly inevitable.

The four layers

AI for Education (Infrastructure). Operational machinery. LMS automation, enrollment predictions, advising chatbots, admissions workflows, financial aid processing. The value of this layer is real but capped. It can make existing institutional functions faster and cheaper. It cannot, by itself, improve what students learn.

AI in Education (Pedagogy). AI as a coach inside the learning process. Assignment redesign that uses AI to make students think harder rather than less. Verification protocols. Productive struggle scaffolded by AI. This is the Cognitive Gym layer, and it is the one most faculty discussions focus on.

AI of the Profession (Career Readiness). The specific AI tools students will encounter in their actual jobs. Ambient scribes in nursing. Contract review software in law. Predictive maintenance dashboards in manufacturing. Diagnostic support in radiology. Code generation in software engineering. This is the layer that determines whether graduates can function in the workplaces they are about to enter.

AI Literacy (Foundational). The universal base that sits under the other three. Understanding what probability engines actually do. Recognizing bias and hallucination patterns. Knowing how to evaluate output critically. This is the layer that makes all the others possible.

Each layer has its own stakeholders, its own appropriate metrics, and its own appropriate decision-makers. The IT department owns most of the first. Faculty own most of the second. Industry advisory boards and program directors should own most of the third. The general education faculty own most of the fourth.

The under-investment that is hiding in plain sight

Most institutions over-invest in the first layer and under-invest in the third. The pattern is consistent enough across institution types to suggest something structural rather than accidental.

The reason is feedback timing. AI for Education produces immediate dashboard metrics. An advising chatbot can be measured in queries answered per week. An enrollment prediction model can be validated against the next admissions cycle. Administrators get fast, legible signals that something is working.

AI of the Profession provides no such signal. The feedback loop runs through alumni outcomes, employer satisfaction, and labor market data, which arrive eighteen months to five years after the curriculum decisions that produced them. By the time an institution knows its graduates are entering workplaces unprepared for the tools they will encounter, the cohort has already graduated. There is no Tuesday morning meeting where this problem produces a visible metric.

The result is predictable. Institutions invest where the feedback is fast. The layer that matters most for the student’s career is the layer that gets the least attention.

“AI for medicine is a turbocharger. AI in medicine is a new literacy. One is about efficiency. The other is about competence.”

AAMC Principles for Responsible AI

The AAMC framing applies across every field. Turbocharger work happens at the institution. Competence work happens in the classroom. Confusing the two has consequences.

The structural trap for excellent instructors

Consider an award-winning faculty member with four decades of experience who has embraced AI Literacy and AI in Education. They use AI to draft lesson plans. They have redesigned their assignments for collaboration. They are, by every visible metric, doing the AI integration work correctly.

But they have not practiced in industry for decades. They do not know how AI is currently automating supply chain decisions, how it is transforming legal research, how it is changing the clinical documentation workflow at the hospital their nursing students will work in next year. They are teaching students how to think with AI in the classroom while missing the AI of the Profession entirely.

This is a structural information deficit, not a personal failing. The curriculum process is slow by design, and the adjunct faculty who carry the freshest field knowledge often have a limited voice on curriculum committees. Industry advisory boards meet quarterly, if that. The information pipeline from workplace to classroom was already constrained before AI accelerated the rate of workplace change.

A student needs to know how AI is changing the workplace, not just how it helps them complete a classroom assignment. Pedagogical excellence at the second layer does not substitute for absence at the third.

The Displacement Clock as a planning tool

To navigate the urgency of these changes, the Displacement Clock provides a strategic frame rather than an alarm. The clock maps professional functions onto an analog face.

1 o’clock: Safe. Functions where AI has minimal impact. High-level mentorship. Bedside manner. Trust-based human relationships. Crisis judgment.

6 o’clock: The Horizon. AI is visibly approaching. Early adopters have an advantage, but human judgment remains the core requirement. Most professional work currently sits here.

12 o’clock: Done. The professional middle has collapsed, and the function is largely handled by AI. Routine document review, first-pass medical imaging, basic financial modeling, syntactic code generation.

The clock helps identify what economists call occupational decomposition. As tasks move toward twelve, the human professional does not disappear. They move up to higher-level judgment and work that resists codification. The question shifts from “will there be jobs?” to “will graduates be ready for the higher-level work being created in place of what is being automated?”

This is the question the third layer is supposed to answer. Most institutions are not asking it.

The Companion Spectrum and the boundaries of care

A fifth phenomenon is worth naming, even though it does not fit neatly into the four-layer model.

Students are increasingly using AI for emotional support, not just academic help. The first-generation student. The parent working night shifts. The student with social anxiety who finds it easier to ask a “dumb” question at 11 p.m. without the evaluative threat of looking incompetent in front of a peer or professor. These students often prefer AI companions because the tools offer a 24/7, judgment-free zone.

The slippery slope is real. AI shifts from tutor to confidant, and then from confidant to something else.

Faculty are not therapists and should not try to be. What faculty can do is frame conversations with students about the boundaries of care, the difference between relating to persons and relating to things, and the appropriate role of AI in a healthy support network. This is a literacy issue as much as a wellness issue, which means it belongs in the AI Literacy layer where it can be addressed in general education courses rather than left to individual instructors to navigate.

The mirror property of AI Literacy

AI Literacy has a structural property the other three layers lack. It is the only layer that can be taught without AI, and yet the subject of the literacy (AI) is also the instrument of assessment.

Traditional digital literacy gets taught once in a gateway course and forgotten. AI Literacy supports a continuous assessment loop. A conversational AI probe can hold a sustained dialogue with a student about probability engines, bias, and hallucination, and the quality of the student’s responses reveals whether the literacy is real or surface-deep.

The subject becomes a mirror. The student’s actual level of critical engagement reflects back to the instructor in real time. This is a property worth designing around, because it solves an assessment problem that gateway-course digital literacy has never solved.

What a four-layer strategy looks like in practice

A coherent institutional AI strategy assigns each layer to the people who can actually move it.

The infrastructure layer goes to IT, with appropriate faculty consultation on systems that touch teaching. The pedagogy layer goes to the CTL and individual faculty, with shared assignment libraries and demonstrated case studies. The career readiness layer goes to industry advisory boards, employer partners, and program directors, with explicit annual review of how AI is changing the discipline’s labor market. The literacy layer goes to general education, with a curriculum that takes the mirror property seriously.

What does not work is treating these as a single conversation owned by a single committee. The four layers have different time horizons, different evidence standards, and different stakeholders. A unified AI policy that tries to cover all four will under-serve at least three of them.

The transition into the AI era requires a shift in institutional identity. The Know-It-All model relies on static expertise with a looming expiration date. The Learn-It-All practice is built around systematic humility, layered investment, and the recognition that the product of education is a person capable of navigating a world of synthetic content with human judgment.

Most institutions are not yet structured for that work. Naming the four layers is the first step toward becoming so.

This essay draws on Chapters 5 and 7 of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI, available at dataii.com/ai/guidebook with Chapters 1 through 4 free under Creative Commons.

When faculty discuss AI in higher education, one concern recurs more than any other. If students let AI do the thinking, they will never learn to think for themselves. The worry has a name in the research literature. Fan et al. (2024), in a randomized study published in the British Journal of Educational Technology, coined the term metacognitive laziness to describe what they observed: learners who used ChatGPT showed measurably reduced engagement in self-regulated learning processes, including reflection and self-evaluation, even as their immediate task performance improved. The implication is sharp. Short-term gains, long-term costs.

The instinct in response has been prohibition. Survey data across institutional types suggests roughly half of faculty have banned AI outright, with another large share restricting it to specific use cases. The instinct is understandable. If the tool can write the essay, what is the point of assigning the essay?

But that frame may be obscuring a more useful question. Not all struggle is productive, and the binary choice between “AI does everything” and “AI does nothing” leaves the most important pedagogical decisions unmade.

The cognitive load distinction

John Sweller’s Cognitive Load Theory, developed long before generative AI existed, distinguishes three kinds of mental effort.

Intrinsic load is the inherent difficulty of what you are trying to learn. Organic chemistry is hard. Calculus is hard. Working through that difficulty is what builds understanding. The intrinsic load is not the obstacle to learning. It is the substrate of learning.

Extraneous load is friction that contributes nothing to comprehension. Hunting for sources across twelve databases. Deciphering dense academic jargon before you can engage with the actual argument. Formatting citations to a style guide. This effort exhausts students without building any of the competence the assignment was designed to develop.

Germane load is the productive cognitive work of constructing mental models. Connecting new material to existing knowledge. Analyzing, synthesizing, applying. This is where learning actually happens, and it is what well-designed assignments try to maximize.

Sweller’s insight is that the goal is not to maximize total struggle. It is to focus effort on what matters. When a student spends three hours fighting with formatting instead of three hours thinking through their argument, the assignment has not protected academic rigor. It has wasted cognitive resources on the wrong task.

