News · 2026-10-12

Brief AI help lowered immediate independent performance in Berkeley experiments

Three randomized experiments found that participants given direct AI assistance performed worse on subsequent problems immediately after the assistant was removed. The Berkeley-linked study also found reduced persistence in some comparisons, measured as a greater tendency to skip questions. Its results concern short-term independent performance after brief help, rather than demonstrated lasting cognitive decline.

Key facts

• Researchers recruited 1,222 participants across two fraction-arithmetic experiments and one reading experiment.
• Berkeley publicized the work October 9; the updated paper was published as a conference paper at the Conference on Language Modeling in 2026.
• Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel Bakker and Rachit Dubey authored the paper.
• The primary sources are the paper and Berkeley’s report.

The important moment in this experiment comes after successful assisted work. A participant can get more answers right with help and still be less prepared for the next question alone. That makes the study relevant to how schools and workplaces evaluate AI assistance: the visible improvement while a tool is available is only one outcome.

The researchers titled their paper “AI Assistance Reduces Persistence and Hurts Independent Performance.” They recruited online participants through Prolific in the United States. The three recruitment totals were 354, 667 and 201. After exclusions, the analytic samples were smaller: 307, 585 and 168. Recruitment counts and analyzed counts should not be collapsed into a single number purporting to describe every reported comparison.

The arithmetic tasks involved short sequences of fraction problems of varying complexity. The reading experiment used paired passages and comprehension questions. In the AI condition, a GPT-5 sidebar was supplied with the current question and its solution. Participants could request complete answers, not only guidance. After roughly 10–15 minutes, the assistant disappeared without warning and participants attempted three more problems unaided.

That design produces a clear causal comparison within the experimental setting: random assignment determines whether someone receives this kind of assistance before the tool is removed. The existing lesson on randomized trials and treatment effects explains why assignment matters. It does not make every broader interpretation causal, especially explanations of why a particular participant chose to use the tool in a certain way.

A concrete analogy is a practice session with the answer key on the desk. The student can complete the worksheet quickly, but the last few problems after the key is taken away test something different: what they can do without it. The analogy helps explain the measurement. It does not show that an AI tutor and an answer key have identical effects, since the study did not compare those conditions directly.

The first fraction experiment found lower unaided solve rates and more skipped questions in the assisted group. The second replicated lower solve rates, but its aggregate skip-rate difference was not statistically significant. The reading experiment again found lower subsequent solve rates and higher skipping. The overall pattern is therefore stronger for immediate independent accuracy than for an identical persistence effect across all three experiments.

Within the AI condition, a subgroup analysis associated direct-answer use with the lower-persistence signal more than hint use. The authors explicitly treat that comparison as cross-sectional and not necessarily causal. People who chose answers may differ from people who chose hints. The result cannot by itself establish that changing everyone’s interface to hint-only would eliminate the effect.

Berkeley’s headline says brief use “erodes” persistence. The full paper provides the necessary limit: the exposure was brief, the assistant was unusually helpful, and follow-up occurred immediately. Whether effects persist for hours or days, accumulate with repeated use or change under different tutoring styles remains unknown. A durable decline in intelligence, motivation or character was not measured.

The strongest methodological challenge in the Hacker News discussion is the absence of other-help controls. Calculators, textbooks, human tutors or answer keys might also change practice and post-removal performance. Another plausible interpretation is that unaided participants simply practiced the exact skill more. These questions do not negate the randomized result; they narrow the mechanism the experiment can establish.

The practical implication is to assess assisted output and independent transfer separately. A tool can improve completed work while changing what its user practices. For education, that makes the form of assistance and the eventual unaided task central design choices. The study supplies a reason to measure those outcomes deliberately, while leaving long-term learning and ordinary sustained workplace use as open empirical questions.

Primary source, verified: read the paper → (arXiv 2604.04721)

Ground Truth. (2026, October 12). Brief AI help lowered immediate independent performance in Berkeley experiments. Ground Truth. https://groundtruth.day/news/berkeley-ai-assistance-immediate-independent-performance.html

@misc{groundtruth:berkeley-ai-assistance-immediate-independent-performance,
  title  = {Brief AI help lowered immediate independent performance in Berkeley experiments},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/berkeley-ai-assistance-immediate-independent-performance.html}
}

Topics: education · human-ai-interaction · evaluation · cognitive-offloading