Brief AI Help Lowered Later Accuracy and Raised Skipping in Three Experiments

Randomized experiments found that people who got about 10 to 15 minutes of AI help solved fewer problems on their own afterward, and in two of three experiments skipped more. The finding is narrower than the headlines suggest, and the way people used the AI appeared to matter.

People who spent roughly 10 to 15 minutes working fraction or reading problems with an AI assistant, and then lost it, solved fewer problems correctly than people who never had help. In two of the three experiments, they were also more likely to click "skip." That is the core of a paper by Grace Liu of Carnegie Mellon University, Brian Christian of UC Berkeley and colleagues at Oxford, MIT and UCLA, which began as a preprint in April 2026 and was presented at the Conference on Language Modeling in October.

The result bears on a worry many people already voice: that leaning on a chatbot wears down the willingness to struggle. The study supplies causal evidence for a narrow version of that worry. It does not show that 10 minutes of AI makes anyone give up on hard problems in general, and the gap between the headline and the data is worth understanding before drawing lessons for school or work.

What the researchers actually tested

According to the paper on arXiv, the work is a series of randomized controlled trials with 1,222 participants in total. The authors describe the design as giving causal evidence that AI assistance reduces persistence and impairs unassisted performance. Random assignment matters here because it rules out the explanation that people who like chatbots were simply weaker or less motivated to begin with.

The tasks were school-style questions. Two experiments used fraction problems, and a third used SAT-style reading comprehension. A secondary summary by the Italian analyst Pasquale Pillitteri, which should be checked against the paper, says the participants were U.S.-based adults recruited through the online platform Prolific. The same summary says people in the AI condition were told they were using ChatGPT, while the sidebar was in fact GPT-5 preloaded with each problem and its solution.

In Experiment 1, the AI group had help for 12 of 15 problems. Then the assistant disappeared for the final stretch, which the brief's table treats as the test phase of about three problems. The control group worked all the problems alone. Researchers compared the two groups on how many of the unassisted problems they solved correctly and how often they skipped.

How long "10 minutes" really is

The "10 minutes" in the coverage is a rounding of the paper's own language. The abstract says about 10 minutes, while the introduction describes sessions of 10 to 15 minutes, and whole sessions lasted 13 to 15 minutes. A secondary summary of the paper says effects emerged after as little as 10 to 15 minutes of use. The honest reading is that the exposure was brief, not that a fixed 10-minute threshold exists.

The authors also offer an explanation for the pattern. They propose that AI conditions people to expect immediate answers, denying them the experience of working through a challenge on their own. That is the authors' hypothesis, introduced with the words "we posit," and the experiments did not test it directly.

What the numbers show

The accuracy effect appeared in all three experiments. The skipping effect was less consistent. The table below lists the unassisted-phase results as reported through the paper and its summaries, using the brief's figures.

In Experiment 1, according to an overview of the paper on alphaxiv, the AI group solved 0.57 of the unassisted problems against 0.73 for controls, a difference with an effect size of d = −0.42. Another summary on Emergent Mind puts the skip rate at 20% for AI users and 11% for controls. The replication in Experiment 2 added a control matched on skill, and the accuracy gap shrank but held: 0.71 against 0.77, with an effect size of d = −0.19.

Skipping did not hold up as cleanly. In Experiment 2, the paper itself says the aggregate difference in skip rate did not reach statistical significance, and the Pillitteri summary gives p = 0.239 for 10% against 7%. Experiment 3 showed the largest skip gap, 0.08 against 0.01 (P = 0.008, d = 0.42), although the same secondary source says answers given in under five seconds also counted as skips there, which may widen what the category captures.

Effect sizes on accuracy range from d of about −0.19 to −0.42, which makes them small to moderate. Pillitteri's summary also reports that about 14% of participants were excluded, leaving 1,060 of 1,222 in the analyses, and that point is consistent with the degrees of freedom in the paper's tests. The brief could not confirm whether the exclusion rules were set in advance, and the authors themselves raise the possibility that exclusions in the AI condition retained lower-ability participants in Experiment 2.

Why "giving up on hard problems" overstates it

The popular framing says AI erodes grit. The study measured something smaller. Nobody was observed abandoning a difficult, open-ended project. Participants worked short fraction or reading questions, and a skip click served as the stand-in for reduced motivation. The paper did not test especially hard problems.

