Perfect homework, poor exam

In Fullerton, California, a maths teacher tells his students something teachers did not used to say. He is “not going to pretend like you aren’t” using AI.¹

He is not alone. In July 2026, The 74 reported that teachers across the United States were rethinking, scaling back or, in some cases, dropping homework, because AI can finish it in seconds.¹ The story reached us in September through local news sites.

The notes in the margin are the journey: where we got something wrong, checked, cut or rewrote. Facts carry numbered sources. My own views are marked as mine. Where nobody knows, I say so.

The shortcut itself is old. In a survey of about 30,000 American high schoolers, a third said they had copied someone’s homework, and a third had used AI as an unauthorised aid.¹ What is new is that it is now free and instant. In a Pew survey fielded in autumn 2025, 54% of American teens had used AI chatbots for schoolwork, and one in ten did all or most of it that way.²

The companies have an answer: a “teach me” mode that asks questions instead of handing over answers. Anthropic, OpenAI and Google each launched one in 2025.⁴⁵⁶ All are optional. Justin Reich of MIT put the problem in one line: the “carefully teach me this stuff” button sits right next to the “do everything for me” button.¹

The young people notice: in one survey of 12 to 29 year olds, 67% now think AI harms their critical thinking, up from 54% earlier in 2025, even as their use rose.³

In early September 2026, New York City and Los Angeles restricted student use of generative AI for the year.⁷ I do not know if they are right. Nobody does yet.

---

What was homework for?

Before the machine, a smaller question: what is homework?

Some of it is practice. Some asks you to learn alone before class. Some asks you to gather facts for a report. In my view, what they share is that you learn by doing: you think, you try, you get it wrong. The paper you hand in is evidence that this happened. It is not the point.

In 1916 John Dewey wrote that the schools that worked gave pupils “something to do, not something to learn”, and that “learning naturally results”.⁸

Does homework even work? The best-known review (2006) found a clearer link with achievement in the upper grades than in the early ones, and no sign that more is better.⁹ A study of more than 7,000 Spanish teenagers found that, among things a student does, the best predictor of a payoff was neither time nor effort. It was autonomy: doing it yourself. (Prior knowledge mattered most of all.)¹⁰ That is the one thing an AI answer takes away.

The worry is also older than AI. In January 1900 the Ladies’ Home Journal ran an article calling homework “a national crime”.¹¹ Its author, the editor Edward Bok, also asked whether a child helped by a parent might come to depend on others.¹¹ That is the AI worry with a parent in the machine’s place.

---

Perfect homework, poor exam

Here is the danger. If the paper is only evidence, the evidence can now be made without the thing it is evidence of. That is no longer only a worry. It has been measured, in two large studies.

In China, a larger study followed 26,811 students for thirty months. When they took up generative AI, homework scores rose 18%, and within six months their monthly exam scores fell by about 20%.¹³ Two cautions. It is a working paper, not yet peer reviewed, and it compares students before and after, not at random. The losses were concentrated among students who outsourced their homework, about 80% of users.¹³

When I first talked this through, I described it from the teacher’s side: the homework is perfect, but the exam does not look good. “Then it actually shows who…”¹⁴ I did not know two large studies, in two different countries, had found something very close.

Why does nobody notice from the inside? Because fluent work feels like learning. In a 2006 experiment, students either re-read a text four times, or read it once and tested themselves three times. A week later the self-testers remembered 61% and the re-readers 40%, yet the re-readers had predicted they would remember more.¹⁵ Performance gains are not learning, as researchers now say about AI.¹⁶

The oldest version of this gap is about 2,400 years old. In Plato’s Phaedrus, an Egyptian king refuses the gift of writing: students will have “the show of wisdom without the reality”.¹⁷ Note who says it: a king, in a tale, told by Socrates, written down by Plato.

---

Same tool, opposite results

The same AI helped some students and hurt others. So what decides it?

Picture three students given the same project and the same AI. The first uses it to learn. The second already knows how, and gets the whole project in one go. The third says “just complete it for me.”¹⁴ The teacher receives three good projects. Inside three students, three different things happened.

My first instinct was simple. It isn’t about the tool. It isn’t only the tool. It is whether the student has a heart to learn, or just wants to finish.

Where I stand now: I think adults grow that heart, partly by designing work where thinking is the only way through, though students’ own habits matter too.

I began this story with an image of my own: “People suffer then they grow just like the steel is being heated in order to smash and become sword.”¹⁴ But not all heat forges. Some just burns. Difficulty helps only when the learner can respond to it, which Robert and Elizabeth Bjork call desirable difficulty.¹⁸ What builds a mind is making an answer and pulling it back out of memory, called retrieval practice. A related finding, the generation effect, is that what you produce yourself is remembered better than what you only read; a 2007 meta-analysis found it across 86 studies.¹⁹ Struggle is not sacred; useful difficulty is.

The other side deserves its best version. In a 2025 Harvard trial of 194 physics students, a tutor designed around good teaching beat an active-learning class, in less time, in one course.²⁰ But the big claims shrink when you look closely. A study of more than 6,000 middle-schoolers found a small, borderline gain on one practised question (p = 0.065), which its authors called “not transformational”.²¹ My reading: the case for AI in learning is strongest exactly where it is designed not to do the work. That supports this piece.

---

How would a teacher know?

If the paper looks the same either way, how would a teacher know?

