Solve 15 basic fraction problems with ChatGPT helping you through the first 11 — then quietly take the assistant away at question 12. That setup drove the most-discussed experiment in a new UC Berkeley paper. On October 9, a multi-institution team from UC Berkeley, Carnegie Mellon, MIT, Oxford and UCLA presented the peer-reviewed version of the study at the Conference on Language Modeling (CoLM).
The experimental design is unglamorous and therefore credible. In the first trial, 354 participants were recruited online: one group solved the problems unaided, the other had ChatGPT open on the side and could even paste the answer. The AI group's accuracy stayed comfortably above the control group's through the first 11 problems — but the moment the AI was withdrawn, that orange line collapsed to the floor. The unaided group kept going, and even improved slightly by question 15. A second trial with 667 participants repeated the pattern almost exactly; a third trial with 201 participants swapped the math problems for SAT reading comprehension and got the same answer. 1,222 randomized participants in total.
This is not simply "laziness"
The authors prefer a sharper word: "productive struggle." They argue that the moments in problem-solving where you frown, change tack, and scrap a flawed approach are exactly what allow the brain to consolidate long-term competence. AI removes that struggle outright. In the short run it looks like saved time — but once the tool is taken away, users have neither the metacognitive feedback of "I have actually understood this" nor any of the new skill that struggle would have produced. That is why "the AI group, once deprived of the tool, got things wrong and gave up" — they never finished the struggle in the first place.
The paper's senior author, Brian Christian (author of The Alignment Problem), puts the result in a way the AI industry will not enjoy hearing: "We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us? Ironic — we showed they often help us in ways that are kind of unhelpful." Anthropic itself has already acknowledged, in a separate report this year, that offloading tasks to chatbots erodes baseline understanding.
A quiet rebuke of current product defaults
The blade here is not aimed at students. It is aimed at the designers of the current default behaviour of every major AI product. Today's frontier assistants are tuned to deliver "the final answer, fast" — which is exactly the behaviour Christian argues against. His proposed alternative is direct: AI should default to behaving like a tutor — guiding, questioning, leaving space — so that productive struggle is returned to the user. Read literally, that would force every "answer-first" agentic workflow to be re-sequenced, and turn long-thinking mode from a marketing flex into a baseline quality bar.
There is also a more uncomfortable audience: researchers themselves. Christian worries that over-reliance on AI tools in science jeopardises what makes a good scientist — staying close to data, being immersed in the material, working as part of a team. When a research group of ten shrinks to two or three people and a chatbot, "something significant is being lost."
So what
The method is unflashy: three randomized controlled trials, one paradigm, an experiment easy enough to replicate. The fact that this passed peer review at CoLM and was picked up by The New York Times, WIRED and Psychology Today means "AI is making people lazy" has graduated from folk suspicion to a published academic concern. With regulators starting to focus on "AI's erosion of human capability" — and the EU AI Act already edging in that direction — 2027 AI products will probably have to answer one small product question in public: is your default response an answer, or a coach?
Paper: arXiv 2604.04721, "AI Assistance Reduces Persistence and Hurts Independent Performance."