A homework sheet has not changed its shape just because generative AI arrived. There may still be ten questions, a paragraph to analyse or an essay due on Monday. What has changed is the number of ways a student can arrive at something that looks finished.
That creates an integrity problem, but integrity is not the whole story. Homework was never valuable merely because a student could submit a correct answer. Its value depended on what happened in the student's head while getting there.
AI changed the shortcut, not the underlying problem
A 2025 study of 548 secondary students across 30 Hungarian classes found that 16% had used AI to complete text-based homework. The study is not an estimate for India, and it should not be treated as one. Its value is narrower: it provides empirical evidence that AI-assisted homework is already a measurable classroom behaviour rather than a hypothetical future problem.
The larger point is that students have always found ways to reduce the effort of assignments that feel repetitive, disconnected or impossible. Search engines changed copying. Answer sites changed how worked solutions circulated. Private tuition changed what could be completed outside school. Generative AI makes the substitution faster and more convincing.
The assignment may be asking the wrong question
If a student can paste a prompt into a chatbot and receive an acceptable answer without needing to understand the material, the machine has exposed something about the task. It may still have a legitimate purpose, but the task is not reliably measuring independent understanding.
This is why the most useful response is not simply to teach teachers how to detect AI prose. Detection turns the classroom into an arms race. A more durable response changes what counts as evidence of learning: a first attempt, a correction, an explanation in the student's own words, a comparison between two answers, a reflection on an error, or a short oral defence of a written solution.
There is a place for AI in homework
UNESCO's guidance on generative AI recommends using these tools to reduce unnecessary pressure around homework and exams rather than increasing it, while also emphasising ethical validation, human accountability and AI literacy. That is a more useful position than either enthusiasm or prohibition.
A student might ask an AI system for three explanations of a difficult idea and then identify which explanation is most accurate. They might ask it to generate a deliberately flawed solution and annotate the errors. They might compare their own paragraph with an AI-generated version and explain which is clearer and why. In these cases, the machine supplies material; the student still has to judge it.
Homework can become evidence of thinking
The strongest redesign is often simple. Ask for the working, not just the answer. Ask what changed between the first and final attempt. Bring one question from home back into class. Require a short reflection on the most difficult decision. Let the task be unfinished in a productive way.
These moves also reduce the temptation to pretend that all homework is independent. A student working with a parent, sibling, tutor, textbook or AI tool is not necessarily learning less. The important question is what responsibility remains with the learner.
The old homework problem remains
There is another reason not to make AI the entire story. Homework can already be excessive, repetitive and unequal. Some students have quiet rooms, reliable internet and adult help. Others share devices, care for siblings, travel long distances or simply have nowhere comfortable to work.
Generative AI can widen those differences, but it can also reveal them. If the only way to make an assignment manageable is to outsource it, the problem is not solely the software. It is also the design and volume of the work.
The homework question after AI is therefore less dramatic than 'How do we stop students cheating?' It is more demanding: what do we want students to practise when nobody is standing beside them? If the answer is judgement, explanation, revision and curiosity, the assignment has to make those things visible.
Homework also carries a hidden assumption: that the finished product is a trustworthy trace of the learner's thinking. That assumption was already imperfect before generative AI. A student can reproduce an answer from a website, copy a model paragraph, ask another person for help or spend an hour following an example without understanding why it works.
The new technology makes that gap more visible because the final product can arrive so quickly. It is tempting to respond by making assignments longer, more complicated or more handwritten. But adding friction is not the same as adding learning. A poorly designed task can remain poorly designed even when it takes twice as long to complete.
A better design starts with the cognitive work. If the aim is retrieval, ask for retrieval. If the aim is explanation, make the explanation observable. If the aim is revision, collect the first and second versions. If the aim is judgement, ask the student to compare alternatives and defend a decision. The task then has a clearer relationship to the learning it is meant to elicit.
There is also a difference between using AI as a source of language and using it as a substitute for judgement. A student may ask for a counterargument, then decide whether it is fair. A student may ask for a worked example, then locate the step that fails. A language learner may compare an AI translation with a teacher's version and explain the difference. These are not automatically good activities, but they keep the learner's evaluation in the centre.