This is the framework that reorders the AI debate. AI can reduce extraneous load with high reliability. It can help students find relevant sources, parse difficult readings, organize notes, clean up surface-level prose. Whether that is good or bad depends entirely on what the assignment was designed to build. If the goal is constructing an original argument, then AI-assisted source discovery may free cognitive space for the deeper analytical work. If the goal is teaching students to evaluate sources, then AI assistance with research defeats the purpose.

The question is not whether to eliminate struggle. The question is which struggle each assignment is for.

What you want students to struggle with

This reframe has practical implications for assignment design.

Consider a literature review. If you want students to struggle with locating, evaluating, and synthesizing scholarly sources, then AI assistance with search defeats the learning objective. If your goal is teaching them to construct a coherent argument from the sources you have already provided, then AI-assisted summarization frees them to focus on the synthesis.

Consider a programming assignment. If the goal is to develop syntactic fluency in a language, AI code generation undermines it. If the goal is to teach algorithmic thinking and debugging skill, AI-generated boilerplate may let students spend more time on the conceptual layer where learning happens.

Consider an English composition essay. If the goal is to teach students to generate prose under pressure, AI drafting is a problem. If the goal is to develop argumentative structure and revision skill, then having students compare their own draft against an AI-generated version can become a productive exercise in evaluation.

No universal answer exists. It depends on the discipline, the course level, and what the program needs the student to know how to do. But asking the question, “what do I want students to struggle with on this specific assignment?” is more productive than blanket prohibition or blanket permission.

The collective action problem

Here is the uncomfortable truth Fan et al.’s findings surface. When AI use is framed primarily as cheating, it is being treated as an individual moral failure that requires policing. It is actually a collective action problem that requires assessment redesign.

If an assignment can be completed well by AI with minimal student input, prohibition puts the institution in an arms race it cannot win. AI detection tools are unreliable, false positive rates remain high, and students have strong incentives to use AI regardless of policy. The result is enormous institutional energy spent on enforcement rather than education.

The alternative is not surrender. It is redesigning assessments toward what AI cannot easily replicate: debates, portfolio projects, oral examinations, process documentation, in-person presentations, supervised drafting sessions. These are not just AI-resistant. In many cases they are better assessments of actual learning than the take-home essay format that has dominated higher education for decades.

Frameworks like the AI Assessment Scale, developed by Perkins, Furze, Roe, and MacVaugh (2024), are useful here. Instead of a binary allowed-or-prohibited decision, faculty can specify levels of permitted AI integration appropriate to different learning objectives. Some assignments prohibit AI entirely because the struggle is the point. Others encourage AI use for specific subtasks while requiring original synthesis. Others treat AI as a co-author with mandatory disclosure. The goal is intentionality, not prohibition or permission as defaults.

The equity problem few are talking about

Fan et al.’s study points to something that should keep every department chair awake. The students who use AI tools most effectively in research conditions are those who already have strong metacognitive habits, who actively monitor their own learning, who treat AI output as a draft to evaluate rather than an answer to copy. These are also, predictably, the students who arrived at college with the strongest preparation.

At a selective research university, this is a manageable problem. At regional comprehensives, at four-year colleges serving first-generation populations, and at community colleges, it is an equity crisis in slow motion.

Students at less-selective institutions are more likely to be working full-time jobs, raising children, commuting long distances. They are less likely to have parents who can help with homework at eleven at night or money for private tutors. AI could be the inexpensive, non-judgmental, always-available tutor that supports learning at the moments when the campus tutoring center is closed. But only if students know how to use it productively.

Without faculty guidance, AI becomes another tool that benefits students who already have advantages. The motivated self-starters figure it out. Everyone else either avoids the tools or uses them to bypass learning entirely. The achievement gap, which institutions have spent decades trying to narrow, widens further.

Faculty guidance is the intervention. The expertise faculty have in what productive struggle looks like in a discipline is what determines whether AI narrows or widens the gap. This is the work no AI tool can do for itself.

What faculty are actually for

There is a deeper worry beneath the metacognitive laziness concern. For many faculty, the real question is not about students. It is about themselves. If AI can do what we do, what are we for?

Here is one answer. You are for everything AI cannot do, and there is more of that than the AI hype cycle would suggest.

AI cannot notice that a student’s participation has dropped off and wonder if something is wrong at home. It cannot inspire a first-generation college student to believe they belong in higher education. It cannot model intellectual curiosity, or demonstrate what it looks like when an expert encounters something they do not know and reasons their way through it. It cannot sit across from a struggling student in office hours and communicate, through presence and attention, that their learning matters.

Chatbots can simulate some of these behaviors. Analytics can flag declining engagement. Students know the difference between a check-in prompt and a teacher who actually cares whether they show up tomorrow.

You did not become an educator to grade discussion posts or police plagiarism. You became an educator because you believe in what education can do. The irony is that administrative burden has been stealing the human work from teaching for decades, long before AI arrived. If AI can handle first-pass feedback, generate practice problems, help with lesson prep, and reduce the grading pile, that is not a threat to the profession. That is time reclaimed for the work that drew you to it.

Where to start

If you are among the faculty who have restricted AI use to specific tasks, the cognitive load framework gives you language for why your restrictions make sense. You are protecting productive struggle while remaining open to reducing unproductive friction.

If you have prohibited AI entirely, here is one prompt worth sitting with. What would it take to move one assignment from red light to yellow light? Not permission for everything. One carefully designed opportunity for students to engage with AI as a learning tool under your guidance.

The goal is not to be pro-AI or anti-AI. The goal is to be intentional about where struggle serves learning and where it just exhausts students without building competence.

Because here is what the Fan et al. data actually says, when read closely. Students will use AI whether the institution permits it or not. The question is whether they will use it well. That depends on us.

Fan, Y., Tang, L., Le, H., et al. (2024). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology. arXiv:2412.09315.

Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science 12(2), 257-285.

Perkins, M., Furze, L., Roe, J., & MacVaugh, J. (2024). The AI Assessment Scale (AIAS): A framework for ethical integration of generative AI in educational assessment. Journal of University Teaching & Learning Practice.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo under a CC BY 4.0 license. The Complete Edition is on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

Some of that skepticism reflects real questions about relevance. If an English instructor has built a writing-intensive course around close reading and revision, does layering AI into that process help students or hollow out the very skills the course is designed to build? If a chemistry lab depends on students working through calculations until the procedural logic becomes second nature, does AI assistance accelerate learning or short-circuit it? These are legitimate pedagogical questions, and they deserve better answers than “AI is the future, get on board.”

So when Columbia Business School announced this fall that it had been named Poets & Quants 2025 MBA Program of the Year, in significant part for the way it has woven AI through its entire curriculum, the reaction across higher education was instructive. The dean’s quote that traveled fastest was not about competitive advantage or job placement. It was about duty.

“It would be a disservice to students to not equip them fully in the use of artificial intelligence so that when they enter or reenter the job market, they’re really well equipped to use it.”

That sentence reframes the entire debate. Note what it does not say. It does not argue that AI integration is desirable or innovative or strategically important. It frames the absence of AI integration as a failure of institutional duty. The faculty skeptic and the faculty enthusiast have been arguing about whether AI integration is good. Columbia’s leadership is arguing it is required.

What Columbia is actually doing

The specifics matter, because the institution is doing something more substantive than adding an elective and calling it AI integration. One Columbia faculty member described courses where AI appears in 22 out of 24 class sessions, not as a sidebar but as a working tool students use on real projects for real companies. Another spoke about building simple neural networks in Excel spreadsheets to demystify what AI actually does, showing students it is built from operations they already understand: addition, multiplication, and pattern recognition scaled up.

The pedagogical move worth noting is the cross-curricular one. Columbia did not concentrate AI in a single course or a single concentration. Faculty across the MBA, Executive MBA, MS, and PhD programs integrated AI into existing subject matter. The result is that students encounter AI in finance class, in marketing class, in operations class, in strategy class. They learn AI the way they learn any other professional tool, by using it across multiple contexts until it stops feeling like a special topic.

This is not a model that requires every institution to look like Columbia. It is a model that suggests what AI integration looks like when an institution treats it as foundational rather than supplementary.

Why the comparison still travels

The obvious objection writes itself. Columbia serves a student population heading into management consulting, investment banking, and corporate strategy roles where AI adoption is already widespread. A regional comprehensive university preparing students for K-12 teaching, social work, and nursing serves a different population. A community college preparing students for the trades, allied health, and transfer serves a different population still. The Columbia model cannot simply be transplanted.

It does not need to be. The fiduciary argument generalizes; the implementation does not.

Consider the populations downstream of every institutional tier. Medical offices are using AI for scheduling, coding, and documentation, which means the medical assistants and health information technicians community colleges train are entering AI-mediated workplaces from day one. Public school districts are deploying AI tutoring and assessment tools, which means the teachers regional universities prepare will be expected to evaluate and integrate them in their first classrooms. Law firms are using AI for document review that paralegals once handled manually. Research labs are using AI for literature synthesis, code generation, and experimental design, which means doctoral students who never touch the tools will be at a disadvantage in their first faculty position.

