What to Teach When AI Can Do the Homework

Every leap in technology produces the same tempting argument: the machine can do X, so why teach X? Calculators were supposed to end arithmetic; search engines were supposed to end memorisation. AI, which can draft essays and solve equations, has revived the argument in its strongest form yet. It deserves a serious answer rather than a reflexive one — and the serious answer is that AI changes surprisingly little about what foundations students need, while adding a genuinely new layer on top.

You cannot check what you never learned

The case for foundations no longer rests mainly on needing to produce the work yourself. It rests on something the AI era has made more important, not less: judgement over machine output. AI produces fluent, confident, sometimes-wrong answers. The only person who can catch the wrong ones is someone who knows the material — a student who never learned the mathematics cannot spot the plausible-looking error in step three, and a student who never learned to write cannot tell a competent AI paragraph from a hollow one. There is a quiet irony here: the better AI gets at producing work, the more valuable it becomes to know the subject well enough to evaluate work. Expertise is shifting from scarce production to scarce quality control, and quality control demands the same foundations, learned the same slow way.

There is a second, less obvious reason. Thinking happens in what you know: working memory is tiny, and it is fluent background knowledge that lets a student hold a complex idea in mind at all. A head empty of facts is not free for higher-order thinking — it has nothing to think with. Offloading the foundations to a machine does not liberate cognition; it hollows it.

The skills that rise in value

• Verification as a reflex: checking claims against sources, solving independently before trusting a solution, treating confidence as no substitute for evidence. This was always good intellectual hygiene; AI has made it a survival skill.
• Questioning: when answers are nearly free, the scarce skill moves upstream to asking the right question — framing a problem, noticing what is missing, probing a fluent answer for its weak point. Classrooms that grade the quality of students' questions, not only their answers, are teaching for this directly.
• Judgement of quality: is this argument actually good? Is this the right method, even if it runs? Evaluation sits near the top of Bloom's taxonomy for a reason, and it has quietly become an everyday workplace skill rather than an academic summit.
• Clear expression of intent: getting good work out of AI is largely the skill of specifying precisely what you want — which is, unglamorously, the old skill of clear writing and clear thinking, now with an immediate payoff.

The durable human skills

Beyond the machine loop sit the abilities that keep their value precisely because machines do not compete there: collaborating with other people, persisting through difficulty, communicating face to face, and caring about whether the answer is actually right. Employers surveying the AI era keep landing on the same list, and it looks less like a technology curriculum than like a description of a well-brought-up, well-taught human being. That is worth saying plainly, because it is reassuring: the schools and tutors already teaching rigour, honesty, and resilience are already teaching AI-era skills.

What this means in practice

1. Keep teaching foundations unapologetically — and give students the real reason: you cannot supervise a machine in a subject you do not know.
2. Add verification to the taught curriculum: exercises where students fact-check, error-hunt, and grade AI output on material they have just learned.
3. Assess more of what AI cannot fake: oral explanation, in-class work, and the ability to defend a piece of reasoning in person.
4. Reward good questions explicitly — they are now at least as diagnostic as good answers.
5. Be honest with students about why effort still matters: the struggle is not the price of learning but the mechanism of it, and no tool has changed that.

The homework AI can do was never the point of homework; it was the visible trace of an invisible process. The task for educators is to keep the process alive while the traces get easier to fake — harder work than before, but not different work. Teach the foundations, teach the checking, and teach students to stay the author of their own thinking. Everything else is tooling.