Every Student Has an AI. Now What Actually Changes?

Somewhere in the last two years the access question stopped being a question. Survey numbers vary and I’d treat all of them sceptically, but every teacher I’ve spoken to reports the same thing from the room: most students, most weeks, for most written work. Whether it’s sixty percent or eighty is an argument about instrumentation. The transition already happened.

Which makes “should students have access” a settled matter and leaves the harder question sitting there, largely unasked: given that they all do, what has to change?

That’s not the cheating question and it’s not the how-do-we-integrate-it question. It’s architectural. Schools perform four separate functions, and AI does something different to each one.

Start with knowledge transmission. Making sure each generation knows what the previous ones worked out is one of the oldest reasons schools exist, and AI doesn’t touch whether it matters. It changes the form. A student who can retrieve any fact in three seconds still needs enough background to know which facts to look for, whether the retrieved answer is credible, and how it connects to anything else they know. Cognitive science calls those structures schemas, and they’re what determines how much new information a person can absorb at all.

AI doesn’t replace a schema. It does make memorising disconnected facts substantially less valuable than it was.

Daniel Willingham has been making a version of this argument since well before any of this, and it holds up better now than it did then: you need to know something in order to think about something. Reading comprehension is largely a knowledge problem. A student struggling with a passage about the American Civil War usually isn’t struggling to decode the sentences. They don’t know enough about the period to fill the gaps the text assumes, and no amount of reading strategy instruction fixes that.

A model doesn’t have this constraint, which is exactly why it matters that the student does. Someone who knows a little about the period can tell whether the answer they got makes sense, can ask a better second question, can catch the confident error. Someone who knows nothing can produce a technically coherent essay on the subject while absorbing essentially none of it.

So: rote memorisation is worth less. Deep conceptual understanding is worth the same or more. That distinction should be restructuring curricula right now and mostly isn’t.

The second function is skill development, and writing is the case worth being precise about, because it’s the one people get wrong in both directions.

Writing isn’t only output. Forcing an argument into prose is a method of thinking: you find out what you don’t actually believe somewhere around the third paragraph. That’s why writing assignments were never really about the essay.

Does AI change this? Partly. The specific skill of producing a competent first draft, prose that covers the right points in roughly the right order, has genuinely lost most of its market value. A student who can’t do that unaided is not obviously disadvantaged any more. But everything that happens after the first draft, editing for precision, finding the weak joint in an argument, deciding which evidence earns its place, defending a claim while someone pushes back, is untouched, and all of it requires the student to have engaged with the material.

The implication is that writing instruction should move from generation to revision. The difficulty is that most curricula are far better at the first, and teaching rigorous revision, teaching someone to read a draft they didn’t write and evaluate it the way a stranger would, is harder to structure, harder to assess, and requires the teacher to be a sophisticated enough reader to model it out loud. A lot of teachers were never trained for that, through no fault of theirs.

The third function is the one AI damages most directly: evaluation and credentialing.

Schools sort people. A grade certifies a level of competence, and that certificate carries information to employers, universities, licensing bodies. Its entire value rests on being a reliable signal.

A grade on a generated essay certifies that the student can operate a language model. That’s a genuine skill and it isn’t the one the grade claims to measure, so the signal degrades into noise. And when credentials go noisy, the institutions that depended on them go looking elsewhere. That’s already visible: over the past few years a run of large employers have quietly reduced the weight on degree class and GPA in favour of portfolios, work samples and technical assessment. Some of that predates any of this. Not all of it.

If universities want their credentials to keep meaning something, the assessments have to certify things a model can’t fake: real-time demonstration, oral defence, hands-on performance, judgment that only comes from having done the thing. All of which is expensive and slow, which means it’s a budget decision rather than a pedagogical one. Most institutions are still deciding. A few are deciding no without saying so.

Underneath all three sits the question that actually determines whether any of this goes well: what happens to the distribution?

The optimistic reading is real democratisation. The student at a school with two overstretched teachers now has expert explanation available at eleven at night. The student whose family could never afford a tutor has something tutor-shaped. The first-generation applicant writing a personal statement has the kind of help that private-school students have always had on tap.

The pessimistic reading is amplification. Students who arrive with strong background knowledge, decent metacognition and a stable place to work are equipped to use this as a learning tool. Students who arrive without those things are far more likely to use it as a substitute for learning. Identical tool, sharply divergent outcomes, sorted by exactly the prior advantages school was supposed to compensate for.

The early evidence is consistent with both, which is the least satisfying possible answer and probably the true one. There are genuine pockets of the first, particularly where access to good explanation was the binding constraint. There are also patterns that look like the second: more sophisticated use on analytical work at the top of the income distribution, more substitution-shaped use at the bottom.

What doesn’t change is that the tool multiplies whatever effort you put in. A student who reads carefully, thinks about what they read, gets genuinely curious and then uses a model to push further gets enormous value. A student who skips to the output gets a finished artefact and nothing else.

Which means the interventions that matter most in an AI-saturated school are not about AI at all. Reading habits. Curiosity. Tolerance for difficulty. Knowing when you’re confused, which is a skill almost nobody teaches explicitly. Those were the things that mattered before, they’re harder to sell than a tutoring platform, and they’re what will determine whether this era is good for students.

There’s a fourth function that vanishes from the conversation whenever it reduces to skills and knowledge, and I think it’s the one at real risk: intellectual identity.

Who does a student believe they are as a thinker? Do they think they can understand hard things? Have they had the experience of sitting inside confusion, trying approaches that failed, and coming out the other side with understanding they built themselves? That experience is where the belief comes from, and it requires the difficulty to be real.

A student who learns that confusion can always be dissolved in ten seconds by asking may end up with a different relationship to difficulty than one who had to sit in it. The long-term research doesn’t exist yet and won’t for years. But the psychology of capability belief is not in dispute: what people believe they can do is built out of what they have done. Consistently succeeding with assistance doesn’t obviously build a sense of what you can manage without it.

That isn’t an argument for withholding anything. It’s an argument that productive struggle, difficulty pitched hard enough to demand real effort and light enough to be survivable, stays essential. The risk of universal access isn’t mainly that students learn less. It’s that they never find out what they could have learned alone, which matters on the day the tool isn’t there, or the problem is too novel for it, or the thing being tested is the nerve to start.

The schools that handle this well will deliberately protect some experiences of unassisted difficulty. Not as nostalgia and not as punishment, but as a specific investment in a student’s belief about themselves. The ones that don’t will graduate people who are excellent at directing a model and quietly unsure what they can do on their own.