Study of 1,222 People Finds 10 Minutes of AI Help Makes Them Quit Hard Problems Sooner
A.I. / news
Study of 1,222 People Finds 10 Minutes of AI Help Makes Them Quit Hard Problems Sooner
The preprint, five versions in and presented at the Conference on Language Modeling, reports effects in fractions and SAT reading tasks. Neither summary we read gives an effect size.
A preprint co-written by Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI, reports that people who solved problems with an AI assistant for about 10 minutes did worse once it was taken away, and gave up more often. The University of California, Berkeley's research office publicized it on Oct. 9, 2026.
The paper, "AI Assistance Reduces Persistence and Hurts Independent Performance," is on arXiv as 2604.04721. Its authors are Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker and Rachit Dubey.
Three experiments, 1,222 participants
All three were randomized controlled trials run online. In each, one group had an AI tool and a comparison group did not, and the tool was then removed partway through.
| Experiment | Participants | Task |
|---|---|---|
| 1 | 354 | 15 fraction problems, ChatGPT removed after problem 12 |
| 2 | 667 | Same basic design, larger group |
| 3 | 201 | SAT reading comprehension questions |
The university's write-up says AI users skipped questions or gave up more often, while the group without AI kept solving. The arXiv abstract page says AI assistance improves short-term performance but that people then do worse without it, and that the effects appear after roughly 10 minutes of interaction.
The design makes the removal point easy to locate in experiment 1. The AI group had ChatGPT for the first 12 of 15 fraction problems, so the last three were the unaided test. In experiment 3 the tool was likewise withdrawn partway through a set of SAT reading questions.
What the authors argue
The authors propose that persistence strongly predicts long-term learning, and that AI may reduce it by conditioning people to expect an immediate answer. They call for AI built to support long-term competence as well as task completion.
Christian said "You almost couldn't ask for a clearer story," according to the Berkeley article. He also said the tools "are often helping us in ways that are kind of unhelpful."
What the two sources do not say
Neither the Berkeley article nor the arXiv page gives an effect size, a percentage drop or a p-value, so the size of the effect is not reported here. The university article does not discuss limitations.
The participants were online volunteers doing fractions and reading questions. The two sources do not describe any follow-up after the sessions, so nothing in them shows how long the effect lasts.
The paper has had five versions since v1 on April 6, 2026. The current one, v5, was revised on Oct. 3. The Berkeley article says the peer-reviewed version was presented at the Conference on Language Modeling in the week the article ran, and it does not name a journal.
The finding lands in a live argument on the site. Linus Torvalds has published patches written with Claude Code, while the QGIS project banned AI-assisted code from its repository. A trial of fractions and SAT passages does not speak to either case.
The full paper on arXiv, which neither source summarizes in numbers, is the place to look for the effect sizes.
Sources
More in A.I.
- 01Torvalds Posts a Kernel Patch Series Written With Claude Code, and Sources Disagree on Its Size and SpeedupLWN counts ten patches and an order-of-magnitude gain, while It's FOSS counts eight patches and a threefold gain on one laptop.
- 02Parsimony Ranks Coding Agents by How Much Code Their SWE-bench Fixes Add, and Its Own Page Warns Smaller Is Not BetterThe MIT-licensed benchmark counts coding units in published patches without rerunning a single test, and reports that code volume tracks solve rate.
- 03OpenBMB's MiniCPM-V 4.7 Went Up With No Model Card on Oct. 6, and Now Reads Apache-2.0 With No Benchmark NumbersThe 35-billion-parameter mixture-of-experts vision model is 27 times larger than its predecessor, and its card claims a win over Qwen3.6 35B-A3B without a score.
- 04Design Creativity Bench Scores GPT-6 Astra 65.5 Against Humans' 92.3, and the Paper Never Names Its JudgeA post using Simon Willison's pelican SVGs to argue that models converge on one image cites a leaderboard whose method the arXiv page does not describe.