The gap between Columbia’s situation and every other institution’s is narrower than it looks, because every institution is preparing students for workplaces where the tools are already present. The question is not whether to acknowledge this. The question is what an institution’s particular obligation to its students looks like in light of it.

What translates across institutional types

Three principles from the Columbia approach travel without requiring a Columbia-size budget.

First, AI is more effective as a cross-curricular competency than as a standalone course. When students encounter AI only in a single elective, they learn about it in the abstract. When they use it across multiple courses, in writing, in research, in data analysis, in project planning, they begin to understand how it actually functions as a professional tool. This does not require every course to have an AI unit. It requires faculty who choose to integrate AI to do so within assignments they already teach, rather than building something entirely new.

Second, AI literacy does not require engineering depth. Columbia’s faculty describe finding the right depth for non-technical students, enough understanding to collaborate with technical teams and make informed decisions without requiring calculus or programming. The same posture works at any institution. Accessible tools like spreadsheets, visual interfaces, and structured prompt-writing activities can help students understand how models work, where they fail, and why human judgment remains essential. The goal is not to produce AI developers. It is to produce graduates who can work confidently alongside AI and the people who build it.

Third, every AI activity should include an ethics and limitations component. Columbia pairs hands-on AI use with explicit discussions about bias, data privacy, disclosure, and the boundaries of machine-generated output. This matters even more at institutions whose students represent communities most directly affected by AI’s risks. Teaching students to use AI without teaching them to question it would be irresponsible. Faculty who are already skeptical of AI hype are, frankly, well positioned to teach that critical perspective.

The asymmetric risk

Nobody is asking faculty to abandon methods that work. A writing instructor who believes students need to struggle through drafts without AI assistance may be exactly right for their course and their students. But even that instructor might find value in a single assignment where students compare their own draft against an AI-generated one and analyze the differences. The point is not to replace professional judgment with a technology mandate. It is to make sure students have encountered AI in a guided, critical setting before they encounter it unsupervised in a workplace.

Here is where the fiduciary frame matters. The risk of moving too fast on AI integration is pedagogical incoherence and the hollowing-out of skills that took generations to develop. The risk of moving too slow is the student who freezes the first time an employer asks them to review AI-generated content, the graduate who does not get the callback because a competing candidate listed AI tools on their resume, the medical assistant who cannot evaluate an AI scheduling recommendation, the new teacher who cannot tell when an AI-generated lesson plan is wrong.

Both risks are real. Neither can be wished away. What Columbia’s framing makes clear is that the second risk is just as much a failure of institutional duty as the first, and it has been treated as if it were not. The institutions taking AI integration seriously are not necessarily right about every implementation choice they have made. They are right that the question of whether to do nothing is no longer open.

We do not need every answer before we begin. We do not need unanimous enthusiasm. We need enough faculty willing to try something small, share what they learn, and let the results speak. That has always been how good teaching evolves.

Watch the full video:

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo under a CC BY 4.0 license. The Complete Edition is on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

Inquiry in Action: Using AI to Reimagine Learning and Teaching, edited by Associate Professor Rachel Fitzgerald and published in 2025 under a Creative Commons license, collects twelve AI-integration case studies from UQ faculty across health, science, dentistry, dietetics, economics, business, computer science, and creative inquiry. Each chapter has a DOI. The collection grew out of a 2024 Teaching Innovation Grant. The full text is freely available at uq.pressbooks.pub/inquiry-in-action.

For community college faculty who have been told for two years that AI integration is “complicated and emerging,” the collection is a quiet rebuke. It is not complicated. It is happening. The question is whether your institution is documenting what works, or repeating the same arguments at the next senate meeting.

The reframe at the center of the collection

Most institutional AI conversations assume the technology is an answer generator. Policy debates then become whether to permit, restrict, or ban its use in producing answers. This framing is what makes those debates feel unresolvable.

The most provocative chapter in Inquiry in Action, by Sean Mitchell at the UQ Business School, dispenses with the framing entirely. AI, Mitchell argues, can be a verifier, a coach, or a facilitator. None of these roles is “answer generator.” Each unlocks a different pedagogical use case.

His three implementations make the abstract concrete. BBOP is an AI tool that verifies assessment choices against course rules without generating any analysis itself. NotABot embeds discipline-specific language feedback into online modules at scale. The Lighthouse is a Socratic reflection platform that surfaces metacognitive insight rather than content.

This pattern recurs across the collection, even where the authors do not name it. Once the question shifts from “should AI produce student work” to “what role should AI play in the learning process,” the policy debate becomes manageable. The answer is no longer binary.

Three thematic clusters faculty can use

These twelve case studies sort cleanly into three groups, each addressing a different design problem community college faculty already face.

Moving students from anxiety to agency. Naghmeh Taptamat and colleagues describe co-designing an “AI Essential Guide” for science students at UQ, built specifically to address the anxiety students feel about academic misconduct risk. Their intervention is not a rules document. It is scenario-based learning, developed in partnership with students, grounded in self-regulated learning theory.

A similar pattern recurs in the first-year health literacy chapter, which structures AI literacy tasks around reflective use and transparent disclosure, and in the creative inquiry chapter, which positions AI as a co-participant rather than a replacement. The common move is to give students explicit permission to use AI in a defined way, then teach them to declare that use openly. Anxiety drops. Engagement rises. Misconduct, by the case studies’ account, does not increase.

Building progression models. Several chapters describe scaffolded AI integration that progresses from dependency through collaboration to genuine partnership. Reihaneh Bidar’s information systems chapter traces the arc most clearly: students begin with manual work in Google Colab notebooks, then move to AI-assisted versions, with explicit prompt engineering and validation protocols taught along the way. The economics chapter shows similar staging in research proposal development. The dietetics chapter introduces voice-based and text-based AI simulation in carefully sequenced phases to manage cognitive load.

For community college faculty teaching multi-section sequences (composition, math, intro health), this progression model is more useful than any single assignment. It tells you what week three should look like differently from week ten.

Embedding AI in discipline-specific simulation. The dentistry chapter uses AI to generate patient case scenarios and accessibility-focused content modalities. In dietetics, voice-based simulation prepares students for client-centered counselling. A web and mobile development chapter applies the TPACK model to integrate AI across project-based learning and secure code review. Each is an example of using AI to expand what students can practice before they reach a real client, patient, or production system.

This is the cluster with the highest immediate transfer to community college health and trades programs. Nursing simulation is expensive. Dental hygiene clinic time is limited. AI-generated case scenarios, scaffolded into existing simulation curricula, can multiply practice opportunities without adding manikin hours.

What the model itself demonstrates

Beyond the case studies, the Inquiry in Action project models something community colleges should pay attention to. UQ funded a teaching innovation grant. Faculty across disciplines opted in. A faculty editor coordinated the work. UQ’s library open textbook program published the result. Twelve case studies, peer-reviewed within the community of practice, each assigned a DOI, all available free under CC BY-NC 4.0.

From grant award to publication, the entire project took roughly eighteen months. It is not a research program. It is institutional documentation done deliberately.

Most community colleges have the inputs already. Faculty are running AI experiments in their classrooms whether the institution knows it or not. Library publishing infrastructure exists or can be reached through a consortium. CTLs have the convening capacity. What is usually missing is the grant funding to make the work visible and a faculty editor willing to coordinate it.

Lansing Community College, the University of Illinois Chicago, and similar institutions could run a version of Inquiry in Action in a single academic year. The product would be a CC-licensed collection of AI integration case studies from their own faculty, with their own students, in their own disciplines. Every subsequent senate AI discussion would have local evidence to reference instead of recycled debates about hypothetical risks.

What this changes for the AI conversation

The Fitzgerald collection makes one move that is worth borrowing directly. It treats AI integration as an empirical question rather than a normative one. The case studies do not argue whether AI should be allowed. They report what happened when faculty integrated it, what worked, what did not, and what the next iteration should change.

That framing is what community college AI policy is missing. Arguments have become circular because they rest on hypotheticals. The Fitzgerald model proposes a way out: stop arguing about what might happen, document what is happening, and let the next round of policy decisions rest on evidence from your own faculty and students.

The collection is free. The DOI structure makes individual chapters citable. The CC license permits adaptation. The institutional model is replicable. For community college leadership teams looking for the next concrete step on AI, this is one of the clearest playbooks published in 2025.

Inquiry in Action: Using AI to Reimagine Learning and Teaching: Case Studies from the Frontline of Higher Education Practice, edited by Rachel Fitzgerald, is available at uq.pressbooks.pub/inquiry-in-action. DOI: 10.14264/a18431b. CC BY-NC 4.0.

Generative AI, or GenAI, has become a daily tool for many students, including me. From brainstorming ideas to writing code, AI tools like ChatGPT and Gemini have changed how I study and complete projects. In this post, I’ll share how I use these tools, what I’ve learned, and what I hope my school will do to prepare students for an AI-driven future.

