Berkeley study: 10 minutes of AI use erodes persistence

Ten minutes using ChatGPT was enough to measurably erode how long participants stuck with a hard problem once the AI was taken away, according to a study of 1,222 online participants reported by Berkeley News.

The paper, "AI Assistance Reduces Persistence and Hurts Independent Performance," was 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, and was presented at the Conference on Language Modeling. Across three randomized experiments, with 354 and 667 participants on fraction problems and 201 on SAT reading comprehension, the setup was the same: one group had ChatGPT access until the AI was taken away at problem 12. The AI-assisted group's accuracy dropped sharply once help was removed; the control group held steady.

"You almost couldn't ask for a clearer story," Christian told Berkeley News, describing the result as supplying "ammunition for a story that a lot of us kind of feel in our gut. I experience it too." Three researchers we follow flagged the Berkeley writeup within hours of publication.

Christian pushes the reframe on the word "helpful."

"We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us?" he said. "We showed, ironically, that they are often helping us in ways that are kind of unhelpful." His closing framing points past the lab: "We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted."