Berkeley, California — October 9, 2026.
Researchers at UC Berkeley and four other institutions have found that using ChatGPT for about ten minutes made people measurably less willing to keep working through hard problems once the AI was taken away — not just worse at them, but quicker to give up. If you use AI heavily at work or in school, this one deserves your attention.
What the researchers did
The team, led by Brian Christian, a research fellow at UC Berkeley’s Center for Human-Compatible AI, with co-authors from Carnegie Mellon, MIT, Oxford, and UCLA, ran three randomized controlled trials with a total of 1,222 participants. The paper, titled “AI Assistance Reduces Persistence and Hurts Independent Performance,” was presented this week at the Conference on Language Modeling.
In the first experiment, 354 participants worked through 15 fraction problems. One group worked alone; the other had ChatGPT open on the side and could ask it for help — or for the answer outright. The AI group started off more accurate, exactly as you’d expect. Then, after 12 problems, the researchers took the AI away. Almost immediately, that group’s accuracy collapsed. A second, larger experiment with 667 participants showed the same pattern. A third experiment with 201 participants, this time on SAT reading comprehension, echoed the results: when the AI was removed, persistence and accuracy dropped.
What the data showed
The striking part wasn’t just that performance fell. Participants who had used AI became significantly more likely to skip questions altogether — they stopped trying. The group that never had AI help stuck with the problems and finished stronger. As Christian put it, “You almost couldn’t ask for a clearer story.”
Christian frames the issue as a question about what “helpful” really means: “We’ve built these systems to be helpful, but are they being helpful to us in a way that’s actually helping us?” He argues the goal should be AI that augments human ability rather than replacing it — preserving what the team calls “productive struggle,” the effort through which people discover their strengths.
Why it matters
Plenty of workplaces now hand employees an AI copilot for everything from drafting emails to writing code. This study suggests there’s a hidden cost: short-term productivity gains may come with long-term erosion of the very skill that makes people good at their jobs — sticking with hard things. The practical takeaway isn’t to ban AI, but to be deliberate about when you lean on it. If the AI does the thinking, make sure you’re still practicing yours.
Frequently asked questions
Does this mean using AI makes you dumber?
Not exactly. The study measured persistence and independent performance, not intelligence. AI-assisted participants performed better while the tool was available — the damage appeared when the AI was removed. The concern is dependence, not raw ability.
Who led the study?
Brian Christian, a research fellow at UC Berkeley’s Center for Human-Compatible AI and author of The Alignment Problem, led the team, with co-authors from Carnegie Mellon, MIT, Oxford, and UCLA.
Where can I read the research?
The paper, “AI Assistance Reduces Persistence and Hurts Independent Performance,” was presented at the 2026 Conference on Language Modeling. UC Berkeley covered the findings on October 9.
How long did the AI exposure last?
About ten minutes — roughly the time it took participants to work through the first 12 problems with ChatGPT’s help before the tool was removed.
Sources: UC Berkeley News, AI Weekly.