Ten Minutes of AI Help, and Then They Gave Up Sooner
A peer-reviewed set of experiments found that people who briefly used a chatbot did worse and skipped more problems once it was taken away. The effect is real in the lab. Whether it lasts in life is unknown.
Does using an AI assistant make you worse at the task once it is gone? In a set of randomised experiments, yes: people who had a chatbot for about ten minutes solved fewer problems afterwards, and in two of the three experiments they also skipped more of them. The effect was measured immediately after the tool was removed, in short online tasks. Whether it lasts, or builds up over months of daily use, was not tested.
Who ran the study, and where was it published?
The paper is titled AI Assistance Reduces Persistence and Hurts Independent Performance. Its authors are Grace Liu, Brian Christian, Michiel Bakker, Tsvetomira Dumbalska and Rachit Dubey. A preliminary draft circulated in the spring. According to a UC Berkeley news story, the updated, peer-reviewed version was presented this week at the Conference on Language Modeling. I read the arXiv version of the paper and the Berkeley story; I did not see the conference version, so figures below are from the arXiv text.
What exactly did they test?
Each experiment had the same shape. Participants were randomly assigned to work on problems either with an AI assistant in a sidebar or without one. The assistant was GPT-5, pre-loaded with each problem's solution, so a participant could type "answer?" and get the answer. Then the assistant was removed without warning, and everyone solved the same final problems alone. Skipping a problem cost nothing, so the authors treat a skip as a deliberate choice not to try.
| Experiment | Task | People analysed | Test solve rate, AI vs control |
|---|---|---|---|
| 1 | Fractions | 305 | 57% vs 74% |
| 2 | Fractions | 585 | 71% vs 77% |
| 3 | SAT-style reading | 168 | 76% vs 89% |
The total of 1,222 is the number recruited. The numbers analysed are smaller because the authors excluded people who failed an attention check or a basic pretest, as the paper describes.
During the AI phase, the AI group did better. Once it was gone, they did worse than people who had never had it.
Did people really give up more?
In two experiments, yes. In the first, the AI group skipped 20% of the test problems against 11% for the control group. In the third, the figures were 8% against 1%. In the second, the AI group skipped 10% against 7%, and the authors report that this difference was not statistically significant.
The authors then split the second experiment's AI group by how people said they used the tool. Those who mainly asked for answers solved 64% of the test problems and skipped 13%. People who used the AI for hints solved 75% and skipped 5%. The authors flag this breakdown as cross-sectional and not necessarily causal, since participants chose how to use the tool. It is still the most practically useful finding: how you use the assistant seems to matter.
Why would this happen?
The paper offers two explanations, and both are hypotheses, not results. The first is that instant answers shift your sense of how long a task should take, so unaided work starts to feel slow. The second is that without working through problems yourself, you lose a sense of what you can actually do. The authors tie persistence to long-term learning, citing earlier research on its link to academic achievement.
What does the evidence not show?
- It does not show lasting damage. Participants were tested straight after the AI was removed. The authors say they cannot speak to whether the impairment lasts hours or days.
- It does not show a cumulative effect. The exposure was roughly 10 to 15 minutes. The "boiling frog" argument, that small effects compound over years, is the authors' concern, and they call for longitudinal studies to test it.
- It does not cover all AI use. The assistant was built to give complete answers on request. The authors say a Socratic assistant that guides instead might avoid the cost, and note that this is untested.
- It is not a study of students or workers. The participants were US-based adults recruited through an online research platform, doing fraction and reading problems for small payments.
- One experiment had a flaw the authors acknowledge. In the first, the exclusion rule could have kept lower-ability people in the AI group, potentially inflating the gap. The second experiment was designed to fix this, and the gap there was smaller, though still significant on solve rate.
Is this the first evidence of its kind?
No. The authors say earlier work was mostly correlational or based on small samples, and cite randomised studies, including one on high-school maths students, that found similar harm to later unaided performance. What this paper adds, they argue, is a measure of motivation (skipping) across more than one kind of task.
What should an ordinary reader take from it?
The sensible reading is narrow. If you use an assistant to practise a skill you want to keep, asking for hints looks safer than asking for answers, and the study's own data point that way. If you manage people or teach them, the study is a reason to test what happens when the tool is unavailable, not a reason to ban it.
A brief session with a maximally helpful chatbot made people measurably worse and less willing to try once it was gone. The open question, and the one that decides whether this is a footnote or a real social cost, is what happens after months, and nobody has run that experiment yet.