MIT and Oxford: Even 10 Minutes With AI Reduces Persistence and the Quality of Solo Work

Researchers from MIT, Oxford, Carnegie Mellon and UCLA have shown that even 10–15 minutes of work with an AI assistant can impair independent problem solving. In the short term, the chatbot improves results, but after it is switched off, people give up more often, perform worse and are less likely to finish the task.

How the experiment worked

The authors ran a series of three randomized experiments with a total of 1,222 participants. They tested not users' abstract impressions but behavior on concrete tasks: arithmetic with fractions and reading comprehension. In the experimental groups, participants were given an assistant based on GPT-5, which was then suddenly switched off, and the researchers watched how results changed without outside support.

Importantly, the researchers measured not only the number of correct answers but also willingness to keep working. To do this, they tracked how often a person skipped a task instead of attempting it. The authors treat such skipping as an indicator of reduced persistence: the participant is not simply making mistakes but ceasing to engage with the task.

• In the first experiment, after the AI stage, participants solved three more tasks without the assistant.
• In the second, the design was repeated on a larger sample, with a pre-test added to remove bias in preparation level.
• In the third, reading comprehension tasks from SAT practice materials were used instead of math.
• In all three cases, the behavior of people with and without AI was compared under identical conditions.

What the tests showed

The picture proved consistent. While the assistant was available, the AI group solved more tasks and gave up less often. But as soon as access was removed, performance dropped sharply: people answered worse on their own and pressed the skip button more often. The same effect appeared not only in arithmetic but also in reading, so the problem is not limited to a single subject.

This is an important detail: the study does not prove long-term "brain degradation" in the medical sense and does not measure neural activity directly. It does, however, provide direct behavioral evidence that even a short session with an AI assistant can weaken the quality of subsequent independent work. For research on the effects of AI, this is much stronger than ordinary surveys in which people simply describe how they feel.

"If even a brief interaction produces a measurable decline, then daily

use of AI over months or years may have serious consequences," the authors write.

When the harm is greater

The most interesting finding came from the second experiment. After the task, participants were asked how exactly they had used the assistant. 61% said they mostly asked it for ready-made answers. 27% used AI for hints and explanations, and 12% hardly used it at all.

The group that took direct solutions from the model later performed worst without AI and skipped tasks more often than the others. For those who asked not for an answer but for an explanation or a nudge, the picture was noticeably better.

The authors state plainly that the risk lies not simply in using AI but in the form of delegating thinking. If the model thinks for you, independent work is harder once it is switched off. If it helps you understand rather than replacing the effort, the negative effect is noticeably weaker.

The researchers offer two mechanisms as an explanation. The first is a shift in the effort norm: when AI regularly solves a task in seconds, work without it begins to feel too hard and slow. The second is the disappearance of "productive friction," through which a person usually not only learns but also understands their own capabilities. Without this experience, both the skill and the confidence that a task is worth pushing through are lost.

What this means

For education, analytics, programming and any other knowledge work, the takeaway is simple: AI is useful as an assistant but dangerous as a permanent replacement for your own effort. For now this is a preprint and concerns short tasks rather than months of office work, but the practical conclusion is already clear: first formulate an idea or try to solve the task yourself, and only then bring in AI for checking, explanation or refinement.