Does using AI for just 10 minutes make you give up faster on hard problems (October 2026 study)?

AI Makes You Quit Hard Problems Sooner: The 10-Minute Study

Yes, if you use AI as an answer machine. In Grace Liu and Brian Christian's trials of 1,222 people, presented at COLM in October 2026, about 13 minutes with an AI helper cut the unaided solve rate from 74% to 57% and nearly doubled skipping. People who asked only for hints held up. John Whitmore's Coaching for Performance predicted it: being told builds dependence, being asked builds capability.

Brian Christian, the author of The Alignment Problem, keeps a notebook on his desk and writes in it every day, because he suspects too much time with AI could cost him his own edge. This month he published the data behind the suspicion. Christian is a research fellow at UC Berkeley's Center for Human-Compatible AI and a co-author, with first author Grace Liu, of "AI Assistance Reduces Persistence and Hurts Independent Performance." Berkeley News reported on October 9, 2026 that the team presented the updated, peer-reviewed paper at the Conference on Language Modeling that week; the latest arXiv version is dated October 13.

The design is simple enough to trust. In the first experiment, 354 people recruited on Prolific worked through 15 fraction problems. Half had an AI assistant (GPT-5) in a sidebar, preloaded with each problem and its solution, so typing "answer?" was enough. After 12 problems the assistant disappeared without warning and everyone solved three more alone. Wrong answers carried no penalty and pay did not depend on score, so skipping a problem measured one thing: whether you were still willing to try. On the unaided test, the AI group solved 57% against 74% for the control group, and skipped 20% of problems against 11%. A second experiment with 667 people added a pretest to remove a selection problem and got the same direction, and a third with 201 people on SAT reading passages repeated it outside maths. The effect showed up after a session of 10 to 15 minutes.

The number I would put on a sticky note comes from how people used the tool. In the second experiment, 61% of the AI group said they mainly asked for direct answers, 27% asked for hints or clarifications, and 12% did not use it. Those who asked for hints solved 76% of the test problems and those who ignored the assistant solved 89%. The direct-answer users skipped 13% of problems, against 5% for hint users and 7% for the control group, and they were the only group that did worse than their own pretest. The authors are careful that this split is cross-sectional, not causal, but it is the part that tells you what to do.

You do not do fraction problems for a living. The mechanism still transfers. The paper's explanation is that instant answers reset your reference point for how hard a task should feel, so working unaided starts to feel like waste, and every shortcut raises the felt cost of the next unaided attempt. For a founder that bites in the places the tool is absent or wrong: the pricing call you have to defend in a board meeting, the negotiation that happens in the room, the contract clause nobody summarised correctly.

John Whitmore described this dynamic thirty years before chatbots. Coaching for Performance argues that telling people what to do creates dependence and suppresses self-belief, while asking raises awareness and responsibility, the two things performance rests on. Today's assistants default to telling. The direct-answer group in the study was being told; the hint group was being coached. So the fix I take from Whitmore is to switch the assistant's mode for the work you need to own, and leave it in answer mode for the rest. Here is the prompt for that switch, run once before writing this page:

Tested October 21, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:

For this problem, act as a coach, not an answer machine. Do not give me the answer, a recommendation or a draft. Ask me one question at a time that makes me do the thinking. If I am stuck after two tries, give me a hint, never the solution. When we are done, ask me to state my decision in one sentence and how confident I am, from 0 to 100%.
The problem: I run a 12-person B2B software company. Our annual plan is $9,600 per customer, and 3 of our 40 customers churned last quarter. I am deciding whether to raise prices 20% for new customers only. Start.

The start of what came back, unedited:

Good, let's work through it. I'll ask one question at a time, and you do the reasoning.
**First question:** What's prompting this idea right now? What specific problem or opportunity do you expect a 20% price increase for new customers to address?

One run is one run: the same prompt gives a different answer next time, and a different model answers differently.

I ran this prompt on October 21, 2026 with Claude Opus 5.5, and the model held the line in its first turn: no answer, no recommendation, one question about what problem the price rise is meant to solve. That is the useful part, because it forces you to state the goal before the options, which is Whitmore's GROW order. Two weaknesses. Its "one question" was really two stapled together, and a single turn tells you nothing about the moment that matters, which is when you get impatient and type "just tell me." That is where assistants tend to cave, and this test did not reach it.

Name the limits before you change your habits. The participants were paid online workers spending about 13 minutes on school-level tasks, not executives on judgment calls, and nobody has measured what months of daily use do. The authors themselves call user-side fixes such as Socratic prompts and usage limits "band-aids" and argue that the models should change. And hint mode is slower; for formatting, boilerplate and anything you never need to do yourself, take the answer. My rule after reading the paper: if you would be embarrassed to be unable to do it without the tool, make the tool ask before it answers.

Sources: Study explores how AI use affects performance on math and reasoning tasks, Berkeley News, October 9, 2026; Liu et al., AI Assistance Reduces Persistence and Hurts Independent Performance (arXiv v6), October 13, 2026; Just 10 minutes of AI use may erode persistence on hard tasks, TechXplore, October 2026; Just ten minutes of using AI as an answer machine can measurably erode problem-solving skills, The Decoder, April 2026.