Ten Minutes of AI Help Hurt Solve Rates in All Three Trials. Giving Up Showed Up in Two.

A fraction problem fills the screen. The easiest ones in this study look like 5/6 minus 1/3, and the hardest chain three steps together. Beside the problem sits a chat sidebar where GPT-5 has already been handed the solution, so a participant who wants the answer can type a single word: "answer?" A tutor would wince at that prompt. The sidebar obliges. Twelve problems into the session, it vanishes without warning, and three more fractions are waiting.

That setup sits at the center of a paper presented this week at the Conference on Language Modeling, and of a UC Berkeley News story published October 9 under the headline "Using AI for just 10 minutes erodes your ability to persist at hard things." The first author is Grace Liu of Carnegie Mellon, with Brian Christian of UC Berkeley and colleagues at Oxford, MIT and UCLA. The paper is more hedged than the headline.

Three trials, one clear pattern and one shakier one

Experiment 1 recruited 354 people on the Prolific research platform. After exclusions, 185 AI-group participants and 122 controls remained. On the three unassisted test problems, the AI group solved 57 percent and the control group 73 percent. The AI group also skipped 20 percent of the test problems against 11 percent for controls, the widest raw skipping gap in the paper.

The authors flag a problem with that result themselves. They dropped anyone who solved fewer than 3 of the 12 practice problems, but AI users could clear that bar by copying answers, and more controls were dropped. The AI group may have kept weaker solvers, which could inflate the gap.

Experiment 2 fixed this. It added a three-problem pretest for exclusions and gave controls a matched sidebar, so both groups saw an interface change when the test began. Of 667 people recruited, 585 were analyzed. The AI group solved 71 percent of test problems against 77 percent for controls, a small effect (Cohen's d of 0.19). They skipped 10 percent against 7 percent. The authors' own test put that skip difference at p = 0.239 and describe it as not significant.

Experiment 3 swapped fractions for SAT reading questions. Of 201 people recruited, 168 were analyzed. The AI group solved 76 percent of test questions versus 89 percent, and skipped 8 percent versus 1 percent (p = 0.008).

Who actually gave up

In Experiment 2, the researchers also asked the AI group how they had used the assistant. Of 308 participants, 189 (61 percent) said they mainly asked for answers. Another 82 (27 percent) used it for hints or clarification, and 37 (12 percent) barely touched it.

The answer-seekers solved 65 percent of test problems and skipped 13 percent. Hint users solved 76 percent and skipped 5 percent. Controls solved 77 percent and skipped 7 percent. So the extra giving up sits in one subgroup, which is also where the authors urge the most caution. People chose their own style of AI use, so this analysis is cross-sectional and not causal by the authors' own account.

What ten minutes can and cannot tell you

The stakes were low. Pay was flat ($2.60 in the first trial, $3.40 in the others), and in the first trial people were told that wrong answers carried no penalty and that payment did not depend on how many they got right. A skip there is closer to a shrug than a failure. The authors tested people immediately after the AI left and say they cannot speak to whether the effect lasts hours or days. Their assistant was also maximally helpful by design, handing over complete answers on request. Whether months of daily use would compound the effect, fade it, or do neither is, in their words, an open question.

Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI and author of the 2020 book The Alignment Problem, keeps a notebook on his desk and writes in it every day, according to Berkeley News, partly to ward off a suspicion that too much time with AI could dull his own scholarly edge. Of the three experiments, he said, "You almost couldn't ask for a clearer story."

That reads strongest for solve rates, which fell in all three trials. For persistence, the cleanest fraction trial gave a smaller signal that did not clear the usual statistical bar.

The obvious objection is the calculator, which also does arithmetic for us. The authors raise it themselves and argue that fractions and reading comprehension are developmental prerequisites, the footing for algebra and critical reasoning. Their proposed mechanism is a hypothesis, not a tested result: when answers arrive in seconds, the expected length of a task shrinks, and unaided work starts to feel slower than it is.

The finding is not alone. The paper's literature review says earlier evidence on AI deskilling was mostly correlational or based on small samples, and cites a 2025 randomized trial of GPT-4 in high school math (Bastani and colleagues, in PNAS) that also found worse unaided performance afterward, partly offset when students used a scaffolded tutor. The direction matches. The size, in this paper's cleanest test, is modest: six points on solve rate.

The narrower version of the headline is still real. About ten minutes of answer-on-demand help left people solving fewer problems in three out of three trials. Giving up showed up in two, and in the one trial that tracked how people used the AI, it clustered among those who asked for answers rather than hints. What nobody has measured yet is the distance between a lab session and a working life.