The experiment, and what it measured
A team led from the University of California, Berkeley has published causal evidence that short exposure to an AI assistant reduces how long people keep trying at a hard problem. Berkeley News reported the result on 9 October, when the peer-reviewed version of the paper was presented at the Conference on Language Modeling. A preliminary draft has been on arXiv since 6 April.
The design is a set of three randomised controlled trials with 1,222 online participants: 354 and 667 on sets of fraction problems, and 201 on SAT-style reading comprehension. In each, one group had ChatGPT available to help, up to and including asking for the answer. At problem 12, the assistant was taken away.
The abstract states the result plainly: “although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up”. The authors stress that the effects “emerge after only brief interactions with AI (approximately 10 minutes)”.
Not accuracy, but quitting
The distinction matters, because it is the part most easily lost in summary. The assisted group did better while it had the tool. What changed after the tool was removed was not only accuracy but willingness to continue: Berkeley News reports that the solve rate dropped sharply and the skip rate rose substantially, while the control group that had never had an assistant held steady and finished stronger.
The authors’ explanation is about expectation rather than ability. AI, they write, “conditions people to expect immediate answers, thereby denying them the experience of working through challenges on their own”. They frame the concern around learning: persistence, they argue, is foundational to skill acquisition and among the strongest predictors of long-term learning.
Their stated conclusion is a design one. The paper’s framing contrasts a mentor — who scaffolds, tracks progress and prioritises growth over the immediate result — with current systems, which it calls “fundamentally short-sighted collaborators”, optimised to answer instantly and completely and never to say no except for safety reasons. The recommendation is that model development should prioritise scaffolding long-term competence alongside completing the task in front of it.
Who wrote it, and what it does not show
The paper is by Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker and Rachit Dubey. Christian is a research fellow at Berkeley’s Center for Human-Compatible AI, and Berkeley News describes co-authors from Carnegie Mellon, MIT, Oxford and UCLA.
“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,” Christian told Berkeley News. “We’ve built these systems to be helpful, but are they being helpful to us in a way that’s actually helping us?”
The limits are the ones that apply to this kind of study. These are short online tasks with a scripted withdrawal point, not classrooms or workplaces, and the paper measures what happens minutes after the tool disappears rather than weeks later. It does not demonstrate lasting damage to learning, and it does not test whether a differently designed assistant — one that withholds answers, or asks questions back — avoids the effect. The experiments also use arithmetic and reading comprehension, tasks with a clear right answer, which is not what most AI use at work looks like.
What is genuinely new this week is the peer review and the conference presentation, not the finding itself, which has been circulating since the spring. What to watch is whether anyone replicates it over a longer horizon, and whether a lab builds the scaffolding assistant the paper argues for and measures the same thing again.
