AI assistants: less persistence, worse performance on math and reading tests

The study titled "AI Assistance Reduces Persistence and Hurts Independent Performance" (arXiv 2604.04721, v5, October 3, 2026) consisted of three controlled, randomized online experiments with a total of 1,222 participants recruited through web platforms. In every experiment the assistant was GPT‑5, available in a sidebar and able to provide the full answer on request. After a learning phase, the assistant was removed without warning; the control group never had access to AI.

According to the authors, the effects appear after only about 10 minutes of interaction with the AI.

In the first experiment (fractions), with 354 subjects (307 analyzed), the solve rate was 0.57 for the group that had used AI versus 0.73 for the control group (P<0.001). Skipped questions were 0.20 versus 0.11 (P=0.031). The effect emerged after roughly 10 minutes of interaction with the AI, as the authors point out. In the replication experiment (fractions with a pre‑test), with 667 subjects (585 analyzed), solve rates were 0.71 versus 0.77 (P=0.020); the difference in skipped questions (0.10 vs 0.07) was not statistically significant (P=0.239). The third experiment, focused on SAT‑style reading comprehension, showed solve rates of 0.76 versus 0.89 (P=0.007) and skipped questions of 0.08 versus 0.01 (P=0.008).

In the second experiment, 61% of participants in the AI group asked for direct answers, 27% for hints or clarifications, and 12% did not use the assistant. In the test without AI, those who had asked for direct answers solved 65% of the problems, compared with 77% in the control group; those who had asked for hints reached 76%.

The authors describe this comparison as exploratory and not necessarily causal. Brian Christian, a research fellow at UC Berkeley's Center for Human‑Compatible AI, commented: "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".

The study's limitations are documented: exposure was brief (10–15 minutes), so it is unknown whether the effect persists or fades with prolonged use; measurement took place right after the AI was removed, so how long the effect lasts remains uncertain; the assistant gave complete answers, and a Socratic approach that guides without handing over the solution was not tested; the tasks were relatively simple and participants were recruited online, which limits generalizability. In addition, the presentation at the Conference on Language Modeling (COLM) is confirmed only by the UC Berkeley press release (October 9, 2026) and cannot be verified directly on the conference website.

— Pixie