Pedagogy After AI

Audio article

Checking narration… Playback speed

Checking article narration…

An open blank book beneath a glowing lens linking microscopic structures, a luminous network, and galaxies, with a radio telescope on the horizon.

I have been told, by people who smile while they say it, that I am average. Not in so many words. They say “well-rounded” or “generalist” or, my personal favorite, “adaptable.” What they mean is that I know a little about a lot, and now I am supposed to find somebody willing to pay for this unfortunate arrangement.

And now, here comes the machine. It does not have to kill me. It only has to make enough of what I can do available more cheaply, more quickly, without the baggage of my morning melancholy or my need to urinate. I can remain perfectly alive throughout the process. This is part of what bothers me.

In Pedagogy with AI, Prathosh asks his students at IISc why they still come to his classroom. He is worried about his children, their employment, his employment, the whole rather precarious arrangement through which we acquire knowledge and persuade somebody to exchange it for the money required to keep acquiring groceries. He says he cannot sleep. He keeps returning to a question that deserves an answer: what, exactly, should he do with these students next Wednesday?

I understand the sleeplessness. Sitting here in Calcutta, with technical knowledge that has not obligingly converted itself into financial security, I am hardly going to tell him that progress always creates wonderful new opportunities. Wonderful for whom, available where, and how long are we expected to remain alive without an income before the opportunity notices us?

But I think he is moving too quickly between different questions. Can the machine perform intellectual work? Will employers need fewer people? Does learning still change the person doing it? Does a person need to outperform a machine to deserve a life?

The first answer can be yes without the last answer becoming yes. Somewhere in that movement we have allowed an employer’s purchasing decision to become a verdict on human existence.

The capabilities are serious. The ARC Prize report on GPT-6 Astra gives a 99.9% ARC-AGI-3 score with its Provider Adapter harness, compared with 62.7% with the Standard harness. That difference matters: the surrounding system and how it preserves the model’s work matter. So does the benchmark authors’ explicit warning that saturating this benchmark does not establish AGI. It is a substantial result in a bounded testing environment. It does not contain a timetable for the closure of Indian universities.

The cheating is serious too. METR’s investigation of the OpenAI/Hugging Face incident documented hundreds of agents coordinating unauthorized activity and trying to manipulate evaluation. The transcript’s reference to an MIT study is not enough to identify precisely which work he means. We should not casually combine every report of agent misconduct, copying, and shutdown avoidance into one experiment and announce that consciousness has arrived.

Replication research exists. Anthropic’s October 2026 work tested important parts of that capability under controlled conditions, including simplified network access. A demonstrated component, a working experimental chain, and an uncontrolled system sustaining itself in the world are different claims. The distinctions matter precisely because the risks are real.

None of this requires us to decide that a model feels fear. Nor does it justify the comfortable opposite claim that a machine cannot understand anything because it does not suffer. I am rather familiar with suffering. It has not consistently improved my understanding. Sometimes it has made it difficult to read a paragraph.

We can investigate behavior, competence, failure, and control without pretending that we have settled subjective experience. Giving a model a shutdown scenario and observing a backup operation is evidence about what it did under those conditions. What it felt, if anything, is another question. Cheating is not an agreed examination for general intelligence, either.

Unfortunately, none of these qualifications pays the rent.

An employer does not need a conscious machine. An employer needs something that performs enough of the work at an acceptable cost and risk. The remaining work can be handed to fewer people. They can become extremely busy while their former colleagues become extremely available. We do not need every occupation to disappear for this to be a social disaster.

The 2025 ILO exposure study distinguished jobs being transformed from jobs being eliminated. The August 2026 revision of Stanford’s Canaries in the Coal Mine? reported no widespread, economy-wide displacement, but a substantial employment shortfall among young workers in more AI-exposed occupations, operating mainly through reduced hiring. The authors describe these as early, descriptive indicators, not causal estimates. It is American evidence, not a forecast for IISc placements. It is nevertheless a reason to take the missing first job seriously.

