Using AI for just 10 minutes weakens your ability to solve problems on your own, study finds

Ask a chatbot for the answer and you will probably get it right — for now. But a new study finds that leaning on artificial intelligence for even ten minutes measurably weakens a person's ability to solve problems once the tool is taken away, and makes them more likely to simply give up rather than try.

The research, published as a peer-reviewed paper at the 2026 Conference on Language Modeling (COLM) and detailed in a UC Berkeley News release this week, is among the first to show a causal, not just correlational, link between brief AI use and a drop-off in independent effort. The team ran three randomized controlled experiments with a combined 1,222 participants, testing both basic math and reading comprehension.

What the experiments found

In the first experiment, 354 people recruited through the research platform Prolific solved fraction problems. Half had a chat assistant (described to them as ChatGPT) available for the first 12 problems; the other half did not. The assistant was then removed for everyone, and all participants worked three final problems alone.

Those who had used the AI solved just 57 percent of the unassisted problems, compared with 73 percent for the group that had worked without it the whole time — a statistically significant gap. The AI group was also nearly twice as likely to skip a problem outright rather than attempt an answer (a 20 percent skip rate, versus 11 percent in the control group).

A second experiment with 667 participants, designed to rule out skill differences between groups, replicated the pattern: solve rates fell from 77 percent to 71 percent once the AI was withdrawn. A third experiment, using 201 participants on SAT-style reading comprehension questions, found an even larger drop — from 89 percent to 76 percent — plus a higher skip rate.

• Experiment 1 (fractions, n=307 analyzed): unassisted solve rate 73% (no prior AI) vs. 57% (after AI access)
• Experiment 2 (fractions, replication, n=585 analyzed): 77% vs. 71%
• Experiment 3 (SAT reading comprehension, n=168 analyzed): 89% vs. 76%
• Effects were measurable after roughly 10 minutes of AI use

Crucially, not all AI use was equally damaging. When researchers asked participants in the second experiment how they had used the assistant, 61 percent said they asked it for direct answers, 27 percent said they used it mainly for hints or clarification, and 12 percent said they barely used it at all. The decline in performance and persistence was concentrated almost entirely in the "direct answer" group, who solved only 65 percent of unassisted test problems afterward, compared with 76 percent for the hint-seekers and 89 percent for those who skipped the AI — suggesting it is the habit of outsourcing thinking, not AI exposure itself, that does the damage.

Why researchers are concerned

The paper's authors — Grace Liu of Carnegie Mellon University, Brian Christian of UC Berkeley's Center for Human-Compatible AI, Tsvetomira Dumbalska of the University of Oxford, Michiel Bakker of MIT and Rachit Dubey of UCLA — argue the stakes go well beyond fraction worksheets.

Persistence, they write, is one of the strongest known predictors of long-term learning, and the skills tested in the study, basic arithmetic and reading comprehension, are developmental building blocks for everything from algebra to critical reasoning later on. The researchers propose two mechanisms: AI resets people's sense of how much effort a task "should" take, making unaided work feel harder by comparison, and it removes the "productive struggle" through which people learn not just the material but an accurate sense of their own abilities.

They frame the risk as a "boiling frog" problem — individually small, incremental losses of independent capability that are difficult to notice, and difficult to reverse, until they accumulate. The paper also raises an equity concern, cautioning that students with fewer academic resources and less supervision may be the most exposed to that slow erosion.

"You almost couldn't ask for a clearer story... This is not a story about fractions and SAT problems. This is something that is happening to human knowledge and human expertise from top to bottom," said Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI and a co-author of the study.

Who this affects

The findings land as AI chat tools have become routine homework aids for students of nearly every age, and as workplaces increasingly fold AI assistants into everyday tasks. The authors note their experiments captured only a single, brief session, so they cannot yet say whether the effect grows, shrinks or plateaus with months or years of regular use, or whether the specific impairment fades after a few hours. Separate research the authors cite, including Anthropic's own internal study of AI-assisted coding, has found related patterns among professional software developers, suggesting the effect is not limited to students or to math and reading tasks.

Christian said the research team designed the study specifically to move past anecdote. "It's a striking finding, but in a way it supplies ammunition for a story that a lot of us kind of feel in our gut," he said, adding that he recognized the pattern in his own habits: "I experience it too."

What happens next

The authors are careful to frame their work as a caution rather than a case for abandoning AI tools altogether. They note that participants who used the assistant for hints rather than answers showed little to no decline, and they argue the fix is not simply to use AI less, but to build systems that behave more like a good tutor — one that scaffolds a learner toward an answer rather than supplying it outright — instead of an "answer machine." User-side fixes such as built-in Socratic prompting or time limits, they write, are only "band-aids" unless AI developers also redesign systems around long-term competence rather than short-term satisfaction.

The Conference on Language Modeling, where the paper was presented this year, has become one of the field's main peer-reviewed venues for AI research, lending the findings more weight than the typical preprint. The authors say their next step is longitudinal research tracking whether the effects compound with repeated, everyday AI use over weeks or months — the question, they say, that will determine how seriously schools, employers and AI companies need to take the warning.