MIT and Oxford: even 10 minutes with AI reduce persistence and the quality of independent work

Researchers from MIT, Oxford, Carnegie Mellon, and UCLA have shown that even 10–15 minutes of work with an AI assistant can worsen independent problem-solving. In the short term, a chatbot improves results, but after being turned off, people more often give up, perform worse, and less frequently complete tasks.

How the Experiment Proceeded

The authors conducted a series of three randomized experiments with a total of 1,222 participants. They tested not abstract user feelings, but behavior on specific tasks: arithmetic with fractions and reading comprehension. In the experimental groups, participants were given an assistant based on GPT-5, which was then suddenly disabled, and researchers observed how results changed without external support. Importantly, researchers measured not just the number of correct answers, but willingness to continue working. To do this, they tracked how often a person skips a task instead of attempting to solve it. The authors interpret such skipping as an indicator of reduced persistence: the participant not only makes mistakes, but stops engaging with the task.

• In the first experiment, after the AI stage, participants solved three more tasks without assistance.
• In the second, the design was repeated on a larger sample and a preliminary test was added to eliminate bias in preparation level.
• In the third, text comprehension tasks from SAT training materials were used instead of mathematics.
• In all three cases, the behavior of people with and without AI was compared under identical conditions.

What the Tests Showed

The picture turned out to be consistent. While the assistant was available, the AI group solved more tasks and more rarely gave up. But as soon as access was removed, the metrics dropped sharply: people answered worse independently and more often pressed the skip button.

The same effect appeared not only in arithmetic but also in reading, meaning the problem is not limited to one academic discipline. This is an important detail: the work does not prove long-term "brain degradation" in the medical sense and does not directly measure neural activity. Instead, it provides direct behavioral evidence that even a short session with an AI assistant can weaken the quality of subsequent independent work.

For research on AI's impact, this is much stronger than typical surveys where people simply report their feelings.

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

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

When the Harm Is Stronger

The most interesting finding emerged in the second experiment. After the task, participants were asked how exactly they used the assistant. 61% said they mainly asked for ready-made answers.

27% used AI for hints and explanations, and 12% barely used it. It was precisely the group that obtained direct solutions from the model that later showed the worst results without AI and most frequently skipped tasks. Those who asked not for answers but for explanations or hints showed noticeably better results.

The authors directly point out: the risk is associated not simply with the fact of using AI, but with the form of delegating thinking. If the model thinks for you, independent work becomes harder after it's turned off. If it helps you understand rather than substitutes for your effort, the negative effect is noticeably weaker.

As an explanation, researchers propose two mechanisms. The first is a shift in effort norms: when AI solves tasks in seconds, work without it begins to seem subjectively too difficult and slow. The second is the disappearance of "productive friction," through which people not only learn but also understand their own capabilities.

Without this experience, both the skill and confidence that the task is worth pushing through are lost.

What This Means

For education, analytics, programming, and any other intellectual work, the conclusion is simple: AI is useful as an assistant, but dangerous as a constant substitute for personal effort. While this is a preprint and deals with short tasks rather than months of office work, the practical conclusion is already clear: first formulate an idea or try to solve the task yourself, and only then engage AI for checking, explaining, or refining.