Skipping is a proxy

A skip can mean a person lost the will to try. It can also mean the person had less practice with that kind of problem, or chose a different strategy. The experiments cannot separate those readings. That is one reason the brief rates the persistence finding at only moderate-to-low confidence, while the accuracy finding in this setting is rated high.

The test came right away

Testing began as soon as the AI was removed, so the study cannot say how long any effect lasts. Writing on October 19, computer scientist Cal Newport speculated that participants who struggled just after ChatGPT was switched off might do much better on a later test in a setting where AI help was clearly unavailable. That is his conjecture, not a tested result, and the paper does not examine it. Live Science notes that scientists do not yet have a full picture of how small instances of offloading thinking affect lasting competence.

An artificial setting

Paid online participants, simple school-type tasks and a sidebar assistant that, per one source, held the correct answers all differ from daily life. Real chatbots make mistakes, and errors could change how people lean on them. No coding or workplace tasks were tested in this paper, so conclusions about workers rest on other studies, discussed below.

Answers versus hints: how people used the AI

The most practical finding concerns how participants used the assistant. According to the alphaxiv overview, 61% of the AI group in Experiment 2 mainly asked for direct answers. The 27% who used it for hints and clarifications did not differ significantly from controls in solve or skip rates on the test. The authors' project page says the declines were concentrated among participants who obtained direct solutions.

Pillitteri's summary adds counts: 189 people asked directly for answers, 82 sought hints or clarifications, and 37 barely touched the tool. It reports that the hint users solved 76% of test problems and skipped 5%, close to the control group's roughly 77% and 7%. The Daily Californian likewise reports that the hint group performed more like the no-AI control.

Other randomized studies point the same way

No independent replication of Liu and colleagues' work turned up in the brief. Two other randomized studies measure related outcomes, learning rather than skipping, in different groups of people.

High school math in Turkey

A June 2025 field experiment in the Proceedings of the National Academy of Sciences, led by Hamsa Bastani, gave nearly a thousand high school math students access to generative AI tutors. The authors concluded that access can improve performance but can substantially inhibit learning when no guardrails are in place. Having the tutor give teacher-designed hints instead of answers largely mitigated the harm.

A separate preprint that cites the work reports a 48% gain in AI-assisted practice performance and a 17% drop on unassisted exams relative to control. Those figures come from that secondary preprint and were not checked in the PNAS paper. The authors state limits of their own: two AI tutors, a single subject and a single high school in Turkey. A PNAS correction notice was issued in August 2025, and its contents were not reviewed for this article.

Junior developers learning a Python library

In a randomized trial posted in January 2026, Anthropic researchers Judy Shen and Alex Tamkin tested how quickly developers picked up a new Python library with and without AI help. Anthropic's write-up reports a statistically significant decrease in mastery in the AI group. A developer-community summary puts the quiz averages at 50% for the AI group and 67% for hand coders among 52 participants (d = 0.738, p = 0.010), with no significant difference in task time of about 23 versus 25 minutes.

The same summary says that people who stayed cognitively engaged, by asking for explanations and checking their understanding, learned well, while pure delegators learned almost nothing. That detail comes from a secondary source and should be read with caution. The sample was small, comprehension was measured shortly after the task, and Anthropic, which sells AI assistants, sponsored the work. It deserves credit for publishing a result unfavorable to its own product, but it is not an independent party.

Habits the evidence supports for students and workers

The research supports context rather than prescriptions, and the habits below are inferences from the studies, not tested interventions. They line up across the Liu, Bastani and Shen-Tamkin results, each of which found that how the AI is used shapes the outcome.

• Try first, then ask. In the use-pattern data, people who sought hints or clarifications resembled people who worked alone, while people who sought full answers did worse afterward.
• Ask for explanations or checks, not finished solutions. Developers who engaged with explanations and verified their understanding learned well in the Anthropic trial, according to a secondary summary.
• Separate output tasks from learning tasks. The measured harms concerned learning and unaided performance, not the quality of AI-assisted work, so the goal of the task matters.
• Build in AI-free practice and measure unaided performance. For teachers and managers, this is an editorial inference: if assisted performance can improve while unaided skill falls, as in the Turkish high school study, only an unaided check reveals the gap.
• Prefer tools that give hints. Bastani and colleagues offer the only randomized test of a design change in this set, and a hint-giving tutor largely mitigated the harm.

