AI Raised Their Homework Scores. Their Exam Scores Fell.

The homework is finished before dinner.

The answers look polished. The steps seem clear. The grade comes back higher than usual, and everyone at the kitchen table can breathe a little easier.

Then the exam arrives.

The room is quiet. There is no chatbot, no search bar, and no helpful prompt waiting in the corner. A question that looked familiar suddenly feels impossible. The homework said one thing. The test says another.

That gap is at the heart of a 2026 study about students using AI for homework. The study does not show that AI is automatically harmful. It shows what happens when the tool does the thinking instead of helping a student practice it.

What the study found

David Strömberg, Victor Lei, and Yanhui Wu tracked 26,811 students in grades 7 to 12 in one county in central China over 30 months. They examined school records across nine subjects. The research appears as CEPR Discussion Paper 21577 and was summarized in a VoxEU column.

By June 2025, roughly 80% of the students reported using generative AI, mostly general-purpose chatbots.

Within six months of starting to use AI, homework scores rose 18%. Homework time also fell by about 30%, from about 64 minutes to about 45 minutes.

That sounds like a clear win until the closed-book exams are included. Over the same six months, monthly closed-book exam scores fell 20%.

The decline on entrance exams happened more slowly. Over about two years, scores fell 24% on high-school entrance exams and 18% on college entrance exams.

The pattern was especially striking among about four in five AI users. They finished homework fast and had high homework scores, a combination consistent with outsourcing the homework. Their exam scores were markedly worse.

This is cognitive offloading in a practical form. The work moves from the learner to the tool. The finished page improves, but the learner does less of the remembering, judging, and problem solving that exams require.

Among AI users, higher homework scores went with worse exam scores. A sudden jump in homework grades can therefore be a warning sign rather than a win.

The problem is not that a student received help. The problem is that the student skipped the thinking that turns help into learning.

The part many headlines skip

The study also found a different group.

The other one in five AI users kept spending about as much time on homework as students who did not use AI. Their exam scores held up. The researchers write that "they learn as much."

That detail changes the question.

The useful question is not simply, "Does AI hurt learning?" A better question is, "What is the student still doing?"

A student can use AI to check a first attempt, ask for a simpler explanation, or practice with a new example. The student can also paste in a question and copy the first answer that appears. Both actions involve AI. Only one keeps the learning work in the student's hands.

AI is not the problem by itself. Skipping the thinking is.

Our team at Gray Sky AI builds Vesela as a thought partner for working through ideas, not for handing over the thinking.

This is not only about school

Adults run into the same choice at work and in daily life.

A message can be written in seconds, but that does not mean the sender knows what they want to say. A plan can sound organized, but that does not mean the tradeoffs have been considered. An answer can feel convincing before personal judgment has had a chance to form.

That is why the question matters beyond students and exams. The skill may stay hidden until the moment it is needed. The ability to explain, decide, remember, or spot a mistake develops through use.

Sometimes AI makes that practice easier. Sometimes it quietly removes the practice.

When a tool seems to make a task effortless, the skill you skipped is worth naming. When an answer sounds right but does not feel like your own reasoning, trusting your judgment becomes part of the work.

The goal is not to reject help. It is to notice which part of the work still belongs to you.

What you can keep doing yourself

• Try the problem first before asking AI for help.
• Ask AI to explain, give a hint, or quiz you instead of finishing the work.
• Explain the answer back in your own words.
• Notice when something feels effortless. That is a reason to slow down and check what you understand.
• If you are a parent, look at time spent and understanding, not just the grade.

A strong homework grade is useful. It is not the whole measure of learning.

The part worth protecting happens before the answer feels obvious: making a guess, getting stuck, changing direction, and finding out why something works.

Keep making that space for your own thinking. Use help to deepen the work, then keep practicing the part only you can do.

Vesela: AI Thought Partner helps you work through ideas without handing over the thinking.
App Store  |  Google Play