UC Irvine study links ChatGPT access to weaker math learning in students

What the researchers looked at

Researchers at the University of California, Irvine wanted to know whether easy access to ChatGPT changes how much students actually learn. As KCRA reported, they teamed up with McGraw Hill, an education company, and analyzed 3.2 million math questions that students had answered in the company's tests.

The questions were fairly simple, the kind of high-school-level problems that AI chatbots solve easily. That choice makes sense: if a tool can answer a problem in seconds, it is the place where shortcuts are most tempting. Lead author Sina Rismanchian said the team was checking whether students' learning behavior changed after ChatGPT was introduced.

The data ran across a decade, from 2015 to 2025. That let the team compare the years before ChatGPT existed with the years after, so the earlier students served as a baseline for how people normally worked through these problems.

Faster, but with a catch

The first result was about speed. Once students had access to AI, they spent significantly less time solving the problems. For high school students the shift was especially large: they needed 30% less time to finish the test than the students before them, who had no AI to lean on.

That sounds like a benefit, but there is a problem. Those tests were not supervised, so nobody could see how the answers were produced. Finishing sooner could mean a student has mastered the material, or it could mean a chatbot did the work.

To tell the difference, the researchers compared against proctored tests, which are supervised and where students did not have AI. There the pattern turned. Starting in the year ChatGPT became available, students got many more of the problems wrong. It appeared that the students who used AI to move quickly had not learned to solve those problems themselves.

Why the struggle matters

Rismanchian pointed to an old idea in learning research: people need to wrestle with a problem, and work through the thinking it takes, in order to learn it. When a chatbot supplies the answer, that effort disappears.

In his words, less struggle comes at the cost of what he called durable learning outcomes. In plain terms, that means knowledge that is still there later, when you sit down for a test and no one is helping.

A rough comparison is a gym where a machine does the lifting for you. The work gets done and the session looks productive, but your muscles are no stronger afterward. The study suggests something similar can happen with math when the thinking is outsourced.

AI can still help, depending on use

The researchers did not conclude that AI has no place in education. Rismanchian said it can enhance learning when used appropriately. One example is a tool that gives students feedback, such as analyzing an essay the student wrote and suggesting improvements. The student still does the hard thinking first, and the AI responds to it.

He summed it up this way: "it depends on the way that the students are using these tools." KCRA's report frames the takeaway as using AI to support learning rather than to replace the problem-solving process.

The difference is practical. Asking a chatbot to solve a problem hands over the struggle. Asking it to check your attempt, or to quiz you, keeps the struggle with you.

What the report leaves open

The KCRA report is a short summary, and it leaves some questions unanswered. It does not say how many students were involved, and it singles out high school students for the 30% figure without giving the same number for other groups. It also does not describe how the researchers accounted for anything else that might have changed over the decade.

The wording is cautious too: the data showed more wrong answers on proctored tests, and it "appeared" that AI-assisted speed meant less real learning. Treat this as a strong signal worth taking seriously rather than the final word. KCRA links to the full study for readers who want the details.

For teachers, parents and anyone training staff, the message is not to ban the tools. It is to notice when they replace practice, and to check understanding in settings where the tool is not available.