Ten minutes with ChatGPT weakens persistence. What does this mean for teams implementing AI?

Researchers from five universities tested on 1,222 people how brief work with a chatbot affects independent problem-solving. The results matter for anyone implementing AI in a team and developing junior employees.

Companies buy licenses for AI assistants because they expect faster work. A new study shows a cost that is rarely discussed during implementations. After about ten minutes of using a chatbot, people who had it taken away solved tasks worse and gave up more often. The most was lost by people who asked AI for ready-made answers. Above all, the study offers guidance on how to set up AI in a team.

What exactly the researchers tested

The paper “AI Assistance Reduces Persistence and Hurts Independent Performance” was created by a team of Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, and Rachit Dubey from Carnegie Mellon University, UC Berkeley, MIT, Oxford, and UCLA. It was published as a preprint on arXiv (opens in a new tab) (a version shared before formal journal peer review) and presented at COLM 2026 (Conference on Language Modeling). The study included three experiments and 1,222 participants from the U.S. recruited online.

In the first experiment, 354 people solved 15 fraction problems. Some had access to ChatGPT, which could give hints or provide results. After 12 tasks, the chatbot was turned off. In the last three tasks, people who had previously used AI answered correctly in 57 percent of cases, while the control group, working independently from the start, did so in 73 percent. The first group also skipped 20 percent of tasks, the control group 11 percent. Participants received no bonus for good results and no penalty for mistakes, so skipping a task directly reflected willingness to put in effort.

The second experiment was repeated with 667 people and supplemented with an initial skills test. The result: 71 percent correct answers after AI was taken away versus 77 percent in the control group. The difference in skipped tasks was no longer statistically significant here for the groups as a whole. The third experiment, with 201 people, concerned reading comprehension in tasks similar to the American SAT exam: 76 percent correct answers in the post-AI group, 89 percent in the independent group, and a skip rate of 8 percent versus 1 percent. So the effect was not limited to mathematics.

What matters is how the employee asks AI

The most interesting data for managers comes from the second experiment. Among people with access to the chatbot, 61 percent said they used it mainly to get ready-made answers, while 27 percent mainly used it for hints and explanations. The former later achieved an average of 65 percent correct answers. The latter 76 percent, almost as much as the control group with 77 percent. People taking ready-made solutions also skipped tasks more often.

This comparison should be approached cautiously, because participants chose their own way of using AI rather than being randomly assigned to it. So the difference may also have reflected differences between people. The authors propose two explanations, which they themselves treat as interpretations. First: when answers appear instantly, independent work starts to seem disproportionately difficult. Second: when AI performs the key steps, the person does not go through the attempts, mistakes, and corrections from which they learn.

Brian Christian of UC Berkeley, a co-author of the paper, suggests in a university statement (opens in a new tab) that AI should work like a tutor who leads to the solution through hints.

What the study does not prove

The experiments lasted several minutes. The researchers did not measure brain activity or lasting changes in cognitive abilities. It is not known whether the effect lasts for hours, days, or weeks, nor whether years of using ChatGPT worsens memory or reasoning. The study shows behavior immediately after the tool is taken away. For a company, this matters because an employee regularly encounters situations without AI or with AI that makes mistakes and whose output must be assessed independently.

Managers are already seeing this problem in their own organizations

The laboratory results match what companies report. In a survey by consulting firm BCG (opens in a new tab) of 70 top executives, published on June 17, 2026, half of respondents admitted that they are already observing deskilling, meaning the erosion of competencies in the organization. More than 60 percent believe that within three to five years it will become a real threat. The areas identified as most at risk were judgment and decision-making, problem definition, creative thinking, and causal reasoning. Empathy, curiosity, and leadership turned out to be less threatened. The report’s authors describe teams that begin to treat AI output as “good enough” and completely skip the critical thinking stage.

A similar signal comes from an IBM study of about 1,500 HR directors and 8,800 employees, published in September 2026. According to CNBC’s report, three out of four employees concerned about AI say it is already straining their competencies, most often critical thinking.

Beginners have the most to lose. According to the AI Index 2026 report by Stanford HAI (Stanford University’s artificial intelligence research center), based on payroll data from HR and payroll company ADP, employment of software developers aged 22-25 has fallen by nearly 20 percent since the end of 2022, while that of developers in their thirties at the same companies has risen by 6-12 percent. This concerns occupations most exposed to AI. If there are fewer juniors, and those who are hired hand off the hardest parts of the job to AI, the company may in a few years have no one from whom to grow its own seniors and leaders.

How to set up AI in a team so it does not undermine skills

From the combination of the study and market data, we draw several editorial conclusions. They do not require giving up AI, only a conscious division of work:

• Separate routine from the competencies on which work quality depends: formatting, summarizing, or drafts can be delegated, but leave judgment and problem diagnosis to people.
• In onboarding (bringing a new person into the role) and with juniors, configure the assistant as a tutor: the system prompt, meaning the permanent instruction for the model, can forbid giving complete solutions.
• In tasks that require judgment, first require an independent attempt, and only then a consultation with AI.
• Include skills development in performance reviews, as BCG recommends, so speed with AI is not the only measured outcome.
• Watch whether the team still verifies AI outputs or accepts them without checking.

IBM is also moving toward coaching. According to CNBC, the company’s chief talent officer Aparna Nair advises managers to move away from a directive style in favor of supporting employees in building competencies, including preserving learning paths at the start of a career.

The balance is simple. AI speeds up work when it takes over repetitive tasks, and weakens a team when it takes over the thinking on which quality depends. Which version wins in a given company is determined by the rules and the configuration of the tools, not by the purchase of licenses alone.