Not with a detector. In a 2023 study, seven popular AI detectors flagged, on average, 61% of essays by non-native English speakers as AI-written.²² Australia’s higher-education regulator wrote in 2025 that reliably detecting AI use “is, at this point, all but impossible”.²³

The signal already exists: the exam. Across a class, strong homework next to weak exams is the gap both studies measured. For one student it is a reason to ask, not proof. The other signal is to ask plainly. My version is five questions: What did you do? Why this way? What if I change one thing? Where could it be wrong? What would you ask next? A student who did the work can say why.

But it costs time. Published university practice puts a short oral check at 10 to 15 minutes per student.²⁴ For one class of 30, that is five to seven and a half hours, about the 5.9 hours a week that teachers who use AI weekly report saving.²⁵ A secondary-school teacher with five classes would need five times that.

It might have to be sampled: a few students, a cold question, the exam you already give. One university, Sydney, splits assessment into two lanes: secure checks where learning must be shown, and open tasks where AI is allowed and taught.²⁶

Why not just ban it? A 2025 study compared 20 English secondary schools with restrictive phone policies and 10 without, among 1,227 pupils. The restrictions cut phone use during the school day, but the study found no difference in wellbeing, and its commentary notes none in sleep or exam results.²⁷ It was observational, not a trial, and only about one pupil in ten took part, so it settles little. It suggests, at most, that a rule at school changes the school day, not necessarily the habit. Check the learning, not the paper.

---

Who can change school?

A good check needs a school willing to adopt it. Schools were not built to change quickly.

I said it bluntly at first: the education system was designed to make people obey mass schooling was born partly as a tool of order. The political scientist Agustina Paglayan argues from data that states built primary schools mostly out of elites’ fear of the masses, often after unrest.²⁸ Her critics say she underplays how teachers and students resisted, and literacy and citizenship were also goals of mass schooling.²⁹

I think change starts where schools are free to try: a teacher can pitch to a board, check the research and run a pilot.

The learning modes in the companies’ products are optional, so a student can switch them off.⁵ I think the real question is who holds the switch. By a lock I mean the rule that says, task by task, how much AI is allowed. My view is that the lock belongs on the assignment, not on the tool: the teacher sets it, and checks the learning where it matters. That only works in a school willing to adopt it, which is why the history above matters.

---

What should stay inside a person?

If a machine can do the task, what should school leave inside a person?

In my view, education is to survive, to grow through the friction of society, and to find your own path.

Here the research pushed back on me. I used to think a map of how things connect could stand in for stored knowledge. The cognitive scientist Daniel Willingham argues that you can only think critically about what you know.³⁰ In a study of baseball, weak readers who knew the game understood a passage about it better than strong readers who did not.³¹ So you need enough connected knowledge in each area to notice when an answer is wrong. With thin knowledge and a fluent, confident AI, you will not notice.

How much is enough?

I don’t know.

Is any skill sacred? I do not think there is a limit: AI may come to do almost everything we can. We do not know. But that does not settle what a child should practise today. The calculator is the precedent: the research backed using calculators alongside ordinary maths teaching, not instead of it.³² Keep the tool, protect the core skill. What to protect is decided per goal, for now.

Promises about classroom technology have a long record. In 1913, Thomas Edison predicted: “Books will soon be obsolete in the public schools.”³³

The old fear also breaks. Socrates’ deeper objection to writing was that a text keeps “a solemn silence” when you question it.¹⁷ AI answers back, fluently, and sometimes wrongly. A calculator does one checkable step; AI can do the whole task. So history does not tell us we will be fine. My reading is that it went best where people redesigned teaching on purpose.

Where this does not fit yet: in Thailand I found no national data on how students use AI for homework. A 2026 survey of 1,777 students found that 94% had used AI to help with their studies. It was not a random sample, so it shows a pattern, not a rate.³⁴ Young children are different: Common Sense Media advises that children under five use no AI chatbots at all.³⁵ For some students with learning differences, AI may be an accommodation, not a shortcut. The evidence is early.³⁶

---

Where I have landed

Here is what I believe now. A finished assignment no longer proves that a child learned. The answer is not to police the tool. It is to check the learning, in class, and to design work where thinking is the only way through.

I will not claim that AI rots the brain: the study behind that headline had 54 people and is still an unreviewed preprint.³⁷ I will not claim that AI tutors double learning, or that I know how much knowledge is enough.

I changed my mind three times along the way. I started with the forge: struggle makes us. I now think useful difficulty does. I thought a map of how things connect could stand in for knowledge. It cannot. And I thought it was not about the tool. It is partly about the tool.

Two things I will be watching: whether the New York and Los Angeles restrictions last the year, and whether any school publishes what happened when it replaced homework with in-class checks.

AI can do the homework. Nobody can do the learning for you.

So what should we leave inside a child, now that the answers are free? I do not know yet.

Keep this for the next time someone shows you a perfect piece of work. Ask them how they made it.

---

How this was made. I talked this story through out loud, in a long conversation. Most of it was lost: 23 of its 133 turns survive word for word, and the rest were rebuilt in the way I think and ask, and confirmed as mine on 7 October 2026. We used tools from Anthropic and OpenAI in the research, and both companies’ learning modes appear above, so we have tried to describe them no more kindly than Google’s. Sources were checked on 1 and 6 October 2026. We will get some things wrong. When we do, we correct them in public.