Is It Cheating to Use AI for Homework? What the Research Says About Learning

Your child is at the kitchen table with a worksheet due tomorrow. They type the first question into a chatbot, and a complete, polished answer appears in seconds. They copy it down, move to the next, and twenty minutes later the assignment is done. No struggle, no false starts, no moment where the idea finally clicks. The work looks good. And you are left with the question every parent is asking right now: is this cheating, and worse, is it quietly teaching them not to think?

It is a fair fear, and you are right to take it seriously. But the honest answer is more interesting than "ban it" or "let it rip." Whether using AI for homework is cheating depends on how it is used, and whether it damages learning depends on one thing above all others: who does the thinking. This article walks through what the research actually says about AI, effort, and learning, and how to tell the difference between a tool that does the thinking for your child and one that makes your child do it themselves.

This is a deeper look at one question from our parent's guide, Is AI Safe for Kids?, where we cover privacy, accuracy, and the bigger picture of raising a child alongside AI.

Is using AI for homework actually cheating?

Let's separate two questions that parents usually lump together. The first is about honesty: is it cheating in the school's eyes? The second is about learning: does it hurt my child's brain? They are related, but they have different answers.

On the honesty question, the clearest rule is this: if a student presents work they did not produce as their own, that is academic dishonesty. School policies vary, so your child should always follow their teacher's rules, but the underlying standard doesn't change. If the assignment is meant to show what the student can do, such as writing, reasoning, or solving, and AI did that work instead, it is a form of misrepresentation regardless of whether anyone catches it.

But here is the part that matters more for learning, and it is where the conversation usually stops too early. The cheating question is about rules and fairness. The learning question is about something else entirely: whether the cognitive work that creates understanding actually happened inside your child's head. A student can follow every rule and still learn nothing, and a student can bend a rule and still learn a lot. The honest answer to "is it cheating?" is "it depends on what the assignment is for." The honest answer to "will it hurt my kid?" is "it depends on who did the thinking."

Try this: Ask your child one question about the assignment they just finished: "Can you explain the main idea to me in your own words?" If they can't, the work may be done but the learning wasn't. That gap is the thing to watch, regardless of whether AI was involved.

The real question: who does the thinking?

Decades of learning science point to the same core principle, and it is the single most useful idea for thinking about AI and school. The brain often learns best by doing the work of producing an answer, not by receiving one. We often remember what we ourselves generate better than what we simply read or are told.

This is called the generation effect, and researchers Norman Slamecka and Peter Graf first documented it in 1978. Across a series of experiments, people who had to generate a target word themselves, by filling in a blank or solving a clue, remembered it better than people who simply read the completed word. The effort of producing the answer was the thing that built the memory. When the answer is handed to you fully formed, you skip the very step that creates lasting learning (Slamecka & Graf, 1978).

This is not a quirk of memory; it is a general principle. The same logic shows up in retrieval practice: the act of pulling information out of your own memory, such as answering a question without looking at the answer, strengthens that memory far more than re-reading it does. In a well-known 2006 study, students who took practice tests retained the material better over the long term than students who simply restudied it, even when the restudiers felt more confident at the time (Roediger & Karpicke, 2006). The work of retrieving is the work of learning. You can read more about how this applies to studying in how to study for a science test when there's too much to memorize.

So the question for AI is sharp and simple: when your child uses AI, are they generating the answer, or is the AI generating it for them? If the AI is doing the producing and your child is doing the copying, the generation effect never engages. The work looks finished, but the step that would have made the knowledge stick never happened. It is the difference between a child who works out a math problem themselves and one who copies a friend's solution. Both have a completed page. Only one has learned the math.

Try this: When your child is stuck, resist the urge to have AI "just give the answer." Have them state, out loud, what they think the next step is before they ask the AI anything. The act of attempting, even incorrectly, is what primes the brain to learn from whatever comes next.

Why offloading the thinking feels productive but isn't

Here is the trap, and it is a sneaky one: using AI to skip the thinking feels good, and it even produces a good-looking result. That combination is what makes it dangerous. A finished, polished assignment feels like learning has happened, when what actually happened is that the hard part was outsourced.

The same dynamic is the subject of our lesson Learning With AI vs. Letting It Think For You, which walks students through the difference in practice.

Psychologists call this cognitive offloading: the tendency to shift mental work onto an external tool so we don't have to do it ourselves. It is not new. The same dynamic was documented over a decade ago in a famous study on what researchers called the "Google effect": when people expect to be able to look information up later, they remember the information itself less well, but they remember where to find it. We offload the remembering to the machine (Sparrow, Liu, & Wegner, 2011). The same review that frames this idea shows offloading is often rational and useful, but it carries real cognitive consequences: you get the immediate benefit of not having to think, while less gets encoded internally, a cost that tends to show up later, quietly (Risko & Gilbert, 2016).