UNESCO's guidance is useful here because it treats generative AI as an educational technology that requires human accountability and AI literacy rather than as a magic solution or an automatic threat. For schools, that means policies should be understandable enough that students know what is permitted and why, while assignments should still be designed so that permitted assistance does not erase the learning objective.
The fairness question matters too. A student who has quiet space, broadband, an adult tutor and multiple devices has a different homework environment from a student sharing one phone with siblings. The technology may magnify that difference. So can a rule that assumes every student can produce the same amount of independent work in the same conditions.
That is why homework reform should not begin with a detector. It should begin with the sentence: this task is asking students to practise ____. The blank forces a teacher to name the learning. Once that is clear, it becomes easier to decide what assistance is legitimate and what would simply bypass the point.
The result may be smaller homework, more purposeful homework or sometimes no homework at all. The point is not to defend the assignment as a tradition. It is to make the learning worth the student's time.
Another weakness of the detection-first response is that it focuses attention on the wrong object. A teacher may spend time asking whether the phrasing sounds machine-generated when the more important question is whether the student can explain the answer. A short oral follow-up can sometimes reveal more about understanding than a software score, while also reminding students that the work belongs to them.
That does not mean every assignment needs an oral defence. The principle is simply to match the evidence to the learning. If the goal is to practise recall, a quick low-stakes check in the next lesson may be enough. If the goal is to build an argument, the student can be asked to identify the strongest evidence and explain why it matters. If the goal is mathematical reasoning, a teacher can inspect the method, not only the final number.
Schools also need a shared language for acceptable assistance. Students should not have to guess whether brainstorming counts as cheating, whether translation is allowed, or whether using an AI system to generate questions is different from asking it to write answers. A policy that names the purpose of the task makes those distinctions easier.
The teacher's workload matters here. Requiring a detailed process log for every piece of homework may simply replace one problem with another. The more useful approach is to use richer evidence selectively, where the learning objective genuinely depends on independent judgement. Not every worksheet needs to become a forensic exercise.
There is a second design opportunity: make some homework collaborative rather than pretending it is always individual. If students discuss a problem with peers, compare explanations or critique an AI response openly, the teacher can treat that interaction as part of the learning rather than as contamination of a supposedly pure product.
That requires honesty from the school. Students need to know when a task is intended to measure independent performance and when tools are allowed. Teachers need enough consistency across departments that rules do not change from one classroom to the next without explanation. Families need to understand that 'help' can range from a useful prompt to doing the work for the child.
The most resilient homework is therefore not necessarily homework that AI cannot touch. It is homework whose educational purpose survives contact with the technology. A task that asks students to judge, revise, explain and reflect can still work in an AI-rich environment because the learning is located in those decisions.
This is why the homework problem existed before the chatbot. The technology has simply made the gap between a finished answer and genuine understanding much easier to see.
That distinction can also change the teacher's response when AI use is suspected. Instead of beginning with accusation, the teacher can ask the student to explain a choice in the work, reconstruct one part, or identify the evidence that supports a claim. The point is not to create a perfect detection system. It is to gather evidence that belongs to the learning objective.
Some subjects make this easier than others. A mathematics teacher can ask for a method, a diagram or a comparison of two solutions. A science teacher can ask why an explanation fits the evidence. A language teacher can ask why one sentence is more precise than another. The disciplinary form changes, but the underlying principle is the same: the student's reasoning should leave a trace.
This also suggests a useful rule for schools: the more easily the final product can be outsourced, the more important it becomes to collect a small amount of process evidence. That evidence does not have to mean surveillance. It can be the normal work of learning made visible.
Teachers can also ask students to annotate their own use of tools. A short disclosure—what was used, for which part of the task and what the student changed—can turn hidden assistance into an opportunity for reflection. It should not become a bureaucratic form for every assignment; it is most useful when the relationship between tool use and learning is itself worth examining.