Homework Grades Up, Test Scores Down: What the New AI Research Means for Your Kid

A new study of 26,811 secondary students found that using generative AI raised homework scores by 18% and cut homework completion time by 30% — while closed-book exam scores fell 20% within six months. The clearest warning sign for a parent isn’t your kid’s grade. It’s homework that suddenly takes almost no time.

That’s a strange thing to have to say out loud. But it’s the most useful finding to come out of AI-in-education research this year, and it changes what you should be watching at the kitchen table.

What the study actually measured

The paper is called The Generative AI Learning Penalty, published in June 2026 by researchers at Stockholm University and the University of Hong Kong through the Centre for Economic Policy Research. They followed 26,811 students in grades 7–12 for 30 months across nine subjects, comparing homework scores and completion times against monthly closed-book exams and high-stakes entrance exams.

The gap between those two things is the whole story:

• Homework scores rose 18%. Completion time dropped 30%.
• Monthly exam scores fell 20% within six months of AI adoption.
• Entrance exam scores fell 18–24%, with the full penalty only showing up after about two years.

Homework got better. Learning got worse. And the delay is the cruel part — a parent looking at a report card in October would have seen improvement, and the bill didn’t come due until much later.

Worth noting the limits: this is one large study in Chinese secondary schools, an exam-heavy system that isn’t identical to a U.S. district. But it’s the largest, longest look anyone has published at the same kids’ homework and closed-book performance side by side, and the mechanism it identifies is not culturally specific.

The mechanism: it’s the clock, not the tool

Here’s the finding that should actually change your behavior. The learning losses were not spread evenly across AI users. They were concentrated in roughly 80% of AI users whose pattern looked like homework outsourcing — the researchers identified them by an unmistakable signature: exceptionally short completion times paired with high homework scores.

The other group — students who used AI but spent about as long on their homework as non-users — showed only small losses.

Same tool. Same subjects. Radically different outcomes, sorted almost entirely by whether the kid still did the thinking.

This is why “is my kid using AI?” turns out to be the wrong question. The useful question is “is my kid still spending the time?” A worksheet that used to take 40 minutes and now takes eight is telling you something that the grade on it cannot.

Why grades stopped being a reliable signal

Most parents are calibrated to read grades as a proxy for understanding. That calibration is now off, and this study shows exactly how far off: the assisted work and the understood work produce the same number on the page.

If you want the plain version — a kid can currently carry a B average and be a full grade level behind, and homework AI widens that gap rather than closing it. We wrote about that gap before this study came out, in why good grades don’t mean grade level. This is the research version of the same warning.

The parent anxiety here is already widespread. In Deloitte’s 2026 Back-to-School Survey of 1,207 K–12 parents, 49% said they worry their child relies on AI too much, while only 33% said their child’s school has guidelines for AI use — and 38% didn’t know either way. Half of parents sense the problem. Very few have anything concrete to look at.

The clock is something concrete to look at.

The industry noticed too — and that’s the real trend

On July 16, 2026, OpenAI added Study Mode to its parental controls, letting a parent with a linked teen account set it to switch on automatically whenever the teen starts a new chat. Study Mode, built with teachers and learning scientists, walks a student through a problem with guiding questions and structured explanations instead of handing over the answer.

That’s a genuinely good change, and it’s worth saying so. But notice what it concedes: the company with the most-used AI assistant on earth now ships a separate mode whose entire purpose is not giving your kid the answer — and lets parents force it on. The distinction between a tutor and an answer machine isn’t a marketing frame anymore. It’s a product setting, at the largest scale in the industry.

Which raises the obvious follow-up for a parent: is the safer behavior a mode you have to remember to enforce on a general-purpose tool, or the only behavior the tool has? That’s the difference between an AI tutor and letting your kid use ChatGPT, and this research is the first serious quantification of what’s at stake in the answer.

What to do this week

Four things, none of which require a confrontation:

1. Time one assignment. Not to catch anyone — just to know your baseline. You’re looking for a collapse, not a number.
2. Ask for the reasoning with the paper turned over. “Explain #4 to me without looking.” Thirty seconds, and it’s the closest thing to a closed-book exam you can run at home.
3. Say the quiet part out loud. Kids genuinely do not know that fast, correct homework can be costing them. Show them the 18%-up/20%-down finding. It lands better as physics than as accusation — and it pairs well with a conversation about using AI to learn versus using it to cheat.
4. Protect the struggle, not the grade. The students who came through this study intact are the ones who kept spending the time. That’s the whole intervention.

Where TutorBuddi sits in this

TutorBuddi is built Socratically on purpose: Buddi asks your kid what they tried, where it broke, and what they’d do next, rather than producing the finished answer. There’s no mode to enable — refusing to hand over the answer is the design, not a setting. That’s the same reason Buddi won’t write your kid’s essay.

The other half of this problem is visibility, and that’s what the parent dashboard is for. Because Buddi remembers your child session to session, it can show you where they actually got stuck and what finally clicked — the information a homework grade has now stopped carrying.

Homework scores were always a rough proxy. This year they became a broken one. The thing worth watching is whether your kid is still doing the thinking — and the good news is that’s visible, if you know to look at the clock.