The deeper reason this feels productive but isn't comes from work on desirable difficulties. UCLA researchers Robert and Elizabeth Bjork have shown for decades that the conditions that make learning feel easy and fluent, such as massed practice, re-reading, or getting the answer handed to you, tend to produce shallow, short-lived learning, while conditions that introduce productive struggle produce deeper, more durable learning. Their key insight is almost uncomfortable: performance during practice and actual learning are not the same thing. A student can look like they are doing great in the moment and be building almost nothing that lasts (Bjork & Bjork, 2020).

This is exactly the "feels productive but isn't" pattern we talk about in our study skills guide for middle schoolers: the student who re-reads their notes for two hours feels like they studied hard, and yet recalls almost nothing on the test. AI supercharges this trap, because the output is not just easy; it is good. The essay sounds smart. The answer is correct. Everything about the experience signals "learning is happening," when the cognitive work that would have created learning was precisely the part that got skipped.

Try this: Watch for the "it felt easy" signal. If your child finishes a tough assignment unusually fast and with no moment of genuine struggle, that ease is a warning, not a win. Real learning usually feels a little hard. Ease on a hard task is worth a second look.

What the newest research shows

The principles above have been stable for decades, but a wave of 2025 research has put fresh evidence behind them. It is worth engaging with honestly, including its limits.

In a peer-reviewed 2025 survey, researchers from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers, adults who use generative AI tools weekly rather than students, collecting 936 real-world examples of how they use them. They found that higher confidence in the AI was associated with less critical thinking, while higher confidence in oneself was associated with more. AI didn't eliminate critical thinking so much as shift it: users spent less effort on the core cognitive work of figuring things out and more effort on verifying, stitching together, and managing the AI's output. That sounds like a fair division of labor, but it means the generative, from-scratch thinking, the part the generation effect tells us builds understanding, was being handed off (Lee et al., 2025).

A second study, from MIT Media Lab researcher Nataliya Kosmyna and colleagues, went further by looking at the brain itself. Using EEG, they tracked 54 participants across three groups writing essays: one using an AI assistant (GPT-4o), one using a search engine, and one using only their own brain. The brain-only group showed the strongest and widest neural connectivity; the search-engine group sat in the middle; the AI group showed the weakest coupling, along with lower reported ownership of their writing and a worse ability to quote accurately from what they had just written. The pattern the authors describe as "cognitive debt," short-term relief from mental effort that may carry long-term costs, was most visible in a small follow-up session of 18 participants who switched conditions. People who had grown used to the AI and then wrote without it showed weaker connectivity and poorer quote recall, while people who had first thought for themselves and then used AI showed higher neural connectivity and better quote recall (Kosmyna et al., 2025).

These findings are striking, and they line up neatly with the older principles. But it is important to state their limits plainly. The MIT study is a preprint, not a peer-reviewed publication; the sample was small (54 people, mostly college-aged, from a handful of Boston-area universities), and the switchback session had only 18 participants; it measured brain connectivity during essay-writing, which is not the same as proving long-term learning loss; and the authors themselves use cautious language throughout, calling the cognitive-debt findings preliminary and asking for larger, longer studies before drawing firm conclusions. The brain patterns are correlations, and the study does not establish that AI use causes lasting cognitive damage. The Microsoft survey, meanwhile, measured self-reported perceptions of effort and confidence: what workers felt, not a direct measure of skill decline.

The honest read is this: the 2025 studies are early warning signals that rhyme strongly with decades of established learning science, not settled proof of harm. They are worth taking seriously precisely because they are consistent with what we already know about effort and learning, but they should be cited with their caveats intact, not treated as the final word.

Try this: When you read alarming headlines about "AI rotting brains," apply the same standard you'd want your child to use: find the original study, check whether it was peer-reviewed, and notice how many people were in it. A single small preprint is a signal to pay attention, not a verdict.

So when does AI help and when does it hurt?

Here is the constructive middle, and it is the most important takeaway for parents. AI is not automatically good or bad for learning. What matters is whether the tool does the thinking for your child or makes your child do the thinking. The same chatbot can be either, depending on how it is used.

• AI as an answer machine (does the thinking for them): "Write me a paragraph about the causes of the Civil War." The student receives a finished product. No generation, no retrieval, no struggle. Learning is minimal, and the work may cross into dishonesty if submitted as original. This is the version to worry about.
• AI as a thinking coach (makes them do the thinking): "I have to write about the causes of the Civil War. Ask me questions to help me figure out my own argument, but don't write it for me." Here the AI acts like a tutor, prompting, probing, and surfacing gaps, and the student is the one generating the ideas, the words, and the connections. The generation effect and retrieval practice both engage, because the cognitive work stays with the learner.