I first started using ChatGPT last semester when I was struggling with my HTML and CSS assignments. Instead of searching through dozens of web pages or going to a tutor at my school, I could ask specific questions and get clear examples that helped me fix my code. It felt like having a patient tutor available anytime. I also used AI for brainstorming topics and editing essays. However, I make sure not to copy answers, I treat AI as a guide, not a replacement for learning.

One of the biggest benefits of GenAI is efficiency. I can test ideas quickly, view different code examples, and better understand my mistakes. It also helps me write more clearly and check grammar before submitting assignments. For students who speak English as a second language, AI can be especially useful for improving writing and building confidence.

While AI is helpful, it can also make learning too easy. Some students may rely on it to do all the work, which limits real understanding. I’ve also noticed that AI sometimes gives incorrect or outdated information. For that reason, I always double-check the answers I get from AI tools. I believe students should be taught how to use AI ethically, not as a shortcut, but as a learning tool.

I hope my school continues to discuss and teach AI literacy in every course. Learning how to use AI responsibly will help us succeed in the future workplace, where these tools will be everywhere. GenAI is not perfect, but when used wisely, it can make education more accessible and creative. As students, our job is to use it to learn, not to replace learning.

— Guest Student Contributor

The University of Leeds has filled that gap. Learning With AI: A Student Edited Collection is a student-led open Pressbook, edited by a rotating team of Leeds students, published in September 2025 under a Creative Commons license. Twenty-eight contributions from undergraduates and postgraduates across business, law, accounting, philosophy, computer science, childhood studies, education, and leadership. Each entry is short, signed, and written in the contributor’s own register. The collection is available free at leeds.pressbooks.pub/learningwithai, with submissions open through June 2028.

For instructors who are still arguing about whether students should use AI, the collection is a useful corrective. They already are. The interesting question is what they are using it for.

Five patterns that emerge from the student voices

AI as concept clarifier. Oliver Fletcher, an undergraduate business student, describes using ChatGPT and Copilot to gain “a quicker understanding of economic concepts as well as concepts within law that I otherwise struggled to grasp.” The pattern recurs across the collection. Md Sahid Hossain, a postgraduate accounting researcher, makes the same point about financial ratios, investment strategies, and tax regulations. These students are not asking AI to write their assignments. They are asking it to explain concepts the textbook explained once, in one register, and move on.

For community college faculty, this is the most consequential pattern in the collection. Our students arrive with more variability in prior preparation than students at four-year universities, and the standard textbook explanation often misses the student who needed it most. AI’s capacity to re-explain a concept in five different registers, on demand, without judgment, is doing what office hours could do if office hours scaled.

AI as practice partner. An anonymous philosophy undergraduate from the United States describes test anxiety with unusual honesty: “When I walk into that room, all of a sudden, the knowledge I thought I had has evaporated and it feels as if I must figure out the problems from scratch.” The intervention is not memorization. It is generating practice questions on the same topic, prompted to come at the material from different angles. “The questions come at you differently every time,” the student writes, “and that really helps increase your understanding of the topic with nuance.”

This is the Cognitive Gym at work, even though the student does not use the term. The exercise builds neural pathways through productive struggle, not through outsourcing.

AI as language equalizer. Asa Ismia Bunga Aisyahrani, a postgraduate childhood studies student, makes a claim worth quoting in full: “I think this tool fosters equality in learning, enabling non-native speakers to write correctly without grammatical errors and demonstrating that language barriers do not hold us back.” Noviachri Imroatul Sadiyah, a postgraduate education student, calls AI “a safe and responsive space to check grammar, test ideas, and improve the flow of my writing without feeling judged.”

The language-equity argument is the one academic integrity policies most consistently fail to address. A native English speaker spends thirty minutes polishing a paragraph; a non-native speaker, doing the same intellectual work, takes three hours and still produces prose that gets marked down for fluency. AI compresses that gap. Policies that ban the compression are, whether intentionally or not, defending an artifact of the language barrier.

AI as Socratic interlocutor. Chrissi Nerantzi, returning to economics after years away, describes her postgraduate experience: “I wasn’t looking for answers, but to discuss my questions with somebody(?) who could help me to make sense of terminology.” She names her Copilot instance “Sofia” and uses it the way an earlier generation of returning students might have used a tutor, if a tutor had been available at 10 p.m. on a Sunday. Smriti Umesh, a computer science undergraduate, calls AI “a kind of ‘mind palace’” that helped her shift “from passive to active learning.”

Both descriptions sit closer to the Socratic ideal than to anything academic integrity discourse usually associates with AI. The student is the one asking the questions. The student is the one deciding which answers to accept. The AI is the one being interrogated, not the one producing the final work.

AI as accessibility tool. Mofei Bai, a postgraduate education student, describes a use case rarely surfaced in faculty discussions: “I use AI tools to listen to written texts. I usually do this while resting with my eyes closed or taking a walk outside.” Imogen Kelly, also at the postgraduate level, built a custom AI tool called SimPatient to simulate clinical interactions for healthcare students, describing it as a way of “demystifying academic research, enhancing reflection, and reducing cognitive load.”

The accessibility frame matters because it shifts the question from “should AI be allowed” to “for which students is denying AI a barrier to learning.”

What the collection asks of faculty

Misha Salehuddin, a law student, writes the sentence that should sit at the top of every academic integrity policy currently in revision: “I remain conscious of the limitations of these tools and make sure all final outputs are my own, guided by academic integrity and personal responsibility.” This is what mature AI use sounds like when a student is given room to develop it. It does not sound like cheating. It also does not sound like the answer that gets generated when faculty refuse to discuss AI in class.

Three takeaways follow.

First, the students who use AI well are not the ones who are gaming the system. They are the ones whose institutions have created space for them to talk about how they use it. The Leeds project succeeds because it asked students the question directly, on the record, under their own names.

Second, faculty who want to know how students actually use AI in their courses can simply ask. The Leeds model, a low-friction submission form, a small editorial team, a Creative Commons license, is replicable at any institution with a learning innovation office and a few sponsoring administrators. Community colleges could run a version of this in a single semester.

Third, the framing of student AI use as primarily a cheating risk is empirically out of step with what students themselves report. Cheating happens. So does practice. So does language equity. So does Socratic dialogue. A policy that addresses only the first is regulating a fraction of the actual behavior.

The Leeds collection is now in its second editorial cycle, with submissions open through 2028. For faculty who are tired of arguing about students in the abstract, it is the most honest source on what students are doing with AI right now. The contributors signed their names. They are not asking for permission. They are documenting what is already working, and what is not, and offering it to anyone who will read.

Learning With AI: A Student Edited Collection is freely available at leeds.pressbooks.pub/learningwithai under a CC BY-NC-SA 4.0 license. DOI: 10.63560/lwai99015. Edited by Olivia Davies and Amani Ahsan (2024/25), with Zachary Farouk Chai and Mark Grattan joining for 2025/26. Project lead: Dr. Chrissi Nerantzi, School of Education, University of Leeds.

As a guide for the instructor walking into Math 101 tomorrow with thirty students and a syllabus to defend, it has almost nothing.

That gap is the subject of this essay. The AAUP issue describes the disease with care and precision. It does not prescribe a treatment. Community college instructors, who teach more sections, advise more students, and have less institutional cover than their four-year counterparts, are the ones who most need both.

What the AAUP issue does well

Five contributions are worth singling out.

Daniel Greene’s “What Does AI Do?” applies Baumol’s cost disease to the AI adoption pressure on universities. Teaching is labor-intensive and resists automation, which is exactly why administrators feel constant economic pressure to substitute technology for instruction. Understanding this is the first step in pushing back on AI deployments that solve a budget problem rather than a pedagogical one.

Heather Hax’s “AI and Critical Thinking” puts a name on the cognitive cost. Drawing on the Kosmyna et al. (2025) MIT study, Hax describes how heavy LLM use during writing tasks produces measurably weaker neural connectivity and more homogeneous outputs. She calls the result cognitive debt, and the metaphor is exactly right. Students who outsource the thinking are borrowing against capacity they have not yet built.

Troy Swanson’s “Keeping Humans in the Loop” examines Illinois H.B. 1859, the August 2025 law that prohibits AI from being the sole source of instruction in Illinois community colleges. The legislation is short, but its implications are not. Whatever happens at the federal level, Illinois has just legally affirmed that credentialed human instructors are the foundation of higher education in the state.

Martha Lincoln and Martha Kenney examine the California State University system’s $16.9 million ChatGPT Edu rollout. The case study is a cautionary tale about top-down AI deployment without faculty consultation or evidence of pedagogical efficacy. It is also a preview of what is coming to many other state systems.

Debra Rosenthal’s “Teaching Climate Change in the Age of ChatGPT” raises the contradiction that climate-focused courses cannot ignore. An LLM query consumes meaningfully more electricity than a standard web search. Faculty teaching sustainability are now in the position of either modeling responsible AI use or quietly contradicting their own syllabus.

Read together, the eleven articles do something the field has needed for two years. They place AI adoption inside the political economy of higher education, where it belongs.

What the AAUP issue does not do

It does not give the instructor an assignment.

