Does AI Make Kids Worse at Math? What the Research Actually Shows
Brian Mwangi
Every new tool in math class has come with the same fear. In the 1970s it was the pocket calculator: children would never learn their times tables. The fear was partly right and partly wrong, and the way it resolved is a useful guide to AI.
A large 1986 meta-analysis of calculator studies found that calculators did not harm basic skills when they were used alongside regular instruction, and often improved problem-solving. What mattered was how they were used. A calculator that frees a student to think about a hard problem helps. A calculator that replaces the thinking the lesson was meant to build doesn't.
AI is the same question with much higher stakes, because AI can replace far more of the thinking than a calculator ever could.
Two studies, opposite results
Two of the most rigorous studies of AI in math and science education point in opposite directions. Reading them together tells you almost everything you need to know.
The study where AI hurt
In 2025, a team led by researchers at the Wharton School published a randomized field experiment in PNAS with nearly 1,000 high school students in Turkey. Students were split into three groups for their math practice sessions:
- No AI: the usual practice problems.
- GPT Base: a standard chat interface to GPT-4, similar to ChatGPT.
- GPT Tutor: GPT-4 with guardrails. It had the teacher's solutions and was instructed to give hints rather than answers.
During practice, both AI groups did much better than the no-AI group. Then came a test without AI. The GPT Base students scored 17% worse than students who had never used AI. The GPT Tutor students roughly matched the no-AI group.
The researchers looked at the chat logs. Many GPT Base students had simply asked for the answer. They performed well while the AI was there, and had learned less once it wasn't.
The study where AI helped
Also in 2025, a Harvard team published a randomized study in Scientific Reports comparing an AI tutor with an active-learning physics class, itself already a strong teaching method. The AI tutor was carefully designed: it followed good teaching practice, guided students through problems step by step, and avoided simply giving away solutions.
Students using the AI tutor learned more than twice as much in less time, and reported feeling more engaged and motivated.
What separates them
Both studies used the same underlying technology. The difference was design:
| AI that hurt | AI that helped | |
|---|---|---|
| Default behavior | Answers the question | Guides toward the answer |
| Who does the work | The AI | The student |
| When student is stuck | Gives the solution | Gives a hint |
| Design goal | Be helpful | Make the student learn |
Why answers make learning worse
Cognitive science has names for what's going on.
The generation effect. People remember information better when they produce it themselves than when they read it. This has been replicated for decades. An answer your child works out sticks; an answer they read fades.
Desirable difficulties. Some struggle is how learning happens. Robert Bjork's research shows that conditions which make practice feel harder, like retrieving from memory instead of re-reading, often produce much better long-term retention. An AI that removes all difficulty removes the learning with it.
Cognitive offloading. A 2025 study of 666 participants published in Societies found that people who relied more heavily on AI tools showed lower critical thinking scores, and that offloading thinking to AI explained much of that link. The study shows a correlation rather than proving cause, but the pattern is consistent with the experiments above.
The common thread: the effort your child spends on a problem isn't a cost to be minimized. It's the thing that produces learning.
Warning signs AI is making it worse
You don't need a study to see this at home. Watch for:
- Homework grades rising while test grades fall.
- Your child can't explain a method they "used" last night.
- Frustration at being asked to try before getting help.
- Speed without understanding: finishing fast but freezing on a slight variation.
We cover these in detail in signs your child is using AI for homework.
How to make AI make your child better at math
The research gives a clear recipe. Whatever tool your child uses, aim for these conditions:
- Attempt before assistance. Your child should try every problem before the AI says anything.
- Hints, not answers. When stuck, the next step should be a nudge, not the solution.
- Explanation, not just procedure. Understanding why a method works is what lets your child use it on a new problem.
- Practice without the tool. Regularly, your child should solve problems with no AI present, the same way they will on a test.
Why we built PennPaper this way
PennPaper is built on the design that produced the Harvard result, not the one that produced the 17% drop.
By default, the tutor asks your child to attempt each step before it helps, and it doesn't hand over final answers. When your child is stuck, it gives a hint or draws the idea on the whiteboard, so they can see why a step works. Sessions are 15 minutes, which keeps practice focused. And the tutor never grades or ranks your child. Its only job is to explain until the idea clicks.
AI doesn't have to make kids worse at math. But it will, if it's designed to answer instead of to teach.
Frequently asked questions
Does using AI hurt children's learning?
It can. A 2025 randomized study found that students who practiced with an unrestricted chatbot scored 17% lower on a later test without it. Students using an AI designed to give hints avoided that harm.
Can AI tutoring actually improve math scores?
The evidence is promising when the tutor is well designed. A 2025 Harvard study, in physics rather than math, found students learned more than twice as much with a carefully designed AI tutor than in an active-learning class. The same design principles apply to math.
Is AI like calculators in math class?
It's a similar question: tools help when they support thinking and hurt when they replace it. AI can replace far more of the thinking than a calculator, so how it's used matters more.
How much AI use is too much?
It's less about time and more about who does the work. If your child can't solve similar problems without the tool, the balance is off.
The bottom line
AI is not good or bad for math learning. Answer-first AI tends to hurt it. Teaching-first AI can help it a great deal. When you choose a tool, choose the one that makes your child do the thinking.
Sources: Bastani et al. (2025), Generative AI without guardrails can harm learning, PNAS; Kestin et al. (2025), AI tutoring outperforms in-class active learning, Scientific Reports; Gerlich (2025), AI Tools in Society, Societies; Hembree & Dessart (1986), Journal for Research in Mathematics Education; Slamecka & Graf (1978), The generation effect; Bjork (1994), Memory and metamemory considerations in the training of human beings