The 1 in 5 AI Users Who Kept Learning

Picture five students on homework night, all with AI one tap away.

Four of them finish fast. The answers look polished. The scores come back excellent. Notebooks close before dinner.

One student is still sitting with the problem.

They spend about as long on the assignment as classmates who never touched AI. Slower, yes. But still doing the work that turns an answer into something they actually know.

When exam day comes, that quiet student holds steady.

A study of 26,811 students in grades 7 to 12 in one county in central China tracked school records over 30 months. The researchers, David Strömberg, Victor Lei and Yanhui Wu, found a striking pattern after students began using AI (CEPR Discussion Paper 21577, summarized in a VoxEU column).

Within six months, homework scores rose 18% and homework time fell about 30%, from about 64 minutes to about 45 minutes per assignment. Monthly closed-book exam scores fell 20%.

We covered the headline result in AI Raised Their Homework Scores. Their Exam Scores Fell.. The short version: homework can look better while learning gets weaker.

But one group did something different.

What the 1 in 5 kept doing

About 4 in 5 AI users finished homework quickly with very high homework scores, a pattern consistent with outsourcing. Their exam scores were markedly worse.

The remaining 1 in 5 kept spending roughly as much time on homework as non-users. Their exam scores held.

In the researchers' words: "when AI students spend the same amount of time on homework as non-AI students, they learn as much."

Among AI users overall, higher homework scores went with worse exam scores. A sudden jump in the homework grade was not the good news it looked like.

The difference was not access to AI. It was what happened after AI entered the assignment. Some students used it to remove the struggle. Others kept enough of the struggle to stay in the problem.

The authors are careful here. The study does not show that requiring more homework time by itself restores learning, and the students who kept spending time may have differed in motivation or skills. Time is not a magic switch.

Still, the pattern points to a practical question: how can you use AI without handing over the part that builds your ability?

Our suggestion is simple. Try first. Put your own answer, outline, calculation, or explanation on the page before asking AI for help. If you are stuck, ask for a hint or an explanation. Ask for a quiz instead of the answer. Then explain the idea back in your own words.

Those habits are our suggestions, not something the study measured. They protect the part of learning a polished answer cannot replace: your own attempt.

Adults face the same choice

This is not only a student problem.

At work, the same fork appears whenever AI can produce something faster than we can learn to do it ourselves. Accept the finished result and move on, or use AI without skipping the skill.

Our AI and judgment survey found that overall, 38% have chosen not to learn something because AI could do it, and another 21% considered it but learned the skill anyway.

That 21% faced the same kind of choice as the 1 in 5 students. Different people, different settings, and no claim that one explains the other. But the decision is familiar: let the tool take over, or stay involved long enough to understand what is happening.

You know the moment. You need to write something, learn a process, or make a decision, and AI hands you a result that looks better than your first attempt.

But if you never make the attempt, you will not know what to ask next time. You will miss the error. You get the output without the judgment that makes it useful.

So the question is not only "Can AI do this?" It is "What part of this do I still need to know how to do?" We dig into that in the skill you skipped.

Meet Vesela. Start your first conversation.
Get Vesela: AI Thought Partner on the App Store  |  Get it on Google Play

Keep the thinking in the room

The goal is not to reject AI. It is to use it in a way that leaves you with more than a completed task. Start with your own effort, use AI to open the next step instead of erasing the whole process, then put the idea into your own words.

That can feel slower. It is also the difference between recognizing an answer and knowing how to reach one.

The 1 in 5 are the hopeful part of this story. Same access to AI, same homework, and their exam scores held. Using AI does not have to mean outsourcing the thinking.

That is why our team at Gray Sky AI builds Vesela as a thought partner that asks questions so people keep doing the thinking. If you want to explore that idea further, read Trust Your Judgment.

FAQ

Does AI make students worse at learning?

Not every student. In the study of 26,811 students, homework scores rose 18% and monthly closed-book exam scores fell 20% within six months of starting AI. The 4 in 5 who finished fast did markedly worse on exams. The 1 in 5 who kept spending about as much time as non-users held their exam scores.

What did the 1 in 5 do differently?

They kept spending roughly as much time on homework as non-users, and their exam scores held. The study does not say exactly how they used AI, and the authors note these students may have differed in motivation or skills.

How can adults use AI without skipping the skill?

Try the task yourself first. If you are stuck, ask AI for a hint or an explanation instead of the answer. Ask it to quiz you, then explain the result back in your own words. These are suggestions from our team, not behaviors measured in the study. The aim is to keep your own judgment in the work.

Want a thought partner that asks questions so you keep doing the thinking? Meet Vesela at vesela.ai.

Meet Vesela: AI Thought Partner. Start your first conversation.
Get Vesela: AI Thought Partner on the App Store  |  Get it on Google Play