The difference is who produces the answer. The same is true in math: an AI that solves the problem and hands over the steps is an answer machine; an AI that asks "what do you think the next step is, and why?" is a coach. We cover the worked-example approach that makes this concrete in how to study for a math test when reading the notes never works.

This is also why the accuracy of AI matters less than you'd think for learning, and more than you'd think for trust. Even when AI is right, if it did the thinking, the learning didn't happen. And when AI is wrong, which it can be fluently and confidently, a child who has been doing the thinking themselves is far more likely to catch it. That is the subject of our companion piece, why AI sounds so sure when it's wrong (and how to teach your kid to catch it).

The practical test is simple and you can use it tonight: after your child uses AI, ask them to close it and explain the idea back to you in their own words. If they can, if they can reconstruct the reasoning, define the terms, and work the next problem, then the AI acted as a coach and the learning is real. If they can't, the AI acted as a crutch and the learning didn't happen, no matter how good the finished assignment looks.

Try this: Set one household rule for AI and homework: "You can use AI, but you have to be able to teach it back to me." That single test sorts the helpful uses from the harmful ones better than any ban.

Practical rules for parents

A few ground rules, drawn from the research above, that make AI a thinking partner instead of a thinking replacement.

1. Blank page first. Before your child asks AI anything, they should spend a few minutes attempting the work themselves: jotting what they know, where they're stuck, what they think the answer might be. This engages generation and retrieval before any offloading happens, and it means the AI becomes a response to their thinking rather than a substitute for it.
2. Tutor, not answer key. Coach your child to prompt AI as a tutor: "Don't give me the answer. Ask me questions to help me get there." A well-designed tool will meet them there, and if it won't, that itself is a signal about the tool. (Our lesson Asking AI Well shows kids how to do this.)
3. Explain it back. Whatever AI helps produce, your child should be able to re-explain in their own words without looking. This is the retrieval-practice test, and it is the single best check that real learning occurred.
4. One non-AI source. For any factual claim AI gives your child for schoolwork, have them confirm it against at least one source that isn't AI-generated. This builds the verification habit the Microsoft study found adults are losing, and it is the same rule we recommend in our guide to AI and kids.
5. Match the tool to the purpose. For practice, retrieval, and feedback, AI can be genuinely useful. For assignments meant to show what your child can produce, the work has to be theirs. When in doubt, ask: what is this assignment actually measuring, and did AI do the part it's measuring?

These are not anti-AI rules. They are pro-learning rules that happen to keep AI in its most useful role, as a coach that makes kids think, not a machine that thinks for them. That is the distinction that turns a worrying technology into a genuinely helpful one.

The bottom line for parents

Is it cheating to use AI for homework? It is cheating when a student submits AI's thinking as their own. But the bigger question is whether learning happened at all, and that depends entirely on who did the thinking. The research, old and new, converges on one idea: we learn what our own brains effortfully produce, not what we passively receive. When AI does the generating, the retrieving, and the reasoning, the learning largely doesn't happen, no matter how good the finished work looks. When AI prompts, questions, and stretches the learner, and the learner does the producing, it can be a real asset.

The newest studies give us early warning signs that align with what learning science has said for decades, and they should be taken seriously but not overstated. The MIT findings are a small preprint with open questions; the Microsoft survey captures perceptions, not proven decline. Treat them as signals, not verdicts, and apply the same critical thinking to them that you want your child to apply to AI.

If there is one thing to take away, it is this: protect the thinking. Keep the cognitive work with your child. A tool designed to make a student think, to elicit their effort rather than replace it, is doing something fundamentally different from an answer machine, and the research says that difference is where learning lives. That is exactly the design principle behind our AI Tutor, which is built to coach students toward their own answers rather than hand them over. Used that way, AI stops being a threat to learning and becomes one of the better tools a kid can have.

Want help putting this into practice? Start with the free 7-Day Study Reset: a simple week to help your child study less, remember more, and use tools, AI included, in a way that builds real learning. And for the full picture of raising a child alongside AI, begin with our parent's guide, Is AI Safe for Kids?.

References

Bjork, R. A., & Bjork, E. L. (2020). Desirable difficulties in theory and practice. Journal of Applied Research in Memory and Cognition, 9(4), 475–479. https://doi.org/10.1016/j.jarmac.2020.09.003

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv. https://arxiv.org/abs/2506.08872

Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In CHI '25: Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM. https://doi.org/10.1145/3706598.3713778

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002

Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning & Memory, 4(6), 592–604. https://doi.org/10.1037/0278-7393.4.6.592

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745

Get the free 7-Day Study Reset

A simple, week-by-week plan to help your child study less and remember more, and to use AI as a thinking coach, not an answer machine. Get the free kit →