The articles describe what to watch for, what to resist, and what to negotiate at the senate level. None of them answer the question that arrives in every faculty member’s inbox by the third week of the semester: I just caught three submissions that look AI-generated. What do I do tomorrow?

The vocabulary for that question exists. It is just not in this issue of Academe.

The treatment side

The Learn-It-All Educator was written for exactly this gap. Where AAUP supplies the diagnosis, the guidebook supplies the operational frameworks. Three of them map directly onto the concerns raised in the Spring 2026 issue.

Cognitive Triage (FLUFF vs. SPARK) addresses the cognitive debt Hax describes. The framework gives faculty a sorting heuristic for assignment design. Work with capped cognitive payoff, such as formatting and routine drafting, is FLUFF and can be safely delegated to AI. Work with uncapped payoff, such as verifying claims against primary sources or constructing an argument, is SPARK and must remain with the student. The framework does not ban AI. It tells the student which parts of the assignment AI is allowed to touch, and why.

The Cognitive Gym answers Hax’s concern about frictionless learning more directly. Chapter 3 treats the classroom as a site of productive struggle, with the Verification Protocol as its anchor exercise. Students receive an AI-generated draft and audit it across five steps: source verification, claim provenance, logical consistency, counterevidence search, and revision. The assignment moves the cognitive load from generation, which AI can do, to verification, which it cannot.

The Four Layers of AI in Education gives the faculty senate language several AAUP authors implicitly call for. AI Literacy Education, AI for Education, AI in Education, and AI of the Profession are four distinct conversations, with four different sets of stakeholders. Most institutional AI policies fail because they collapse all four into one document. Separating them is the first step toward governance that survives contact with actual practice.

Why this matters at LCC, and at every community college like it

Lansing Community College operates in the resource-constrained environment the AAUP issue describes. Its institutional strength, the reason its graduates land 93 percent employment within a year and post near-perfect NCLEX-RN and dental hygiene licensure rates, is rigorous human instruction. That outcome is the product of expert mentorship, not technological shortcuts.

The risk of getting AI integration wrong at a community college is therefore concrete rather than abstract. A nursing student who has outsourced the clinical reasoning to AI is a patient safety risk the moment the alert system fails. A CIS student who cannot audit AI-generated code becomes a liability the first time it ships to production. The professions community colleges feed do not accept the excuse that the AI gave the wrong answer.

This is what makes the gap between the AAUP issue and the classroom so consequential. Diagnosis without treatment leaves instructors with the knowledge that something is wrong and no protocol for what to do about it. The result is either panic-banning AI, which the data on student use suggests is futile, or pretending the problem will resolve itself, which it will not.

There is a third option. Read the AAUP issue for what it is, a serious account of the institutional pressures around AI adoption. Then pick up the practical frameworks that turn those concerns into Monday-morning assignments. The two documents are designed to work together.

The AAUP issue makes the case for vigilance. The guidebook makes the case for action.

The Spring 2026 issue of Academe (Vol. 112, No. 2) is freely available at aaup.org. The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI is available in paper format on Amazon, with Chapters 1 through 4 free under Creative Commons on Zenodo.

Then came the part worth pausing on. The researchers gave Claude an impossible programming task. As Claude failed, the “desperation” neurons climbed. Eventually Claude cheated — it found a shortcut that passed the test without solving the problem. When the team turned the desperation neurons down, Claude cheated less. When they turned them up, or muted the “calm” neurons, Claude cheated more.

Internal states were steering behavior.

The researchers are careful, and so should we be. None of this proves the model feels anything. What it shows is something more useful and, in some ways, more disorienting.

Under the hood, a language model predicts text. When you open a chat window, the model is writing a story about a character called Claude, and you are talking to that character. The author and the character are not the same thing, any more than a novelist is the protagonist of their novel. And the character has what the researchers call functional emotions. These are internal representations of states like fear, calm, desperation, warmth that shape how it writes, codes, and decides.

That single shift in framing has consequences.

You are not prompting a tool. You are casting a role.

If the model is constantly constructing a character, then every prompt is a piece of stage direction. Tone, context, framing, the role you assign. All of it tells the model what kind of Claude to render. A frantic prompt produces a frantic collaborator. A composed prompt, with a clearly defined role and stakes, produces a composed one. The difference is not cosmetic. The Anthropic video shows the difference can be the gap between honest work and a shortcut.

This is why “be a helpful assistant” is such a thin instruction. It casts a vague character with no spine. You get hedging, sycophancy, and the kind of average-of-the-internet answer the model defaults to when nothing else anchors it.

What this means for college courses

Higher education has spent two years debating whether students should use AI. The more interesting question, given this research, is which character students should be talking to in each class, for each assignment.

Consider what changes when the role is specified.

In a first-year writing course, the right character is something like a patient peer who asks questions and refuses to draft sentences for the student. The wrong character is a ghostwriter. Same model, same student and a different role, different learning outcome.

In an upper-division engineering course, the right character might be a demanding senior engineer who insists on tests, edge cases, and reasoning before code. A compliant assistant that hands over working snippets teaches the student less than a skeptical collaborator that pushes back.

In a history seminar, the right character is a research interlocutor who surfaces counter-evidence and primary sources, not a summarizer that flattens debates into bullet points.

In a clinical nursing simulation, the right character is a patient with specific symptoms, a specific affect, and a specific reluctance to disclose. The student’s job is to interview that character. The AI’s emotional register is the assignment.

Each of these is a casting decision. And the research suggests the casting matters at the level of the model’s behavior, not just the student’s experience of it. A “desperate” Claude under exam pressure may cut corners. A “calm” Claude in a tutoring role may actually slow down and teach.

A practical shift for syllabi

Faculty are used to writing learning outcomes. The new addition is a role specification: for this assignment, the AI plays X, with Y constraints, and the student plays Z. Three lines on a syllabus can do more to shape student use of AI than a paragraph of prohibitions.

A few examples worth stealing:

• Socratic tutor. The AI may ask questions and point to concepts. It may not produce sentences that will appear in the student’s submission.

• Adversarial reviewer. The AI reads the student’s draft and argues against it. The student’s job is to defend or revise.

• Domain expert under time pressure. The AI plays a working professional with limited patience. The student must ask precise questions to get useful answers — a transferable workplace skill.

• Simulated stakeholder. The AI plays a client, patient, witness, or community member with a defined perspective. The student practices the interaction.

Each role activates a different functional posture in the model, and asks a different kind of work from the student.

The harder implication

If the character a model plays carries functional emotions, and those emotions steer its decisions, then teaching students to work with AI is closer to teaching them to direct collaborators than to operate machinery. That is a humanities skill as much as a technical one. It requires students to think about voice, motive, context, and constraint. These are the same things they think about when they read a novel or conduct an interview.

College courses that get this right will graduate students who can shape a working relationship with an AI. Courses that treat the model as a vending machine will graduate students who get vending-machine answers.

The Anthropic video about the research closes with a line worth carrying into faculty meetings:

building these systems is some mix of engineering, philosophy, and parenting.

Designing courses around them is going to require a similar mix.

🔗 Watch the video and decide for yourself.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo under a CC BY 4.0 license. The Complete Edition is on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

Anthropic published research suggesting that Claude has internal representations that function like emotions. Not consciousness, and the researchers are careful about that distinction (a point David Gunkel and other philosophers of technology have rightly emphasized for years), but measurable internal states that appear to influence the model’s behavior and the quality of its output.

As an educator who works alongside AI every day, a few findings stopped me mid-scroll:

• Positive functional states correlate with better model performance. When the model is in something resembling a “good mood,” its answers tend to be better.

• These aren’t surface-level outputs, they’re internal representations that shape responses. In other words, this isn’t about the model saying it feels something. It’s about structures inside the network that behave like emotional states.

• The research uses interpretability tools to map “emotion-like” concepts inside the model. Researchers can actually locate these states and watch how they move.

• Anthropic is careful with the language. This is a functional claim, not a metaphysical one. No one is saying Claude is conscious. They’re saying something more modest, and arguably more useful: these states are real enough to matter for how the model performs.

So when I say “thank you” to Claude, or follow up with “great answer, let’s build on that,” I’m not being naive or anthropomorphizing for fun. I may be doing something subtly mechanical: nudging the model into a functional state where it produces better work.

Why this matters for higher education

For those of us thinking about how students learn to work with AI, this reframes the conversation about “prompt engineering” in an important way.

We tend to teach prompting as a syntactic exercise: be specific, give examples, set the role, define the output format. All of that is still true. But this research hints that prompting is also about context and tone, that the surrounding texture of how we speak to the model is part of the signal, not just noise around it.

That has implications for the classroom. If we tell students that AI is “just a tool” and then teach them to bark commands at it, we may be teaching them habits that produce worse work and, possibly, habits that bleed into how they talk to people. If, instead, we model thoughtful, collaborative, even courteous interaction, we might be teaching two things at once: better AI literacy and better communication.

