Key takeaways
- Three randomized trials with 1,222 participants found AI help raised accuracy at first but hurt performance once it was removed.
- People who used ChatGPT for about 10 minutes skipped or quit hard problems more often than those who never had AI help.
- Researchers urge tutor style AI defaults, and users can protect persistence by trying first and asking for hints, not answers.
A new AI persistence study from researchers at UC Berkeley and partner universities suggests that leaning on a chatbot for as little as 10 minutes can make people quicker to give up once the help goes away. Across three randomized experiments with 1,222 online participants, people who had ChatGPT on hand solved problems more accurately at first, then stumbled and quit more often when the tool was taken away. The finding is getting wide attention this week because it lands on a question millions of students and workers face every day: does AI make us better at our work, or just better while the AI is switched on?
What the AI persistence study found
The paper is titled "AI Assistance Reduces Persistence and Hurts Independent Performance." It is posted on arXiv, and according to Berkeley News, an updated version was presented at the Conference on Language Modeling (COLM) in early October. The arXiv listing names five authors: Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker and Rachit Dubey. Christian, a research fellow at Berkeley's Center for Human Compatible AI, is best known as the author of the 2020 book *The Alignment Problem*. Berkeley News says the team also included scholars from Carnegie Mellon, MIT, Oxford and UCLA.
The headline result is simple. AI help raised performance in the moment. But once the help disappeared, the people who had used it did worse than people who never had it, and they were more likely to stop trying.
How the experiments worked
Berkeley News describes three trials:
- Experiment 1 (354 people): Participants worked through 15 basic fraction problems. One group had no assistance. The other had ChatGPT open and could ask it for answers. After 12 problems, the AI was removed.
- Experiment 2 (667 people): A larger repeat of the fraction task, with the same setup of taking AI away partway through.
- Experiment 3 (201 people): SAT style reading comprehension questions, again with the AI pulled partway through.
In each case, the AI group started out ahead. After the tool was withdrawn, their accuracy fell quickly. In the second experiment, according to Berkeley News, the AI group gave more wrong answers or simply quit, while the group that never had help kept going and finished more successfully. Skipping rose sharply in the AI group after the switch. The reading test in the third experiment showed the same pattern for both accuracy and persistence.
The short time frame is what makes the result striking. These were not people who had relied on chatbots for months. The effect showed up after roughly 10 minutes of use.
Why it matters
Persistence is one of the quiet engines of learning. You get better at hard things partly by sitting with them when they feel slow and frustrating. The authors argue that this is exactly what quick AI answers can short circuit.
In the paper's summary, the researchers say AI may be training people to expect immediate answers. Once that expectation sets in, slower and more effortful work starts to feel inefficient, even when it is the only option. Berkeley News also notes that the team links reliance on AI to the loss of "productive struggle," the kind of effort that helps people figure out what they are good at and what they enjoy.
Christian put the stakes plainly in the Berkeley News piece: "This is not a story about fractions and SAT problems." He also asked a pointed question about the design of today's assistants: "We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us?"
His concern reaches beyond classrooms. According to Berkeley News, he worries in particular about science and academia, where overreliance could weaken researchers' closeness to their data, their immersion in the material and their teamwork.
Who is affected
Students and parents
Students are the most obvious group. Homework help from a chatbot can feel like progress, and in the moment it often is. But exams, interviews and real projects tend to arrive without an AI sitting beside you. If a few minutes of easy answers lowers the urge to push through a hard problem, that cost will show up later. Parents weighing tools like a teen version of ChatGPT may want to read our coverage of Common Sense Media's warning on ChatGPT for teens alongside this study.
Workers and managers
Office workers are in a similar spot. A TMC Insight analysis of the study points to Microsoft's 2025 Work Trend Index, which surveyed 31,000 workers in 31 markets and found that 75% used AI at work. As assistants and AI agents take on more routine tasks, companies may be measuring the output produced with AI while missing what happens to people's skills when the tool is not available. TMC Insight suggests that employers track independent performance and retention, not just finished output.
People building AI tools
The study also puts pressure on developers. Christian argues that assistants should default to instructive, tutor style responses rather than handing over the answer. "We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted," he said, according to Berkeley News.
What the study does not show
It is worth being careful here. TMC Insight stresses that the research does not show permanent cognitive harm, and it does not show that every use of generative AI weakens human ability. The tasks were short and fairly narrow: fractions and reading questions in an online setting. Open questions remain, including whether the effect fades over time, whether it varies by task, and whether experienced users respond differently.
The study also measured what happened right after the AI was removed. It did not follow people for weeks or months. So the fair reading is that brief AI use can change how people approach the next hard problem in the short run, and that this deserves more research before anyone draws sweeping conclusions.
How to use AI without losing your grit
None of this means you should stop using chatbots. It does suggest using them more deliberately. A few habits drawn from the researchers' advice and the TMC Insight analysis:
- Try first, then ask. Make a real attempt or write a rough draft before you open the assistant. You keep the struggle and still get a check on your work.
- Ask for hints, not answers. Tell the chatbot to explain the next step or point out where your reasoning went wrong, instead of solving the whole thing. Our guide to writing better AI prompts can help you phrase those requests.
- Critique the output. Treat the AI's answer as something to evaluate, not something to copy. Look for errors and gaps.
- Work without it on purpose. Set aside regular stretches where you solve problems with no assistant at all, especially for skills you need on your own.
- Test what stuck. Students and trainers can use delayed recall quizzes to see what someone actually learned, not just what they produced with help.
For workplaces, the same idea applies at a bigger scale. If AI is shaping how people are evaluated, there is a strong case for transparency, a point we made in our opinion piece on why workers deserve to know when AI judges them.
What to watch next
The bigger question is whether AI companies will change how their products behave by default. Several assistants already offer study or tutor modes, but the default experience is still built to deliver fast, complete answers. Research like this gives educators and regulators a concrete data point when they ask for learning focused designs, especially for younger users.
Expect follow up studies too. The authors call for AI development that supports long term competence, not only finishing the task in front of you. Longer trials, workplace studies and tests across different subjects would show whether the 10 minute effect holds up, grows or fades. You can follow more of this coverage on our AI news page.
Frequently asked questions
What did the Berkeley AI study find?
Across three experiments with 1,222 participants, people who used ChatGPT for about 10 minutes did better at first but performed worse and gave up more often once the AI was removed, compared with people who never had AI help.
Who ran the study?
The paper lists Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker and Rachit Dubey. Berkeley News says the team included researchers from UC Berkeley, Carnegie Mellon, MIT, Oxford and UCLA.
Does this mean AI harms your brain permanently?
No. The study measured short term effects right after AI was taken away. It does not show lasting cognitive damage, and the researchers and commentators say more work is needed.
How can I use AI without losing persistence?
Attempt problems on your own first, ask the chatbot for hints instead of full answers, check its output critically, and practice regularly without any assistant.
Sources
Share this story
Spotted an error? Tell us
