Ten minutes with AI, and people skip hard problems more often. Part 1: what experiments with 1,222 participants showed
Topics: AI, Education, Science
Short answer: in three experiments with 1,222 participants from the US, researchers from Berkeley, Carnegie Mellon, MIT, Oxford and UCLA gave one group of people an AI assistant and the other group none, and then removed the assistant without warning. After roughly ten minutes with AI, people solved the following problems worse than those who had worked on their own from the start, and in two experiments out of three they skipped problems statistically significantly more often. In the second experiment, those who took ready-made answers from the AI did worse than both the control group and those who took hints. Those who took hints differed less from people who worked without AI, but this part of the analysis does not prove cause. Whether the effect lasts an hour, a day or a month later, the study does not know. Below is what exactly they did, what numbers they got and what to take from it.
Who tested what
The paper is called "AI Assistance Reduces Persistence and Hurts Independent Performance." The authors are Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel Bakker and Rachit Dubey. According to the Berkeley news story of October 9, 2026, the work was presented at the COLM conference, and the paper itself carries a note that it was accepted to that conference.
Brian Christian, a researcher at the Center for Human-Compatible AI at Berkeley, is known for his 2020 book "The Alignment Problem." According to the news story, he writes in a notebook every day: it helps him quiet the suspicion that working with AI too much might make him lose his own scientific edge.
How the experiment works
All three experiments are set up in a similar way. I take the first one as an example.
- Participants. 354 people from the US, recruited through the online research platform Prolific. They were randomly split into two groups.
- Problems. 15 fraction problems of varying difficulty. Wrong answers were not penalized, and pay did not depend on how many problems were solved.
- AI group. For the first 12 problems there was a side panel with an assistant running on the GPT-5 model. The assistant had been given each problem and its solution in advance, so a person could simply ask "answer?". Then the assistant was removed without warning and three more problems followed.
- Control group. Solved the same 15 problems on their own.
- "Skip" button. Everyone had one. Since there was no penalty for mistakes, the authors chose the share of skipped problems as their measure of persistence and motivation. Whether a person tried to solve a problem before pressing the button, the experiment does not show.
The main comparison is on the last three problems, identical for both groups: both groups solved them without help.
What came out
| Experiment | Participants | Problems | Average share of correct solutions on the last three problems: with AI vs. without AI | Average share of skips on the last three problems: with AI vs. without AI |
|---|---|---|---|---|
| 1 | 354 | fractions | 57% vs. 73% | 20% vs. 11% |
| 2 | 667 | fractions, with a pretest | 71% vs. 77% | 10% vs. 7%, the difference is not statistically significant |
| 3 | 201 | reading comprehension, questions from open practice materials for the American SAT exam | 76% vs. 89% | 8% vs. 1% |
According to the authors, while the AI was available, it improved results in the moment. As soon as it was removed, the AI group in all three experiments solved the last three problems worse than those who had worked on their own from the start. The number of participants in the table is before exclusions; the percentages are calculated on those who remained. In all experiments, those who failed an attention check were excluded. In the second and third, so were those who failed the pretest problems, before any AI. In the first, the skill-based exclusion was done on problems within the experiment itself, and the authors themselves call this a possible distortion: the AI group may have retained weaker participants who got correct answers from the assistant. That is why the second experiment added a pretest.
In the second experiment the difference in skips turned out not to be statistically significant, meaning such a difference could also have arisen by chance. The authors write that this may be because people used the AI in different ways, and they analyze this separately.
Ready-made answers vs. hints
At the end of the second experiment, participants in the AI group were asked how they had used it. The split below rests mainly on their own answers:
- 61% mostly took ready-made answers from the AI;
- 27% took hints and explanations;
- 12% did not use it at all. The authors checked this against the logs of chats with the assistant.
From here the difference became more visible. On the last three problems, the average share of correct solutions was 65% among those who took ready-made answers, 77% in the control group, 76% among those who took hints and 89% among those who did not use AI. The average share of skips: 13% among those who took answers, 7% in the control group, 5% among those who took hints and 7% among those who did not use AI. According to the paper, those who took ready-made answers did statistically significantly worse than each of the three other groups.
An important caveat from the paper itself: this is a comparison of groups that formed on their own, based on participants' answers, not on random assignment. So it does not prove that ready-made answers caused the drop. On the pretest, before working with AI, these groups did not differ statistically significantly in results, but even that does not turn the comparison into proof of cause: the groups may have differed in something the pretest did not measure.
What the study does not know
The authors list the limitations themselves.
- Short exposure. People worked with AI for about ten minutes. Whether the effect builds up with daily work or, on the contrary, weakens, is unknown.
- Testing right away. Participants were tested immediately after the assistant was removed. Whether the effect holds for hours or days is unknown.
- Maximally helpful AI. The assistant gave a full answer on the first request. Whether an AI that acts like a tutor and does not hand out ready-made answers would help is an open question. The authors call the hint data from the second experiment only a tentative argument in favor of testing this.
- Problems. Fractions and reading comprehension questions. Christian says in the Berkeley news story that this story is "not about fractions and SAT questions," but the experiments themselves tested only these kinds of problems.
The authors' explanation: AI conditions people to expect an instant answer and deprives them of the experience of working through difficulty on their own. This is their hypothesis, not a measured fact.
What to take from it
These are my conclusions, not data from the paper.
- Learn with hints, not answers. If the goal is to learn, ask the AI for a hint or for a check of your own solution, not for a ready-made answer. The study does not prove that this protects you, but in its data those who took hints differed less from those who worked without AI than those who took answers.
- Who does it: the person who is learning, whether a new employee or a schoolchild.
- How to check: once a week, solve a similar problem with no AI at all and write down whether you managed and how long it took.
- Tell "getting it done" apart from "learning." When a task just needs to get done and the skill is not needed, a ready-made AI answer is a normal tool. When an employee will later need to manage on their own, for example a newcomer in their first weeks, agree on where they work without ready-made answers.
- Who does it: the newcomer's manager.
- How to check: a task from regular work that the newcomer does without AI at the end of the probation period.
- Do not draw big conclusions from the study. It showed an effect right after ten minutes on fraction problems and reading comprehension questions. It says nothing about what happens to skills after a year of daily work with AI.
How the experiments work in more depth, what the numbers mean statistically and how the second experiment corrected the first, I go through in the continuation: Part 2: method, statistics and limitations.
When this does not concern you
If you use AI for tasks you do not need to be able to do yourself, for example a draft of a letter that you will proofread anyway, this study changes almost nothing. It is about the skill of solving a problem without help.
Summary
In three experiments with 1,222 participants, after roughly ten minutes with an AI assistant, people without it solved problems worse than those who worked on their own, and in two experiments out of three they skipped problems statistically significantly more often. In the second experiment, those who took ready-made answers did worse than both the control group and those who took hints, but this part of the analysis does not prove cause: the groups by usage style formed on their own. How long the effect lasts and what happens with long-term AI use is still unknown.
I work on AI agents and automation. If you want to look at my projects or discuss your own task, visit my portfolio.
Sources
- UC Berkeley Research, Jason Pohl, 09.10.2026: Using AI for Just 10 Minutes Erodes Your Ability to Persist at Hard Things. https://vcresearch.berkeley.edu/news/using-ai-just-10-minutes-erodes-your-ability-persist-hard-things
- Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, Rachit Dubey: AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721, version 5 of 03.10.2026. https://arxiv.org/abs/2604.04721