Ten Minutes of AI Help Made People Quit Sooner. Asking for Hints Changed That.

People who leaned on a chatbot for about 10 minutes did worse and gave up more often once it was taken away, according to a set of randomized trials with 1,222 participants presented this week at the Conference on Language Modeling (COLM). The paper, whose authors include Brian Christian of UC Berkeley's Center for Human-Compatible AI, found that GPT-5 assistance raised short-term scores but cut persistence and unaided performance. In the first experiment, people who had AI solved 57% of the test problems on their own, against 73% for people who never had it. The catch is in the details: the damage was concentrated among people who asked the AI for answers, not hints.

Twelve Fraction Problems, Then the Chatbot Vanished

The setup was deliberately ordinary. Participants recruited on Prolific, all US-based and paid $2.60 to $3.40, worked through 15 fraction problems. One group had a GPT-5 chatbot in a sidebar, labelled "ChatGPT", which they could ask for anything, including the answer. The control group worked without it. After problem 12, the assistant was removed without warning, and everyone did the last three problems alone. Participants were asked not to use outside AI or Google.

With the chatbot available, the AI group was ahead. Without it, they fell behind almost immediately. In Experiment 1 (307 analysed participants), the AI group solved 57% of the final problems versus 73% for controls, and skipped 20% of them versus 11%. The paper reports both differences as statistically significant.

A larger replication with 585 analysed participants added a pretest to rule out the possibility that weaker students had simply been hiding behind the AI. The gap shrank but held: 71% versus 77% solved. The skip-rate difference, 10% versus 7%, was not statistically significant this time. That is a real caveat, and it is the main reason the headline "AI makes you give up" needs the qualifiers below.

The third experiment moved from math to SAT reading comprehension, where answers submitted in under five seconds counted as skips. After the AI disappeared, the AI group solved 76% of passages versus 89% for controls, and skipped 8% versus 1%.

The Users Who Asked for Answers Paid the Price

The most useful finding is about how people used the tool. In Experiment 2, 61% of the AI group said they mainly asked for direct answers, 27% mostly asked for hints or explanations, and 12% didn't use the chatbot at all.

Those groups ended up in very different places. The answer-seekers solved 65% of the unassisted test problems and skipped 13%, against 77% and 7% for the control group. The hint-seekers solved 76%, roughly the same as people who never had AI. The people who ignored the chatbot scored 89%. According to The Decoder's summary, only the direct-answer users got worse compared with their own pretest.

So the study is less "AI rots your brain" and more "outsourcing the thinking has a cost you pay later." And the finding isn't about one product. The study used GPT-5, but any chatbot that hands over answers on request can be used the same way, and ChatGPT now holds less than half the AI market.

Why Ten Minutes Is Enough

The authors offer two explanations. First, AI resets people's sense of how hard a task should feel: after instant answers, working through a problem by hand feels slower and more frustrating than it did before. Second, the chatbot removes the "productive struggle" that builds skill and gives people an honest sense of what they can do alone.

"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. He was clear that fractions aren't the point: "This is not a story about fractions and SAT problems."

His worry is about where the habit leads. If people lose the patience to stay with hard material, he argued, that could weaken work that depends on it, including science itself, which relies on closeness to data and long immersion in a problem.

What the Study Doesn't Show

The experiments were short, low-stakes and unpaid for performance. There was no penalty for wrong answers and no bonus for right ones, so skipping costs participants nothing, which may exaggerate how quickly people give up. Nobody was followed for weeks or months, so the paper can't say whether the effect fades, builds up, or reverses with practice. The authors themselves frame the cumulative effect of daily AI use as an open, urgent question rather than a result.

They also argue that user-side fixes such as Socratic prompting or usage limits are band-aids. Christian's suggestion is that AI systems should default to behaving "more instructive, like a tutor" instead of handing over answers. That puts the responsibility on the companies building these tools, at a moment when they are mostly competing on speed and helpfulness. It is a familiar tension: the same industry that is citing AI as a reason to cut jobs is shipping products that, on this evidence, can quietly make the remaining humans less willing to do the hard part themselves.

The practical takeaway for anyone using a chatbot today is simple: ask it to explain or give a hint, not to give the answer. In this study, that one habit was the difference between falling behind and keeping up with people who never used AI at all.

Sources

• arXiv: AI Assistance Reduces Persistence and Hurts Independent Performance (Liu, Christian, Dumbalska, Bakker, Dubey)
• Berkeley News: Using AI for just 10 minutes erodes your ability to persist at hard things
• EurekAlert: Using AI for just 10 minutes erodes your ability to persist at hard things
• The Decoder: Just ten minutes of using AI as an answer machine can measurably erode problem-solving skills
• Project page: AI Assistance Reduces Persistence