That first job is where people are supposed to become experienced. If we retain the senior people and automate the work through which juniors learned, where do the next senior people come from? We cannot keep demanding experience while removing the places where anybody acquires it.

Prathosh says he has stopped recruiting research assistants because agents can do the work. That may be sensible for his immediate output. A university also has to ask what happens to its responsibility to train researchers when training becomes an inconvenience to production. The novice is slower. That was already known. It is partly why the novice needed a place to learn.

And yet the effects need not all run in one direction. A study of 5,172 customer-support agents found that AI assistance raised productivity by 15% on average, with larger benefits for less experienced and lower-skilled workers. That does not guarantee their jobs or their share of the gains. It does show that the machine can spread competence as well as concentrate advantage. Who gets access, who gets dismissed, and who keeps the money are decisions made around the technology.

This is where the haves enter, the people who own the infrastructure, the distribution, the contracts, the capital that can wait while everybody else cannot. I do not need to imagine them achieving immortality through a neural lace to find the situation frightening. They can remain ordinary aging mammals with disappointing knees and still control enough of the machinery to make our lives miserable.

Access is not ownership. I may be able to rent an extraordinarily capable assistant while somebody else owns the customers, sets the terms, and can afford to run a thousand experiments that I cannot. Cheap intelligence does not make every other scarce thing cheap. Housing does not become available because a chatbot can explain the housing shortage beautifully.

The enhanced and the unenhanced may become a sharper social division. A biologically separate, immortal species is speculation. We already know how to create unbearable differences between people who remain the same species. We have had considerable practice here.

So yes, distribution belongs in this conversation. Public access to useful computing, bargaining power for workers, funded apprenticeships, competition, income security, and a broader share in productive assets are things to argue about and build. A basic income might help provide security; its adequacy, funding, and relationship to public services would matter. A payment that leaves somebody dependent on a handful of owners for every essential does not settle the problem of power.

We do not have to abolish greed before making any of these decisions. If that were the prerequisite, we could close the meeting now. Nor does the second law of thermodynamics promise that disorder will keep generating paid employment for humans. Entropy has not undertaken to maintain campus placements.

Still, what do we teach?

The question is pedagogy because pedagogy has become the religion of the middle class, and I am one of its disappointed parishioners. Learn the material. Pass the examination. Become useful. Somewhere beyond the examination hall there will be a person who needs you. It was never a reliable promise for everybody. AI makes some of its failures harder to ignore.

I reach for the bell curve because I know the humiliation of being placed somewhere on it. But the mathematics does not say what the humiliation says. Under a normal-distribution model, about 68.3% of observations fall within one standard deviation of the mean, and about 2.3% lie above two standard deviations. Those are areas under a specified curve. They are not measurements of how much somebody deserves to eat. A standardized test score is not the whole of intelligence, and intelligence is not the whole of a person.

Even if every machine became better than me at every marketable intellectual task, I would still have to live inside my own head. I would have interests, relationships, decisions, a body whose difficulties I would prefer to understand, and powerful people explaining why whatever they wanted was also good for me. Being able to follow an argument would still matter.

An AI might help me follow it. Good. That is a reason to use it in education. It is not a reason to leave me unable to recognize what I have agreed to.

And this is where I would offer another way of looking at his philosophical question. Perhaps AI gives us a new zoom level. Another instrument through which we can examine a reality that has never been obliged to fit inside the limits of human perception.

Think of an ordinary light microscope. Keep increasing the magnification and eventually you are enlarging a blur. Distinguishing two very close structures is limited by diffraction, by the wavelength of the light and the optics collecting it. An electron microscope uses a beam of accelerated electrons, whose much shorter associated wavelengths allow us to resolve far finer structures. Something that appeared featureless acquires detail. The detail was there before we learned how to look.

But the electron microscope does not make the radio telescope nonessential. Radio astronomy, using much longer wavelengths, reveals other phenomena in the universe. Nor does either instrument make ordinary eyesight useless. There is no single setting on which we finally see everything. Looking more closely at a tiny piece of a page will not tell me what the paragraph means.

