Study: 10 minutes with ChatGPT weakens persistence in independent work

After a short period of working with a chatbot, experiment participants made mistakes more often and gave up on tasks when the AI was taken away. The results differed depending on whether they asked for ready-made answers or for hints.

Researchers from Carnegie Mellon University, UC Berkeley, MIT, the University of Oxford and UCLA conducted three experiments involving 1222 people. They suggest that about 10 minutes of using ChatGPT worsens performance and persistence when people then have to continue working without it. The paper “AI Assistance Reduces Persistence and Hurts Independent Performance” has been available on arXiv since April as a preprint (opens in a new tab), that is, a preliminary version of a scientific publication. Its latest version, dated October 3, was presented at the COLM 2026 (Conference on Language Modeling) conference, dedicated to language models. On Friday, October 9, the study was discussed by the university outlet Berkeley News.

In the first experiment, 354 people solved 15 fraction problems. Some had access to a chatbot, which was turned off after the 12th task. In the last three tasks, the group that had previously used AI answered correctly in 57 percent of cases, while the control group did so in 73 percent. The first skipped 20 percent of the tasks, the second 11 percent.

The second experiment involved 667 people and an initial skills test. After the AI was taken away, the group that had previously used the chatbot gave 71 percent correct answers, while the control group gave 77 percent. The difference in skipped tasks was not statistically significant here. Among people with access to AI, 61 percent said they mainly asked for ready-made answers, and they later achieved 65 percent correct solutions. Another 27 percent asked primarily for hints and achieved 76 percent. The authors note that participants chose for themselves how to use the chatbot, so this comparison does not prove a causal relationship.

In the third experiment, 201 people solved reading comprehension tasks in the style of American SAT exams. The group with earlier access to AI achieved 76 percent correct answers, while the control group achieved 89 percent. Skipped tasks accounted for 8 percent and 1 percent respectively.

The study did not measure lasting changes in cognitive abilities. The effect was observed immediately after the AI was taken away, and it is not known whether it persists longer. The authors point to two possible explanations: instant answers make independent work seem disproportionately difficult, and people lose the opportunity to practice through trial and error.

For teams using AI assistants, the results suggest that the mode of working with the tool may matter - asking for an explanation instead of a ready-made solution. However, the experiments concerned school tasks, not professional ones.