The Homework Completion Machine: When Schoolwork Is the Only Thing Your Kid Uses AI For

What this is: a look at one specific, narrow finding — not a verdict on AI in schools generally, and not an argument that homework help is the "worst" use of AI for kids. It's about what happens when the only thing a kid ever does with AI is outsource a task someone else designed, with no structure around it. The OECD has data on that pattern. A comment from Amplify's Dan Meyer on a September panel describes the same pattern showing up at home. Parents who only ever see their kid open AI for homework should know what that combination tends to produce — and what the alternative looks like.

What the OECD actually found

The OECD's Digital Education Outlook 2026 looked at what happens when students get access to general-purpose generative AI for schoolwork. The headline finding is specific and worth stating precisely: when a task is outsourced to GenAI without pedagogical design around it, performance goes up but real learning does not follow. Worse, the advantage can disappear or reverse once the AI access is taken away.

The report cites a field experiment in Türkiye high schools: students using GenAI on practice problems scored roughly 127% higher on that practice work. The same students then sat an exam without AI access and scored roughly 17% worse on the exam without it. The tool made the homework look done. It didn't make the student more capable of doing the work unassisted.

The OECD's broader point cuts the other way too — when GenAI is designed into the pedagogy (the task is structured so the student still has to reason, check, and revise), the research found more sustained learning and stronger critical thinking, creativity, and collaboration. The technology isn't the variable. The design around it is.

That distinction matters for this post, because it's easy to flatten "AI for homework is bad" out of this finding, and that's not what the research says. The finding is about unstructured task outsourcing, a pattern that happens to show up most often in exactly how homework-AI gets used in practice: open the chat, paste the problem, get the answer, turn it in.

What a panel comment adds

In September, at a webinar on who sets the rules for K-12 AI policy, Dan Meyer of Amplify (formerly of Desmos) said on the panel that in the classrooms he works in weekly, AI is close to a non-event — teachers don't need to fear missing out on it. He also said, on the same panel, that kids with home access to frontier AI models mostly use that access to get their homework done for them, and that he's observed this pattern raise homework completion while lowering test scores. He didn't cite a study for that observation, and we're not treating it as one — it's a working educator's read from the field, offered on a panel, not research.

Put next to the OECD field data, though, it's a plausible read. A kid who opens AI only when a worksheet is due and closes it when the worksheet is turned in isn't practicing judgment about the output — they're practicing getting the assignment off their plate. The OECD's Türkiye numbers show what that produces at scale: inflated practice performance, reduced unassisted performance. Meyer's panel comment describes the same shape showing up informally, at home.

Why this isn't an argument against homework help

It would be easy to take this post and conclude "don't let your kid use AI for homework." That's not the finding, and we're not making that argument. The OECD explicitly found that GenAI designed into the learning task — where the student still has to do the reasoning and the tool scaffolds rather than substitutes — produces real gains. A kid using AI to get unstuck on one step of a math problem, then finishing the rest themselves, is a different pattern than a kid pasting the whole problem set and copying the output. Both involve "AI for homework." Only one of them is the pattern the OECD and Meyer are describing.

The actual risk is narrower and more specific: if homework-completion is the only context in which your kid ever uses AI, you've picked, by default, the use the OECD found produces completion without learning when it isn't designed well — and most homework-AI use isn't designed at all, because the kid is the one deciding how to use it, with no pedagogical scaffolding in the loop.

The alternative: something the kid chose, that has to work

Here's a pattern that sits structurally apart from homework outsourcing, without claiming to be a replacement for it or a cure for the effect Meyer described: AI used on something the kid picked for themselves, where the output has to actually function, and the kid is the one who can tell whether it does.

A kid building a platformer game with AI can't paste in a vague request and walk away satisfied, the way they can with a homework answer. If the jump height is wrong, the game is unplayable and they know it immediately — nobody has to grade it. If the dragon moves slower than the kid imagined, that's not a subjective call, it's visibly wrong against what they wanted. The judgment is built into the task, because the task is self-grading: does it work the way I wanted, or not. That's a structurally different relationship to AI output than "did this look like a correct homework answer that a teacher will accept."

What we actually see, and its limits

On Xyplor, kids revise their AI-built creations by giving feedback — "make the dragon faster," "the jump is too floaty," "that's not what I asked for." We tagged the first 71 classified kid revisions internally and found that 55% were fixes — the kid rejecting what the AI produced and asking for something different, rather than just adding a feature or starting over.

We want to be precise about what that number is and isn't. It's an internal tag count on a small early sample (N=71), not a study, and the proportion may move as volume grows. It tells us that more than half of the revisions in that sample were the kid rejecting what the AI produced. It does not tell us that kids who do this learn more, retain more, or are protected from the completion-without-learning pattern the OECD and Meyer described. We haven't measured transfer, retention, or any academic outcome, and we're not claiming to.

What we think is true, more modestly: a context where the AI's output has to satisfy the kid's own stated intent, and where failure is visible and immediate rather than graded by someone else later, is a different relationship to AI than pasting a worksheet. Whether that difference compounds into anything measurable over years is a question we don't have an answer to yet.

What this means practically

A few things that follow from the OECD finding and the panel comment, without overclaiming:

• Notice if homework is the only context your kid uses AI in. That's not inherently harmful, but it is the use the OECD found produces performance without learning when unstructured — and most home homework-AI use is unstructured by default, because no teacher is in the loop shaping how it's used.
• If AI is going to scaffold homework, the scaffolding matters. A kid using AI to explain a concept they're stuck on, then doing the rest themselves, looks different from a kid pasting the whole assignment. Ask which one is actually happening.
• Consider giving your kid at least one context where AI output has to satisfy something they chose, not something assigned. This can be Xyplor, it can be something else — the structural property that matters is that the kid can tell, on their own, whether the output is right.
• Don't expect a creative-AI habit to cancel out a homework-AI habit. We don't have evidence either way on that, and anyone who tells you otherwise is further ahead of the research than it actually is.

The honest summary: one narrow, well-sourced finding says that AI used only to finish an assigned task, without design around it, inflates performance and erodes it when the tool is taken away. Amplify's Dan Meyer says he's seeing that pattern show up at home. Neither claim tells you to abandon homework help, and neither tells you that a creative AI tool is a fix. What they do tell you is worth checking: what is the only thing your kid currently does with AI, and is it a task someone else assigned, or one they chose and have to answer for themselves?