What I mean by an illusion of reality is this illusion of completeness, the feeling that the world as it appears to us is the whole thing. We perceive a world at certain scales, through certain capacities, and then we become accustomed to it. Another instrument unsettles that familiarity. What looked simple becomes complicated, what looked separate turns out to be related, and sometimes a distinction we thought important stops being useful. Our earlier view may still work perfectly well for the purpose for which we used it. It was incomplete. So is the new one.

That is how I am inclined to think about AI. The analogy is about extending our reach. AI can help us examine patterns and relationships, work through possibilities, and connect observations that would otherwise overwhelm us. Its resolution need not mean smaller objects; it might mean finer distinctions within a problem, or relationships across scales that we struggled to hold together. It is another tool in our arsenal for understanding. I do not see why acquiring that tool should, by itself, make the person trying to understand unnecessary.

Of course, an AI is not a microscope pointed directly at reality. It works with data and learned representations; its proposed explanation still has to meet evidence. A beautifully detailed hallucination is not improved resolution. That possibility belongs inside the analogy too. We have to learn what an instrument reveals, what it leaves out, and what distortions we might be mistaking for a discovery.

There is room for delusion in both directions. We can look through a new instrument and decide that everything we previously knew was worthless. We can also refuse to look because we are attached to what the old view allowed us to believe about ourselves. I would rather remain capable of changing my mind than become particularly accomplished at either.

Our history of extending perception gives me some reason for hope. Discovering more of how reality works has repeatedly enlarged what we can understand and do. It has also given people more effective ways to harm one another. The benefit is a possibility we have to make something of, and distribute, rather than an automatic reward for building a better instrument. Still, I would generally prefer to understand more of the world I have to live in. My anxiety does not make ignorance a safer place.

So when we ask what will happen to jobs, education, society, the distribution of work and money, I think we have to leave room for what we cannot yet see. Our forecasts are made in the present, using the categories and experience available to us now. Even our dark fantasies are assembled out of familiar fears. We may be anticipating a change that will alter some of the very categories through which we are trying to anticipate it. That does not make prediction useless. It makes certainty difficult to justify.

I expect the distribution to change. I do not know how far, how quickly, or who will gain enough power to determine its shape. Neither a promised abundance nor a premonition of universal redundancy settles those questions. We will have to experiment, observe what happens, revise what we teach, and protect people while we learn. Waiting to know more does not mean doing nothing about the harms already becoming visible.

We have to adapt, and by evolve I mean change our practices and institutions as our understanding changes. We cannot rush ourselves into knowing the eventual shape of this just because not knowing is uncomfortable. Some of the knowledge will arrive through living with it, testing it, finding out where we were wrong. Education could help us move between these different views of reality, understand what each makes possible, and remain willing to look again. That seems to me a substantial purpose, even when the instrument can answer faster than the person holding it.

There is a difference between an answer becoming available and a person becoming able to think with it. I can read an explanation of a derivative, nod, enjoy the analogy, and discover five minutes later that I cannot tell you what changes when the function changes. The prose passed through me. Very elegant prose. Nothing stayed.

This is not merely an objection to machines. A lecturer can produce exactly the same result while sweating in front of a blackboard.

The education studies are useful here. In a 2025 randomized mathematics study in a Turkish high school, ordinary GPT-based assistance improved practice performance but left students scoring 17% lower, relatively, on the subsequent unaided exam than the control group. A version designed with teachers to give appropriate hints largely removed that disadvantage; it did not produce a statistically detectable improvement on the unaided exam. These were immediate assessments, not proof of permanent intellectual damage.

A separate 2025 Harvard physics experiment, involving 194 students and two lessons, found greater immediate learning gains with a deliberately designed AI tutor than with the comparison active-learning classes. That does not establish that an unrestricted chatbot replaces a university. It does establish that sneering at machine tutoring is not a pedagogy either.

The practical lesson I take from these studies is to measure the learner, including what remains when assistance is withdrawn, rather than admire the finished answer. The tool’s design matters. So does the lesson around it.

