AI assistants are getting better at answering questions. That is exactly the ability that may be quietly eroding another one. A peer-reviewed paper presented at the Conference on Language Modeling earlier this month, by a team spanning UC Berkeley, CMU, MIT, Oxford, and UCLA, reaches an uncomfortable conclusion: ten minutes of ChatGPT assistance is enough to make participants give up on hard problems almost immediately after the AI is taken away.
The paper first appeared on arXiv in April (arXiv:2604.04721, "AI Assistance Reduces Persistence and Hurts Independent Performance"), authored by Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, and Rachit Dubey. Across all three experiments, the combined sample reached N=1,222, a solid size for this kind of online behavioral study.
The same story, three times
To rule out task-specific quirks, the team ran the same basic design on three different problems. In each case, an "AI group" and a "no-AI group" worked through 12 problems with their assigned condition, then the AI tool was removed, and the researchers measured what happened next.
Experiment 1, with 354 participants, used 15 basic fraction problems. The AI group started with a clearly higher accuracy — unsurprising, since the model can usually produce the right answer. But once the AI was pulled after problem 12, the AI group's solve rate dropped off a cliff, and the skip rate climbed in lockstep.
Experiment 2 scaled up to 667 participants, with the same task structure. The result was just as clean: participants who had been relying on AI either got things wrong or gave up entirely once the helper was gone, while the no-AI control group held steady and finished stronger.
Experiment 3 swapped in 201 participants and an SAT reading-comprehension task. Without a fixed answer key, the AI could not just "spoil" the problem, yet the pattern repeated. The moment the AI was removed, persistence and accuracy both fell.
A single experiment with fraction problems could still be explained as people depending on ChatGPT's particular style of answers. SAT reading has no canonical solution, so the result there points to something deeper than a specific answer dependency.
What the paper actually argues
The authors avoid the easy framing of "AI is too smart" or "users are lazy." Their explanation is more structural. Today's AI systems, they write, are fundamentally short-sighted collaborators — optimized for instant, complete, never-refuse responses (barring safety concerns). The ideal human collaborator is more like a mentor: scaffolds learning, tracks progress, and is willing to say "don't look at the answer yet."
When a system trained to deliver answers in seconds intervenes for only ten minutes, people's tolerance for slower, deliberate work shifts. Brian Christian put it more directly in UC Berkeley's official write-up: when you get used to AI returning answers in seconds, anything that takes longer starts to feel inefficient. The paper uses a specific term — "productive struggle" — for the kind of effortful problem-solving that, the authors argue, is what actually pushes people toward their real interests and strengths. AI removes that struggle, and in doing so, helps you by helping you skip the part that would have helped you most.
The paper is not alone in this framing. It cites Anthropic's own published research showing that frontier labs already acknowledge how offloading tasks to a chatbot degrades a person's baseline mastery of a subject. Anthropic used the phrase "baseline mastery" — the more often you hand a task to AI, the more your own floor on that task drops. Christian's proposed response is a different default behavior: instead of optimizing for the shortest path to a correct answer, AI should default to instructive, tutor-style responses, even at a small efficiency cost.
Beyond fractions: what the researchers are really worried about
The numbers in the paper are restrained, but Christian's framing in the Berkeley interview is not. He calls the result "a story about human knowledge and human expertise from top to bottom, from grade-school students all the way to the leading experts in the world." His sharpest concern is scientific research: the process by which ten people brainstorm, sit with the data, and write a breakthrough paper is being compressed into "two or three researchers and a chatbot." Short-term efficiency rises, but the institutional muscle of a field may be quietly atrophying.
The New York Times ran a related feature in late September, summarizing the worry as "AI is shrinking the window in which we learn." WIRED and Psychology Today covered the early preprint back in the spring, suggesting the cognitive-impact question has moved from academia into mainstream conversation.
The paper also surfaces a quieter but important point: persistence is one of the strongest predictors of long-term learning — more predictive than short-term accuracy. That makes the trade-off uncomfortable. "AI helps you get 12 out of 12 right in ten minutes" is a real short-term win, but it may not come close to offsetting "the next time you hit a hard problem, you give up faster." The very metric users tend to judge AI by (did the answer come back fast and right?) is not the metric that determines whether they actually learned anything.
What the industry is already trying
Read closely, the paper functions as a quiet endorsement of a specific product direction: AI assistants that deliberately preserve struggle. That is in tension with how most current assistants are evaluated, where "answer accuracy" and "user satisfaction" dominate the dashboard, but several teams are already experimenting.
Anthropic's own coding-skills research pushed the field toward measuring whether a user can still perform the task independently after the AI is gone, rather than only counting successful task completion. OpenAI has been reported to be working on "tutor-style" modes that prompt the user to think instead of answering outright. Google's LearnLM family is in the same neighborhood, packaged as an education product rather than a developer tool.
Seen together, these signals point at what may be the defining product question for the next generation of AI assistants: the speed at which you deliver an answer is also the speed at which you drain the user's growth. The paper does not argue that AI assistants are useless — Experiment 1's first 12 problems show they genuinely help. What it provides, for the first time at this sample size and experimental rigor, is causal evidence that short-term help may come at the cost of long-term persistence. Whether a tool actually helped you cannot be judged by whether its answer was right; it has to be judged by whether you can still do the problem after it is gone.
The team is continuing to follow the 1,222 participants. If the "AI aftereffect" persists for days or weeks, the implications for education, coding training, and clinical skill acquisition — all of which lean heavily on the kind of "painful practice" the paper describes — extend well beyond fraction problems.