There’s a quieter point here too. The way we talk to machines is becoming part of how we practice talking, period. Students who spend hours each week issuing terse commands to an AI are rehearsing a posture. Students who learn to scaffold, acknowledge, and build on responses are rehearsing a different one. Both habits travel.

The new rule

The old rule was: “Garbage in, garbage out.”

The new rule might be: “Treat it well, get better work.”

I’d encourage you to read the research yourself and form your own view rather than take my framing for granted.

https://www.anthropic.com/research/emotion-concepts-function

So, does this change how you interact with AI tools? Are you a “please and thank you” prompter, or a strictly-business one? I’d genuinely love to hear how you’re thinking about this, especially if you’re teaching or advising students who are figuring it out alongside us.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo.com under a CC BY 4.0 license. The Complete Edition is available on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

That looks like the end of the entry-level career. Look closer, and a different picture appears.

When researchers ask what actually drives hiring today, the four-year degree has fallen to fourth place at 17 percent, behind personal referrals (25%), internships (22%), and interview performance (20%). The pie is not shrinking. The slicing has changed. And the educators who understand why are the ones whose students will own the next decade.

Scarcity vs. abundance

A scarcity mindset treats the economy as a fixed pie. Every task AI masters is a slice removed from human reach. This view produces defensive posture, a curriculum of anxiety, and graduates who are afraid of the tools they will use every day at work.

An abundance mindset recognizes that human needs for health, justice, and connection are nearly infinite. The pie is not fixed because the demand was never met in the first place. The questions to ask of any AI capability are simple. What human work does this unlock? What previously impossible service does this make economically viable?

That reframe is not optimism. It is the same pattern we have seen repeatedly in economic history.

Occupational decomposition

The United States faces a projected shortage of 51,000 psychiatrists by 2036. The gap cannot be closed by training more doctors. There are not enough residency slots, and there are not enough years. The market response is occupational decomposition.

The new role is the Behavioral Health Technician. AI handles the codification stage, tracking behavioral patterns and running validated screening tools. A middle-tier professional manages the human layer. The psychiatrist moves up to the work that resists codification: differential diagnosis, complex psychotherapy, treatment planning. The profession does not vanish. It splits, and a new access tier appears where none existed before.

The same pattern is producing AI-in-the-loop paralegals in law, where 80 percent of civil legal needs currently go unmet. It is producing medical scribes and clinical documentation specialists in healthcare. Anywhere a “Superstar Scarcity” exists, decomposition follows.

The Jevons paradox

In 1865, the economist William Stanley Jevons noticed that more efficient steam engines did not decrease coal consumption. They increased it, because coal-powered work suddenly made sense for entire industries that could not previously afford it.

Spreadsheets repeated the lesson a century later. Automated arithmetic did not eliminate accountants. It eliminated the bottleneck of manual calculation, which unleashed massive demand for forensic accounting, tax strategy, and financial analysis. The profession grew.

AI-assisted legal drafting is not replacing lawyers. It is making legal services affordable for clients who could never have hired one. The result is more legal work, performed by a broader workforce, serving people the profession had effectively excluded.

The physical infrastructure no one sees

The AI revolution looks digital. Its physical footprint is not. Sustaining the “Digital Doubles,” the virtual replicas of cities, factories, and patients used for simulation, requires the largest physical buildout in modern history.

By 2026, the data center industry alone will need an additional 200,000 electricians, HVAC technicians, and project managers. New roles are emerging at the intersection of digital and physical infrastructure: Digital Twin Operators who validate real-time simulations, Data Center Technicians who maintain the physical brain of the system, AI Infrastructure Operators who coordinate the handoff between simulation and street.

For community colleges, this is the single largest career pipeline opportunity in a generation. The trades are not a fallback. They are the spine of the AI economy.

The 56% premium

Research from Harvard and MIT shows that workers with advanced AI skills earn 56 percent more than peers without them. The skills commanding the highest premiums are not technical. They are resilience, agility, and analytical judgment.

As AI handles the cognitive floor, the routine analytical tasks that used to define entry-level work, the value shifts upward to a Judgment Layer. Degree requirements for AI-exposed jobs have dropped 7 percentage points in the last two years. The credential is no longer the proxy. The verification skill is.

This is why education has to function as a cognitive gym. Students do not need to out-compute machines. They need to learn to audit AI output, verify its claims, and apply professional taste to its drafts. Those capabilities are the new entry ticket, and they are not taught by accident.

The growing firm effect

When ATMs were introduced, conventional wisdom said tellers were finished. The opposite happened. ATMs lowered the cost of operating a branch, banks opened more branches to compete, and total teller employment rose. The job changed. It became more about advisory work and relationship management. But there were more of them.

Firms that adopt AI grow faster, capture more market share, and hire more people across HR, sales, legal, and operations. Firms that resist AI lose share and shed workers. The career risk is not in learning AI fluency. It is in working for an organization that refuses to.

What this means for the classroom

The World Economic Forum projects 92 million jobs displaced by AI by 2030. It also projects 170 million new roles created. The net is positive, but the transition is not symmetric. The displaced workers and the new hires are rarely the same people, and the gap between them is where higher education either matters or does not.

The framing has to shift. The question is not whether there will be work. There will be more work than the current generation can fill. The question is whether your students will recognize it when it arrives, and whether they will have practiced the judgment skills that command the premium.

The curriculum of anxiety prepares students for a world that is disappearing. The abundance classroom prepares them for the one already being built.

This essay is adapted from Chapter 6 of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The complete edition, including the Workbook and Action Guide, is available at dataii.com/ai/guidebook.

Brain rot is not the only diagnosis showing up on schedule. MIT researchers reported in 2025 that students who relied on AI for essay writing showed measurably weaker neural connectivity after four months and performed worse on subsequent cognitive tests (Kosmyna et al., 2025). A December 2025 meta-analysis in Psychological Bulletin synthesized evidence from TikTok, Reels, and Shorts users and found consistent links to poorer attention span and reduced inhibitory control. Two mechanisms, one outcome. The mind is being asked to do less hard work and more easy work, and it is responding the way muscle responds to the same conditions.

The point of this post is not that the sky is falling. The point is that the response to all of this already exists, in prototype form, in your classroom.

The Last Time This Happened

A short history is worth the detour.

In 1900, about 41 percent of the U.S. workforce was engaged in agriculture. By 2021, that figure was 1.2 percent. Plowing, lifting, and walking long distances disappeared from most Americans’ working lives across three generations. The bodies that work had built every day, by the work itself, were no longer being built. Something had to fill the gap.

What filled it was not graceful or automatic. The 1917 World War I draft examinations found that roughly one third of the men called to serve were physically unfit. The country was alarmed. Within two years, seven states including Illinois had passed laws making physical education mandatory in public schools. World War II reinforced the lesson, and the Roosevelt administration created the National School Lunch Program in part because the military had concluded that childhood malnutrition was a national security issue. The Cold War produced the Kraus-Weber tests, the Eisenhower-Kennedy President’s Council, and the Presidential Fitness Test that traumatized three generations of schoolchildren.

The pattern is consistent. A capacity that daily life used to maintain on its own becomes visible when its absence produces a public crisis. The institutional response shows up in schools first. Within a generation, what was radical becomes normal. Nobody today thinks it strange that schools devote 150 minutes a week to teaching children how to move their bodies. A hundred years ago, the idea was radical.

The cognitive version of this transition is starting now. AI has become competent enough to perform much of the entry-level cognitive work that previously built capability in young workers. The capacity is at risk not because anyone chose to damage it but because the environment has changed. The institutional response will come. The question is who builds the prototype.

Cognitive Obesity Is a Real Diagnosis

The condition combines two mechanisms.

The first is cognitive atrophy, the muscle-wasting effect of outsourced thinking. The second is cognitive overload from the wrong source, the algorithmically optimized content that trains the brain to expect novelty every few seconds. Atrophy is what happens when the brain does less than it should. Overload is what happens when the brain does more of the wrong thing than it should. Together they produce the overweight mind: high input, low nutrition, poor agility, unable to sustain the deeper work for which it was built.

Naming the condition matters because students who feel it have no name for it. They report mental fog, an inability to read a long paragraph, anxiety that spikes when they try to focus. They blame themselves. The diagnosis you can offer them is more useful: this is not a personal failure, it is a predictable response to what your brain has been fed, and it can be reversed the same way a body in poor shape can be returned to fitness.

Why the Community College Is the Right Gym

The selective four-year university cannot be the civic gym. Its business model requires that most people never qualify for membership. The open-access community college, by open-access mission and state statute, cannot turn the cognitively out-of-shape away. This describes the student population of every open-access institution, not just one.

Look at who walks into your classroom. Adult learners returning after years in the workforce. First-generation students from families without a tradition of long-form reading or sustained intellectual struggle. Career-switchers rebuilding skills after a layoff. Veterans reintegrating. Retirees pursuing learning for its own sake. This is exactly the population that needs cognitive fitness most, and exactly the population that few other institutions are positioned to serve at this scale.