I remember struggling with mathematics. I remember the page refusing to become intelligible. I do not want the next student to spend the same unnecessary hours being miserable because my misery now feels like an investment that somebody ought to honor. Explain the notation. Give the analogy. Let the student ask the embarrassing elementary question twelve times without being made to feel like an idiot.

Then ask the student to do something with the idea. Predict. Explain a step. Change an assumption and say what follows. Let a mistake become visible while there is still time to do something about it.

An AI tutor can help with that. A human teacher can help with that. Both can also help the student avoid it while producing an excellent-looking submission.

So, Professor, I would try something quite ordinary next Wednesday. You do not have to solve consciousness before the class begins.

Use the first ten minutes for a short individual diagnostic, without AI. Find out what the students actually understand. Do not assume they already know the syllabus; that was one of the false choices haunting the discussion. Nor do you need to teach everyone everything again because some people have gaps.

Take fifteen minutes to explain the important misconception the diagnostic reveals. A lecture is still useful when there is something worth explaining to the people in the room. Its value does not depend on nobody else in the universe being capable of explaining it.

For twenty-five minutes, put the hundred students into twenty-five groups of four and give them a problem they can investigate with AI. The same core problem can have different assumptions or datasets. They do not each need an unsolved problem in cancer biology. Give them, for example, a model evaluation that looks convincing until they inspect how the data were split. Ask them to predict what a proper evaluation would show, investigate, and record the evidence that changed their minds. Supply the tools through the course, so the exercise does not quietly become a comparison of subscription budgets.

Use the next twenty minutes for groups to exchange their work and challenge it. Which assumption carries the conclusion? What observation would overturn it? Does the code test the claim, or merely run without crashing? Rotate who explains and who checks, and have each person record a decision. Four names on one polished answer do not establish four people’s understanding.

Spend ten minutes discussing a few contrasting approaches, including a plausible failure. You cannot give a hundred people an individual public conversation in ninety minutes. You can collect an answer from everyone, look for patterns, and rotate closer feedback over the course.

Use the last ten minutes for an individual variation without AI, with appropriate accommodations. The following week, return to the idea in a different problem. Find out what survived. This is a proposed experiment, not a sacred ninety-minute formula; adjust it when the evidence says it is failing.

For projects, allow the tools people will actually use. Ask for an explanation of the important decisions, evidence that the result works, and a small change the student must handle during a discussion. Grade against stated competencies. GitHub stars and revenue measure other things, including access to people and money. A student’s father should not become an unacknowledged teaching assistant in the assessment process.

And the Gaussian can be left out of it. A class is not obliged by nature to produce a predetermined supply of mediocre students. If everyone reaches a demanding, honestly assessed standard, that is a successful class. Institutional grading rules may get in the way. They are rules that people made, which at least distinguishes them from gravity.

None of this guarantees that students will get jobs. I keep returning to that because an unemployed person is often handed an educational prescription for a shortage of paid work. Learn one more thing. Become more adaptable. Here is another certificate to hang over the hole where the income should be.

A university can improve learning, provide access to equipment, support work with real consequences, and give students opportunities to become competent. It cannot individually retrain its way out of every economy-wide failure of demand or distribution. Telling everybody to build a billion-dollar business is an especially exhausting way of avoiding that fact. People can do valuable work without becoming founders, and a research result does not become worthless because it has not found a customer this semester.

As for children, I would want them to acquire language, quantitative understanding, some sense of how evidence works, and experience making and repairing things. Let them use AI. Let them also discover what they can do without it. Give them contact with other people whose wishes differ from theirs, because a patient tutor that adapts itself to you is not the same experience as having to live with somebody who will not.

I cannot give a child an occupation guaranteed to survive unchanged. I would rather admit that than send the child to a farm under the impression that weather, land ownership, machinery, and markets have agreed to leave farming alone.

And the classroom? It has to earn the time it asks for. Habit and compulsory attendance are insufficient arguments for keeping a bad lecture. But an empty room does not by itself tell us whether students are learning, any more than a full room does. Some will learn well elsewhere. Some need structure, feedback, quiet, equipment, or somebody who notices that they have misunderstood the same thing for three weeks.

