Berkeley Study: Brief AI Use Erodes Problem-Solving Grit

Ten minutes with ChatGPT is enough to measurably weaken a person's willingness to keep grinding at a hard problem once the AI is taken away, according to a UC Berkeley-led study reported by Berkeley News. Across three randomized controlled trials with 1,222 online participants solving fraction questions and SAT reading comprehension items, the groups given AI help showed immediate drops in accuracy and higher give-up rates the moment the tool was withdrawn, while the control group held or improved its performance.

Brian Christian, a research fellow with the UC Berkeley Center for Human-Compatible AI, co-authored the paper with collaborators from Carnegie Mellon, MIT, Oxford and UCLA. "You almost couldn't ask for a clearer story," he told Berkeley News of the results, which he framed as more than an academic curiosity: "It's a striking finding, but in a way it supplies ammunition for a story that a lot of us kind of feel in our gut. I experience it too."

The design was straightforward. A first experiment put 354 participants on fraction problems; a second, larger run used 667 participants on the same type of task; a third gave 201 participants SAT reading comprehension. After an initial run of problems, AI access was pulled, and the researchers watched what the formerly-assisted participants did next. The paper, "AI Assistance Reduces Persistence and Hurts Independent Performance," was presented at the Conference on Language Modeling in October 2026.

Christian's reading of the data is not confined to classrooms. "This is not a story about fractions and SAT problems," he said. "This is something that is happening to human knowledge and human expertise from top to bottom and from grade-school students all the way to the leading experts in the world." It lands into a growing AI ethics beat — 89 stories on our tracker over the last 90 days — where the question is increasingly whether assistive systems are actually assisting. Christian put it bluntly: "We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us? We showed, ironically, that they are often helping us in ways that are kind of unhelpful."