WHAT THEY FOUND
● 1,222 participants across three randomised controlled trials — 354, 667 and 201.
● The AI group started out more accurate. Then the AI was taken away.
● Their solve rate dropped sharply and their skip rate rose. The no-AI group held steady or improved.
● Ten minutes was enough to produce the effect.
What they actually did
Three randomised trials. Participants were split into groups, one with ChatGPT available — including the option to simply ask for the answer. Partway through, the AI was removed, and the researchers compared what happened next.
| Experiment | Participants | Task |
|---|
| 1 | 354 | 15 basic fraction problems; AI removed after 12 |
| 2 | 667 | 15 basic fraction problems |
| 3 | 201 | SAT reading comprehension |
They measured three things: accuracy, persistence, and how often people skipped a question or gave up.
The headline result is not that the AI group was worse. While they had the tool they were better. The finding is what happened on the other side of the removal: solve rate down, skip rate up, against a control group that was flat or improving.
Who did it
The team includes Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI, with colleagues from Carnegie Mellon, MIT, Oxford and UCLA. The paper is "AI Assistance Reduces Persistence and Hurts Independent Performance" (arXiv 2604.04721). A preliminary draft circulated in spring 2026; the peer-reviewed version was presented at the Conference on Language Modeling in October 2026.
Christian, on the result: "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." And: "We showed, ironically, that they are often helping us in ways that are kind of unhelpful."
What it does and does not show
Worth being careful here, because this study is going to be over-quoted in both directions.
It does show: a measurable, randomised, replicated-across-three-trials drop in independent performance after short AI assistance on a bounded task, with a control group ruling out the obvious alternative explanations.
It does not show: any long-term effect. The tasks were fractions and SAT reading — bounded problems with correct answers, nothing like the open-ended work most people use AI for. And the design removes the tool, which is not the condition most people operate under. If the AI is always there, "performance without it" may be the wrong thing to measure.
The Berkeley write-up reports no numeric effect size and lists no limitations from the researchers, so the paper itself is the thing to read before leaning on this hard.
The useful version of this finding is narrow: if you need to be able to do something without the tool, practise doing it without the tool. That is not new pedagogy. It is just newly measured.
Why it matters commercially
Every AI product sold on productivity is implicitly claiming the user ends up ahead. This measures a case where the user ends up ahead while using it and behind afterwards — and the gap is invisible from inside the product, because the usage metrics look excellent right up until someone has to work unaided.
For anyone training staff on AI tools, the actionable read is about sequencing rather than restriction: build the skill first, add the assistance second. Reaching for the answer before trying is where the effect lives.
Sources