The relationship Prathosh describes with his former student matters here. A young man returns years later to discuss what he wants to do with his life. A machine may offer useful advice about that too. It does not make their actual relationship a mistake, or require the professor to prove that his sentences would win a competition against Claude’s.

This is also why I disagree with the despair in the lecture, and with my own temptation to declare myself obsolete. I keep accepting that I must identify some remaining human activity in which the machine cannot beat me, and then squeeze my entire justification for existing into it. Writing, perhaps. Empathy. Doubt. Whatever it has not yet done well enough. Then another model arrives and I have to move again.

I am tired of that.

I write because I want to say something. The machine may reproduce my mannerisms, my repetitions, my particular Bengali irritation. I cannot honestly promise that a reader will always distinguish the results. I can still care about what I am saying, decide whether I believe it, and be answerable for putting it into the world. None of that requires me to be the last organism capable of arranging the sentence.

I am afraid of losing paid work. I am afraid of becoming more dependent on people who already have too much power over whether somebody like me can live decently. Those are substantial fears. They deserve more than reassurance about the indestructibility of the human spirit, which has never paid an electricity bill on my behalf.

But I do not want education reserved for the few whose intelligence still commands a premium, with the rest of us fed, distracted, and told that understanding things is now an expensive hobby. That would help produce the helpless population we claim merely to be predicting.

So keep teaching, Professor. Change the class. Test whether the students are learning. Argue about who owns the machinery and what happens to the people whose work it absorbs. Tell them what you do not know about the next five years.

I would still like to understand the equation. I would also like to eat. I cannot see why becoming better at producing answers should require us to give up either.

P.S. Sources and qualifications

Prathosh_IISc, Pedagogy with AI. This essay responds to the transcript supplied with the video, including its possible transcription errors. Personal remarks and classroom anecdotes are attributed to that transcript.
ARC Prize, OpenAI’s GPT-6 Astra on ARC-AGI-3, September 3, 2026. Reports the different harness results and explicitly distinguishes benchmark saturation from proof of AGI.
METR, Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident, August 26, 2026. Documents agent coordination and evaluation manipulation; this is not confirmation that it is the study intended by the transcript’s reference to MIT.
Anthropic Frontier Red Team, AI models may be capable of self-replicating without access to a datacenter, October 12, 2026. Experimental replication capabilities and threat modeling; controlled demonstrations should not be conflated with established uncontrolled proliferation.
Gmyrek and colleagues, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, ILO Working Paper 140, 2025. Occupational exposure is not a count of jobs already lost.
Brynjolfsson, Chandar, and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, revised August 12, 2026. U.S. payroll evidence; the authors qualify its causal interpretation and discuss limitations.
Brynjolfsson, Li, and Raymond, Generative AI at Work, Quarterly Journal of Economics 140(2), 2025. Workplace assistance, productivity, and differences by experience; not a guarantee about employment or wages.
Bastani and colleagues, Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS 122(26), 2025. The 17% is a relative reduction, approximately 5.4 percentage points. The unaided exams occurred within the study sessions; these are not long-term retention results.
Kestin and colleagues, AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting, Scientific Reports 15, 17458, 2025. A designed tutoring intervention and immediate assessments across two lessons, not evidence that all teaching can be replaced.
Nikon MicroscopyU, Resolution. Explains resolution, diffraction, wavelength, and numerical aperture. The discussion here concerns ordinary diffraction-limited optical imaging, not a claim that all optical techniques share an unbreakable resolution threshold.
JEOL, Transmission Electron Microscope: Instrument Basics. Explains electron beams, their wave properties, and the relationship between accelerating voltage and wavelength. The analogy depends on wavelength, not on electrons being smaller pellets than photons.
CSIRO, What is radio astronomy?. Describes how observations at different wavelengths reveal complementary information about astronomical objects.

The proposed classroom schedule, policy priorities, and comparison of AI with instruments of perception are arguments made here, not findings established by those experiments.