Christian told the Daily Californian that long-term effects are a crucial area for future study. Liu said the team began the project after noticing its own growing reliance on AI and wanting to examine it cognitively. The authors also call for AI systems that scaffold long-term competence rather than only supply answers.

What it means, and what to watch

The cleanest reading of the evidence is modest. A brief spell of AI help, mostly used to get answers, left people less accurate on their own right afterward and, in two of three experiments, quicker to skip. The authors warn that if such effects accumulate with sustained use, current AI systems risk eroding human persistence. That sentence describes a hypothesis, not a result.

Three things will decide how much weight the finding deserves: a delayed test that shows whether the drop fades, a study that randomly assigns hints-only versus answer-giving AI, and replication by researchers outside the original group. Until then, the studies agree on one point. The harm, where it appears, seems tied to handing over the thinking, not to the mere presence of a tool.

Frequently asked questions

Does 10 minutes of AI use make people give up on hard problems?

Not in the broad sense the headlines suggest. In randomized experiments, people who had about 10 to 15 minutes of AI help solved fewer problems afterward than controls, and in two of three experiments skipped more. The tasks were short fraction and reading questions, and persistence was measured as clicking skip.

Who ran the study and where was it published?

The paper is by Grace Liu of Carnegie Mellon, Brian Christian of UC Berkeley, Tsvetomira Dumbalska of Oxford, Michiel Bakker of MIT and Rachit Dubey of UCLA. It appeared as an arXiv preprint in April 2026 and was published as a conference paper at the Conference on Language Modeling, where it was presented in October.

Does it matter how AI is used?

It appears to. In the second experiment, participants who used AI for hints and clarifications performed about like controls, while those who mainly sought direct answers did worse. That comparison was not randomized, but a randomized trial with Turkish high school students found a hint-giving tutor largely mitigated learning harm.

Does the effect last?

Unknown. The test came immediately after the AI was removed, and no study in the brief followed people over days or months. Cal Newport has speculated that the effect could be short-lived, but that has not been tested.

Does the finding apply to workers?

This paper tested neither coding nor workplace tasks. A separate randomized trial of 52 mostly junior software engineers found lower mastery of a new Python library among those who used AI, but it was small, measured comprehension shortly after the task, and was sponsored by Anthropic.

Sources

1. AI Assistance Reduces Persistence and Hurts Independent Performance , arXiv (Liu, Christian, Dumbalska, Bakker, Dubey) (2026-04-01)
2. AI Assistance Reduces Persistence and Hurts Independent Performance (project page) , Authors' project website
3. AI Assistance Reduces Persistence and Hurts Independent Performance (paper overview) , alphaxiv
4. AI Assistance Harms Persistence & Performance , Emergent Mind
5. 10 Minutes of AI Reduces Persistence, Berkeley Study Says , Pasquale Pillitteri (2026-10-10)
6. Using AI for just 10 minutes erodes your ability to persist at hard things , UC Berkeley via EurekAlert (2026-10-01)
7. AI Assistance Reduces Persistence and Hurts Independent Performance (COLM 2026 listing) , Bespacific
8. 10 minutes is all it takes for AI assistance to erode your motivation, study shows , Live Science (2026-10-01)
9. UC Berkeley researcher discovers just 10 minutes of AI use impairs problem-solving abilities , The Daily Californian (2026-10-01)
10. The Good Thing About Hard Things , Cal Newport (2026-10-19)
11. Generative AI without guardrails can harm learning: Evidence from high school mathematics , Proceedings of the National Academy of Sciences (2025-06-25)
12. Correction for Bastani et al., Generative AI without guardrails can harm learning , Proceedings of the National Academy of Sciences (2025-08-20)
13. Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build , arXiv
14. How AI assistance impacts the formation of coding skills , Anthropic (2026-01-28)
15. Devs Learning With AI Scored 17 Points Lower in Anthropic's Trial. Juniors Code Solo First , DEV Community

This article was drafted with the help of AI and reviewed by Jaroha editors before publication.