Homework is the easiest place to see what AI has done to education, because homework is where the evidence is now daily and undeniable. Teachers know. Parents know. Students know. The only people still pretending it's fine are the ones writing the policies.
What homework was supposed to do
Before we talk about what to do, let's be clear about what homework was trying to accomplish:
- Consolidation. Practice the skill you learned in class, so it sticks.
- Extension. Apply the skill in a slightly new context, so it generalises.
- Exposure. Encounter new material you'll build on tomorrow.
- Assessment signal. Show the teacher, asynchronously, what you understood and what you didn't.
Of those four, only the last is structurally broken by AI. But it's the one we spent the most bandwidth on, and it's the one parents and admissions inspectors are trained to notice.
What AI actually breaks
AI doesn't break consolidation. If a student uses AI to generate ten practice problems and works through them, they've consolidated. If they use AI to check their answer to a problem they solved themselves, they've consolidated.
AI doesn't break extension. Most of the richest extension prompts are now better with AI — "explain this idea to someone three years younger; now critique your own explanation" is a prompt that barely worked before AI and works beautifully now.
AI doesn't break exposure. A student reading an AI-summarised primer on tomorrow's topic is not a worse-prepared student. They're a differently-prepared student, and in almost every case, a more-prepared one.
AI breaks the assessment signal. Specifically: the signal we got from asking students to produce written artefacts alone, at home, without supervision, and then inferring their understanding from the artefact. That signal is gone. Pretending it isn't is the single most expensive error schools are currently making.
The redesign is simpler than you think
Once you accept that take-home written artefacts are no longer reliable assessment signals, the redesign falls out naturally:
-
Move assessment in-class. Short, frequent, low-stakes. Not more tests — more writing-under-observation, more thinking-out-loud, more whiteboard explanations. If a student can explain it at the whiteboard, they understand it. The whiteboard was the original assessment tool and it still works.
-
Repurpose homework for the first three functions. Consolidation, extension, exposure. Design the homework to assume AI assistance, not to forbid it. "Generate ten variations of this problem, solve three, and identify the one you found hardest" is a homework task that uses AI honestly and produces a better student.
-
Use the homework for conversation, not grading. The point of collecting it isn't to mark it. The point is to start tomorrow's lesson with "I noticed that half of you got stuck on problem four — let's talk about it." You don't need a grade to run that loop.
The policy implication
If this framework is right, then "AI policy" isn't really about AI. It's about which part of the educational pipeline each kind of work is supposed to serve. Assessment moves in-class. Practice moves to homework, with AI explicitly welcomed. The "no AI on homework" policies being drafted in faculty meetings right now are trying to solve a problem that doesn't exist — while leaving the actual problem (eroded assessment signal from unsupervised written work) completely untouched.
We are three years into this. Schools that spend the next year redesigning homework the way described above will finish next academic year with better-taught, better-assessed, better-served students than they had when they started. Schools that spend the next year writing more-restrictive AI homework policies will have the same teachers, the same students, the same assessment blind spots, and a slightly worse relationship with their parents.
The homework problem is a gift. It's the clearest, smallest, most tractable version of the AI-in-education question. Schools that solve it first will find the rest of the curriculum much easier to move.