A team including UC Berkeley researcher Brian Christian and Grace Liu (Carnegie Mellon University) reports a series of randomized controlled trials showing that roughly 10–15 minutes of AI-assisted problem solving can raise short-term accuracy—yet once the assistant is removed, unassisted solve rates fall and skip/give-up rates rise. The paper, AI Assistance Reduces Persistence and Hurts Independent Performance, was presented in peer-reviewed form at COLM 2026 (Conference on Language Modeling). Berkeley News covered the conference presentation on 9 October 2026; Christian also spoke on the work at Berkeley on Wednesday, 7 October 2026.
Across three experiments the team recruited 1,222 US-based online participants. In AI conditions, a sidebar GPT-5 assistant was available; participants could ask anything, including for the answer.
Fractions: better while helped, worse when alone
Experiment 1
Experiment 1 recruited 354 people (307 after exclusions for attention checks and basic fraction ability). During a learning phase of 12 fraction problems, the AI group could use the assistant; then AI was withdrawn without warning and both groups solved 3 test problems. On the test:
- Mean solve rate: 0.57 (AI) vs 0.73 (control);
- Mean skip rate: 0.20 (AI) vs 0.11 (control).
During learning, AI users were more accurate and skipped less; after removal, they lagged participants who never had AI.
Experiment 2: replication with cleaner design
Experiment 2 recruited 667 people (585 in the final sample), adding a pretest-based exclusion rule and giving controls a sidebar of pretest worked solutions during learning to reduce interface asymmetry. Test solve rates were 0.71 (AI) vs 0.77 (control)—a significant gap—while the overall skip difference (~0.10 vs ~0.07) was not significant.
Usage patterns mattered. Among AI-condition participants, about 61% (189) said they mainly used AI for direct answers; about 27% sought hints or clarifications; about 12% reported little or no use. Direct-answer users had the lowest test solve rate (mean 0.65) and the highest skip rate; hint users tracked closer to controls. Pretest solve and skip rates did not differ significantly across usage groups and controls, so the test gap emerged after AI use.
Reading: the same pattern
Experiment 3 used SAT reading-comprehension items (recruited 201; final 168). After AI removal:
- Solve rate: 0.76 (AI) vs 0.89 (control);
- Skip rate: 0.08 (AI) vs 0.01 (control).
The authors argue the effect is not limited to arithmetic reasoning but also appears in meaning-making and critical reading.
Persistence and design implications
Skipping is treated as a persistence/motivation measure: participants were told wrong answers did not reduce pay, so choosing skip signals deliberate disengagement. Both the paper and Berkeley News note that instant complete answers may reset expectations toward “seconds-scale” completion and remove productive struggle—the effort through which people build skill and self-knowledge. In the Berkeley News piece, Christian asks whether systems optimized to be helpful are helpful in ways that actually help people long-term, and argues for augmenting rather than replacing human ability—for example, defaulting toward tutor-like guidance instead of answer-dumping.
Limitations are explicit: exposure was brief and testing immediate after removal; whether effects accumulate or fade is open; the study AI was maximally helpful (full solutions on demand). The hint-user pattern in Experiment 2 tentatively supports exploring scaffolded or Socratic designs that preserve help without the same persistence cost.
Primary sources:
- Berkeley News, Using AI for just 10 minutes erodes your ability to persist at hard things, 2026-10-09, https://news.berkeley.edu/2026/10/09/using-ai-for-just-10-minutes-erodes-your-ability-to-persist-at-hard-things/
- Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R., AI Assistance Reduces Persistence and Hurts Independent Performance, arXiv:2604.04721, https://arxiv.org/abs/2604.04721
