The headline is accurate, but incomplete
The paper is unusually useful because it asks a question most AI benchmarks avoid: what happens after the assistant disappears? Across three randomized experiments, people worked on fractions or reading comprehension. Some had a GPT-5 sidebar during the practice phase; others did not. Then the sidebar vanished for the final three questions.
The direction was consistent. Assistance lifted performance while it was present. On the immediate unaided test, prior AI users solved fewer questions. In the first fraction experiment they also skipped more. The reading experiment reproduced both effects. That supports a narrow causal claim: under these short experimental conditions, answer-ready AI assistance reduced performance and persistence on the next comparable tasks.
It does not support the broadest viral versions of the story. Nobody's brain was shown to be damaged. The paper did not measure durable skill loss. It did not follow participants into classrooms or workplaces. It tested the next three questions immediately after a tool was removed without warning.
The replication shrank the result
Experiment 1 produced the dramatic numbers: a 57% unaided solve rate after AI versus 73% in control, and a 20% skip rate versus 11%. But the authors themselves identified a selection problem. Their exclusion rule could retain lower-skill participants in the AI group because those participants could submit AI-provided correct answers, while comparable control participants were removed.
Experiment 2 fixed that with a pretest and balanced the interface change for controls. The accuracy gap remained, but became smaller: 71% versus 77%. More importantly for a headline about quitting, the aggregate skip-rate difference was 10% versus 7% and was not statistically significant. That is not a failed replication; it is a more precise description of the evidence. The performance effect survived. The overall persistence effect did not clear the conventional significance threshold in that experiment.
The third experiment moved from arithmetic to reading. Here the gap returned clearly: 76% versus 89% solved, and 8% versus 1% skipped. That makes the pattern harder to dismiss as a fractions-only quirk, while still leaving the same short-term limit in place.
Direct answers look worse than hints — cautiously
The most useful product signal comes from Experiment 2. Among people assigned AI access, 61% said they mainly asked for direct answers, 27% sought hints or clarification, and 12% said they did not use the assistant. The direct-answer group had the weakest later performance and higher skipping than the hint group.
That comparison was not randomized. People chose how to use the assistant, so differences in confidence, motivation or ability could partly explain the outcome. The paper says this explicitly. Treat it as a design hypothesis, not proof that changing a chatbot from answers to hints will erase the effect.
Still, it is the right hypothesis to test. An assistant optimized for immediate completion can look excellent on today's dashboard while quietly removing the practice that tomorrow's task requires. A hint-first system may feel less helpful in the moment and produce better independent performance later. Most product evaluations never measure the second half.
What the study changes
The paper does not justify banning AI from learning. It does justify changing the scorecard. If a tutor, coding assistant or workplace copilot claims to build capability, assisted task completion is not enough. Test users again without the tool. Measure whether they attempt hard problems, not only whether the final answer is correct. Separate direct-answer use from explanation, critique and staged hints.
The hard part is product incentives. A model that withholds an answer, asks the user to try first, or waits before helping can score worse on instant satisfaction. Yet that friction may be the feature. The study's strongest contribution is not the ten-minute warning. It is evidence that "helpfulness" and human capability can point in opposite directions unless the evaluation deliberately checks both.
Read the footnotes
The authors describe the work as randomized controlled evidence, and that is fair for AI access versus control within these tasks. The finer claims need finer labels. The direct-answer subgroup result is observational. One replication's overall skip-rate difference was non-significant. Participants were online adults, not schoolchildren. Exposure was brief, and the final test was immediate. Durability and cumulative effects remain open questions.
Those caveats do not erase the finding. They locate it. Ten minutes of answer-ready AI changed what people did on the next three unaided questions. Whether a different assistant, a longer relationship, or a better-designed learning workflow produces the same cost is exactly what researchers and product teams should test next.
Sources
- [1] Liu et al., “AI Assistance Reduces Persistence and Hurts Independent Performance,” arXiv v5 (October 3, 2026)Read source
- [2] Study project pageRead source
- [3] NewsNation report on the study (June 1, 2026)Read source