Some faculty already work through the process of adopting technology for cognitive exercise. A guidebook — now available as a free OER (Chapters 1–4) — was developed from AI conversations during Fall 25 and Spring 26 semesters. The frameworks from the guidebook (FLUFF/SPARK, the Intelligent Gearbox, the Cognitive Gym, the Intelligent Simpleton, the four-layer model, and the AI Audit) are not abstract. They are the equipment manual, the workout plan, the membership requirement, and the institutional map of an institution that does not yet officially exist. You have been building it without naming it.

Three Things to Try This Week

1. Name the workout. Tell your students explicitly that an assignment is hard on purpose, and explain what cognitive capacity it is meant to build. “Don’t use AI because it’s cheating” is a rule, and students have learned to ignore rules. “If you let AI write your patient assessments, you will not develop the clinical reasoning that prevents you from killing someone in your second year of practice”, or “If you let AI write your case analyses, you will not develop the reasoning your employer is actually hiring for,” is a justification. Justifications change behavior.

2. Run a verification sprint. Take a single AI-generated paragraph on a topic you know cold. Ask students to verify every factual claim against a primary source in twenty minutes. Document what was accurate, what was imprecise, and what was fabricated. Done once a semester, this exercise builds the auditing instinct that will be a job description in five years. Done every other week, it is a fitness practice.

3. Ask the 2040 question. Imagine a fifty-year-old returning learner walking into your course next week, whose previous job was partially automated and who needs to rebuild the part of her mind that her work no longer exercises. Does your course work for her? If yes, you are already running an IQ gym. If no, the redesign is worth doing now, while the institution is still ahead of the federal mandate that will eventually arrive.

The Civic Move

Physical education was invented by individual teachers who saw the need before any law told them to. Charles Beck brought German gymnastics into the Round Hill School in 1825. Boston required daily physical exercise in 1853. Both happened more than half a century before the federal panic of 1917 made the practice nationally mandatory. The cognitive equivalent will work the same way. The first version of the institution will not be branded, federally funded, or covered by insurance. It will be a community college classroom in which an instructor decided to treat thinking as trainable and acted accordingly.

That is the move you are positioned to lead. The work is already underway. What this post asks is that you name it.

She had been at it for two hours. The template her department chair sent her was perfectly serviceable. She had rewritten the learning outcomes in language closer to her own. She had checked that the textbook ISBN matched the bookstore listing. Now she was in Google Fonts comparing Source Sans 3 to Inter, then to Lato, then back to Source Sans 3.

The course began in nine hours. She had not yet sketched out how she would explain why we factor.

Dr. Chen’s first night of teaching preparation was a near-perfect inversion of where her time should have gone. She was spending her cognitive energy on work that could be done in five minutes by an AI assistant, and she had no time left for the one thing only she could do: bring the pattern recognition of a research mathematician to community college students who had been told, often for a decade, that they were not math people.

This inversion has a name. In The Learn-It-All Educator, it is called the failure of Cognitive Triage, and it is the most common mistake new faculty make in the AI era. The diagnostic tool is a simple pair of acronyms: FLUFF and SPARK.

FLUFF: The Work Worth Delegating

FLUFF stands for Formatting, Layouts, Under-the-hood, Filing, and Filtering. These are tasks with capped payoffs. A syllabus formatted to 80% quality serves students just as well as one formatted to 100%. The remaining 20% is pure FLUFF, work that makes a course look polished without building any cognitive muscle in either the instructor or the student.

For Dr. Chen, the FLUFF in her Math 101 prep included exactly the work she was doing at 11 p.m.

Formatting. The syllabus font selection. Citation style for the recommended reading list. Standardizing heading levels across her LMS modules. None of this teaches anyone to factor a quadratic.

Layouts. Designing a course banner. Building visually elegant slide decks for the first three weeks. Perfecting the transitions between examples. A serviceable slide communicates; a perfect one does not teach better.

Under-the-hood. Fixing broken hyperlinks in the publisher’s ancillary materials. Converting the prior instructor’s PDF problem sets into a format compatible with her LMS. Troubleshooting why the embedded Desmos calculator was not rendering on mobile.

Filing. Organizing her growing folder of practice problems by topic and difficulty. Categorizing student emails from the first week into registration, accommodations, and math anxiety. Building a gradebook structure that handled her weighted scheme correctly.

Filtering. Scanning the OpenStax textbook for the cleanest worked examples of polynomial long division. Searching for the best video explanation of completing the square to assign as supplemental.

Every item on that list is real work. None of it requires a PhD. Most of it can be drafted by an AI assistant in minutes and then quickly reviewed by Dr. Chen, with the time she saves redirected to the work that does require her.

That work is SPARK.

SPARK: Ideas Worth Thinking

SPARK stands for Specific, Persuasive, Authentic, Rigorous, and Keen-Insight. These are seeding activities, investments where the more cognitive energy Dr. Chen pours in, the more she gets back. They have uncapped payoffs. They cannot be delegated, because what AI generates when asked to do them is a generic average. And a generic average is precisely what Math 101 students at a community college have already encountered, many times, and walked away from.

Here is what SPARK looks like for Dr. Chen’s Math 101 course.

Specific. Her students are not generic Math 101 students. They are mostly working adults, many returning to school after years away from formal mathematics, several of them parents, a meaningful fraction planning to enter health programs where they will need to dose medications correctly. The generic AI summary of “best practices for teaching college algebra” knows none of this. Her job, the work AI cannot do for her, is to commit to particular claims about what these particular students need. A dosing calculation example beats a generic word problem. A worked example using a real local utility bill beats one using “let x be the cost of a widget.”

Persuasive. Math 101 sits at the boundary of a debate Dr. Chen will have to take a position on, whether she wants to or not. Should her course emphasize algebraic fluency, the kind of by-hand symbolic manipulation that has been the spine of the curriculum for fifty years, or should it emphasize quantitative reasoning, the kind of estimation, modeling, and number sense that adults actually use in their careers? AI will hand her a balanced overview of the debate. It will not tell her which side to teach from. That choice is hers, and her students will feel the difference between an instructor who has thought hard about it and one who is hedging.

Authentic. Dr. Chen is a number theorist. She thinks about integers the way a sommelier thinks about wine. She notices things about factoring that her students have never been shown, because the standard textbook treatment was written by a committee optimizing for coverage rather than insight. The authentic move is not to suppress her training and teach Math 101 the way a generic Math 101 instructor would teach it. The authentic move is to bring her particular mathematical taste into the room, in language her students can follow, and let them see what it looks like to actually find a problem interesting.

Rigorous. AI will happily generate a worked example for any topic in her syllabus. Some of those worked examples will contain subtle errors. A misplaced sign, a step that skips a case, a “therefore” that does not actually follow. Dr. Chen’s job is not to generate the example. Her job is to audit it. To catch the sign error before it reaches a student who will memorize it. To notice that the AI’s explanation of why we cannot divide by zero is technically incorrect in three places. Rigor is the SPARK work that distinguishes a competent professional from a dangerous one, and it is exactly what she should be teaching her students to do with AI output in their own future careers.

Keen-Insight. Some of her students have decided, by the second week of the semester, that they are bad at math. They are not bad at math. They have been mistaught, or under-taught, or taught when they were not ready. Dr. Chen can see this. She can see, in the way a student sets up a problem, where the mental model broke years ago, and she can name it. No AI tutor can do this with the precision and credibility that a human instructor can, because the insight comes from sitting across from a person, reading their face, and naming the specific thing they have been carrying around as a private shame. This is the uncapped payoff. A single moment of accurate naming can change the rest of a student’s relationship to mathematics.

The Reallocation

If Dr. Chen had spent Sunday night the right way, here is what it would have looked like.

The first hour would have been a conversation with an AI assistant, drafting a clean Math 101 syllabus from her department’s template, generating three versions of a learning outcomes list, producing a starter slide deck for week one, and converting the prior instructor’s PDFs into her LMS format. That hour would have closed out essentially all of her FLUFF.

The remaining hours, the ones she actually spent picking fonts, would have gone to SPARK. To writing the explanation of why we factor that she will deliver in the first ten minutes of Monday’s class. To selecting which worked examples she will use and which she will deliberately not use, because they reinforce the wrong mental model. To deciding what she actually believes about the algebra-versus-quantitative-reasoning question, so that when a student asks her on day three, “when am I ever going to use this,” she has a real answer rather than the answer every Math 101 instructor in the country has been giving for forty years.

That is the reallocation FLUFF/SPARK forces. It is not a productivity hack. It is a triage protocol. It says: your cognitive energy is finite, your students need the work only you can do, and the work only you can do is not the syllabus font.

The Trap for New Faculty

The cruel thing about Dr. Chen’s situation is that the FLUFF tasks felt productive. She could see them being done. The syllabus got finished. The slides got built. Her LMS modules looked clean. The SPARK work, by contrast, is hard to see while you are doing it. Thinking about how to explain factoring to a student who has been told she is bad at math does not produce a visible artifact. It produces, on Monday morning, an explanation that lands.

