4.5/5 - (37 votes)

Getting the right answer from an AI chatbot is not the same as learning how to get it. In a large classroom trial, students who could ask GPT-4 anything during practice solved more problems, then scored 17% lower than classmates with no AI once the tool was taken away.

What the researchers actually tested

The study, published in PNAS in 2025 by a team led by researchers at the University of Pennsylvania, ran at a large high school in Turkey during the fall 2023 semester. Nearly 1,000 math students in about fifty 9th, 10th and 11th grade classes took part in four 90-minute review sessions.

Each session had two halves. First, students worked on practice problems with one of three setups: no AI at all, a plain chat interface built on GPT-4 that worked much like ChatGPT, or a “tutor” version of the same model. The tutor was told not to hand out answers, and its prompt included teacher-written solutions and common mistakes, so it could guide students step by step with hints. Then everyone took a short exam on their own, with no tools.

Better practice scores, worse exam scores

During practice, AI helped a lot. Compared with the no-AI group, grades on the practice problems were 48% higher with the plain chatbot and 127% higher with the tutor version.

The exam told a different story. Students who had the plain chatbot did 17% worse on the unassisted exam than students who never had AI. The tutor group did about as well as the no-AI group: no harm, but no gain either, despite their much stronger practice scores.

See also  Can pink noise during sleep boost your memory? What research says

Why? The researchers looked at how students used the tools. With the plain chatbot, many simply pasted in the question or asked for the answer and copied it. They called this using the AI as a “crutch.” With the tutor, students more often asked for help or tried their own answer first. The plain chatbot was also unreliable on these problems: it gave a correct answer only 51% of the time. But the team’s analysis pointed to the copying habit, more than the errors, as the main reason scores dropped.

One more finding is worth sitting with. Students in the plain chatbot group did not think they had learned less or done worse. The tutor group believed they had done better on the exam, when they had not. In other words, it is hard to feel this cost while it is happening.

Why struggling first can pay off

This fits a wider line of research on what education researchers call productive failure. A 2021 meta-analysis in the Review of Educational Research pooled 53 studies comparing two orders: trying a problem first and getting instruction afterward, or getting instruction first and practicing afterward. Overall, trying first came out ahead, with a moderate effect, and the gap was larger when the struggle phase was well designed. The pattern did not hold for every group, though: for younger children in grades 2 to 5, instruction first did better.

The idea is an old one. The Socratic method works by asking questions until a person notices what they do not know yet. Asking an AI for the answer before you have wrestled with the problem can skip exactly that step.

See also  Many students hide their AI use, a 47-country survey finds

How to use AI without skipping the hard part

  • Attempt the problem first. Write down your own approach, even if it is wrong, before opening a chatbot.
  • Ask for a hint, not a solution. You can tell the AI directly: give me one hint and do not solve it for me.
  • Check its work. On these math problems, the plain chatbot was wrong almost half the time.
  • Test yourself without it. A practice quiz with no tools shows what you can really do on exam day.
  • Read your syllabus. AI rules vary by course, and your instructor’s policy is what counts.

What this study can’t tell you

The authors are clear about the limits. This was one subject (math), at one high school, in one country, using GPT-4 in fall 2023, when students were still new to these tools. College courses, essay writing and newer models may give different results. Each exam also came right after its practice session, so the study says little about what students remembered weeks later.

Still, the main lesson is simple and low risk to try: let the AI help you think, not think for you. If you are struggling in a course, your campus tutoring or learning center can help you build that habit with a real person.