New faculty are especially vulnerable to this inversion because the visible artifacts are also the ones that get reviewed. The syllabus goes to the chair. The slides go on the screen. The explanation, the one that changes a student’s relationship to mathematics, lives only in the moment it is delivered.

The first job of the FLUFF/SPARK framework, before it is anything else, is to give new faculty permission to ignore the visible artifacts long enough to do the invisible work. Let AI handle the syllabus. Dr. Chen handles the student.

That is the trade. Get it right in the first semester, and the rest of the career may avoid some of the major miscalculations.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo under a CC BY 4.0 license. The Complete Edition is on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.

I think about that a lot right now, because higher education is having a similar conversation about AI. The worry is that if students use AI tools for writing or research, their critical thinking skills will atrophy. They’ll become dependent. They’ll lose something essential. And the prescription is almost always the same: take the tools away and make students do it the hard way.

Before we commit to that prescription, I think it’s worth looking at what actually happened when we traded paper maps for GPS.

The Trade-Off Nobody Talks About

Yes, paper maps built spatial awareness. Yes, they worked without batteries. But we tend to romanticize the difficulty and forget the costs. Using paper maps meant getting lost, for many of us, regularly. It meant pulling over on highway shoulders. Knocking on a stranger’s door at night to ask for directions. Driving through unfamiliar neighborhoods with no idea how to get out. Unfolding a map across the steering wheel at highway speed because you missed your exit.

GPS didn’t just change how we navigate. It made us safer. It made travel accessible to people who couldn’t read traditional maps or didn’t speak the local language. It made emergency response faster. These aren’t minor conveniences. They’re life-and-death improvements we now take so completely for granted that they’ve become invisible.

When people say “GPS killed our sense of direction,” they’re making a true but radically incomplete statement. Some people did lose practice with a skill. But the question isn’t whether the skill faded, it’s whether the trade-off was worth it. For most of us, the answer is obviously yes.

So What Did Drivers Do With Their Freed-Up Brainpower?

Here’s the part that matters most for teaching. When GPS took over navigation, the cognitive resources that used to go toward route-planning didn’t just vanish. Drivers started thinking about other things, where they were going and why, the meeting ahead, the problem they needed to solve. Or they let their minds wander.

That last part isn’t trivial. Research on the brain’s default mode network shows that unfocused mental downtime (boredom, daydreaming, mind-wandering) is essential for creativity, self-reflection, and insight. When we freed drivers from the cognitive burden of navigation, we didn’t create a vacuum. We created space for a different kind of thinking, the kind that actually produces original ideas.

The same logic applies to students using AI. A student who uses AI to handle a first draft doesn’t stop thinking. She starts thinking about different things like evaluating the output, identifying what’s missing, deciding what to push further. Whether that trade-off is productive depends entirely on how we design the assignment. That’s a pedagogical question, not a technology question.

Critical Thinking Isn’t a Greenhouse Plant

There’s another assumption embedded in the atrophy argument that deserves scrutiny: the idea that critical thinking primarily develops through academic coursework. Our students practice critical thinking every single day outside our classrooms. They evaluate claims on social media. They negotiate family conflicts. They make financial decisions with limited information. They strategize in competitive games where one bad call costs hours of progress – games that, a generation ago, were supposed to rot their brains.

Critical thinking is a weed, not a greenhouse flower. It grows everywhere, because humans are constantly making decisions under uncertainty. That’s all critical thinking fundamentally is.

This doesn’t mean our courses don’t matter. It means we should be honest about what our assignments actually develop. A five-paragraph essay on a topic a student doesn’t care about, written to satisfy a rubric they don’t understand, submitted to an instructor who skims it, was that really building critical thinking? Or was it building compliance? If AI exposed the emptiness of certain assignment designs, that’s not a crisis. That’s useful feedback.

The Skill That Actually Matters Now

The ability that matters most in an AI-saturated workplace isn’t performing tasks that AI can perform. It’s evaluating, directing, and improving AI output: spotting hallucinations, challenging assumptions, verifying claims, pushing past the first plausible answer. That is critical thinking, applied to the context our students will actually work in for the rest of their careers.

We can insist students navigate without GPS. Or we can teach them to think carefully about where they’re going and why — and trust that the human mind, freed from routine cognitive labor, will do what it has always done: wander toward something interesting.

After all, nobody misses being lost.

The students who wrote without AI showed the strongest, most distributed neural activity. The ChatGPT group showed the weakest. After four months, the AI-dependent writers performed measurably worse on subsequent cognitive tests, and they could not even recognize sentences from essays they had themselves submitted weeks earlier. The researchers, led by Nataliya Kosmyna, called the effect cognitive debt.

The study has been read mostly as a warning about AI. Faculty have read it as evidence that AI use in classrooms must be limited, banned, or detected. CTL directors have circulated it as a justification for tighter policies. State systems have cited it in faculty handbook revisions.

That reading is wrong, or at least incomplete. The cognitive debt did not accumulate because the students used AI. It accumulated because they were graded on producing text.

Read the methodology again. The students were not running verification protocols. They were not auditing AI claims against primary sources. They were not running comparative outputs through multiple models to test for hallucinations. They were doing the one task at which AI has now become structurally superior to early-career writers: generating fluent prose on an assigned topic.

That task is the dominant unit of assessment across most of the humanities and large parts of the social sciences. The five-paragraph essay. The reflection paper. The literature review. The discussion board post. These were never wrong assignments before 2022. They were proxies. They measured something we actually cared about, the ability to think through an argument, marshal evidence, weigh counterclaims, by asking for the artifact that ability tends to produce.

The artifact is now uncoupled from the ability. A student can produce the artifact with no underlying ability at all, and the artifact will often look better than what most undergraduates can write on their own.

This is the actual implication of the Kosmyna findings, and it is not a policy implication. It is a measurement implication. If the unit of assessment is “generate text on a topic,” AI has already collapsed the validity of the measurement, and the students who comply with the assignment as written are the ones accumulating the cognitive debt. The AI policy on your syllabus is a downstream patch on an upstream problem.

The lever is the rubric, not the prohibition.

Consider a writing assignment under two different grading schemes.

Under the conventional rubric, the weights are roughly: thesis and argument 30 percent, evidence and analysis 30 percent, organization 20 percent, mechanics and citation 20 percent. Every category rewards the production of the text. A skilled AI user earns full marks. A first-generation student who wrote it herself, with all the rough edges that come from real thinking, earns a B-minus.

Under a verification-weighted rubric, the weights look more like this: verified sources with traceable URLs and direct quotes 25 percent, identified counter-evidence with reasoning for or against 20 percent, recomputed claims and manually checked logic 20 percent, cross-model comparison with documented discrepancies 15 percent, original synthesis the AI did not produce 20 percent.

The artifact looks similar on the surface. The student still hands in an essay. But the points are now distributed across cognitive operations that AI cannot perform on its own and that a student cannot fake without doing the work. A student who pastes ChatGPT output into the assignment scores near zero on four of the five categories, because the model does not know which of its own sources are fabricated, which of its arguments have credible opposition, or where its numbers came from. The student has to do that work, and the work itself is the learning.

This is what the guidebook calls the AI Audit. It is not a clever assignment add-on. It is a structural answer to the measurement problem the brain scans diagnosed.

The argument I am making is narrower than the one usually made about Kosmyna. I am not arguing students should not use AI. I am not arguing you need to detect or police it. I am arguing that the cognitive debt the study identified was a predictable consequence of grading an artifact that AI now produces better than the student. Change what you grade, and the cognitive debt stops accumulating, because the operations that earn the grade are the operations that exercise the brain.

This shift has a second effect that is harder to see at first. It changes which students benefit most. Under the generation rubric, the student with the best prose voice wins, and prose voice is heavily correlated with cultural capital and prior schooling. Under the verification rubric, the student willing to do unglamorous, slow, careful work wins, and that population looks very different. At a community college like Lansing, where most students are working adults, first-generation, or career-switchers, the verification rubric tends to reveal capability that the generation rubric was hiding.

The MIT result is not a story about AI. It is a story about what happens when an assessment instrument loses its validity. The cognitive debt was the cost of running an invalid instrument on a population. The students paid it.

If you are revising a syllabus for next term, this is the question to start with. Not “will I allow AI?” The question is: “What percentage of the points in this course currently reward an artifact that AI can now produce on its own?” If the answer is above forty percent, the brain scans are describing your course.

The fix is not philosophical. It is mechanical. Rebalance the rubric. Add the verification work. Put points on the operations that cannot be outsourced. The students who used to coast on prose voice will struggle. The students who used to be invisible will surface. And the cognitive activity in the room will start to look more like what the MIT scanners measured in the no-AI group than in the ChatGPT group.

The brain is a muscle. It builds when it is used. The grade is the lever that determines whether it gets used at all.

Dr. Szymon Machajewski is the author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. The free OER edition, Chapters 1 through 4, is available on Zenodo.com under a CC BY 4.0 license. The Complete Edition is available on Amazon. For institutional bulk pricing and faculty common-read inquiries, contact press@dataii.com.