Ten minutes of AI help and what happened next: the 2026 persistence study
After 10 to 15 minutes of help from an AI assistant, people solved fewer problems on their own and were more likely to give up. That is the result of three randomised experiments with 1,222 adults, posted in 2026. People who said they used the AI for hints did about as well as people who had no AI. It was posted as a preprint in April 2026 and accepted at the COLM 2026 conference, which is peer reviewed.
The study in one table
| Title | AI Assistance Reduces Persistence and Hurts Independent Performance |
|---|---|
| Who | Liu, Christian, Dumbalska, Bakker and Dubey |
| Status | Conference paper, COLM 2026 (peer reviewed). First posted April 2026, sixth version October 2026 |
| People | 1,222 adults in the United States, recruited online |
| Design | Three randomised experiments. Tasks were fractions and SAT reading questions |
| Source | arxiv.org/abs/2604.04721 |
What they did
People were put into groups at random. One group worked on problems with an AI assistant for 10 to 15 minutes. The other group worked with no AI. Then the AI was taken away and everyone got a short test of three problems to do alone. People could also skip a problem, which is how the researchers measured giving up.
What they found
| Experiment | Solved alone, after AI | Solved alone, no AI | Skipped |
|---|---|---|---|
| Experiment 1 | 57% | 74% | 20% after AI, 11% with no AI |
| Experiment 2 | 71% | 77% | No clear difference |
The authors' summary is that people "perform significantly worse without AI and are more likely to give up". The gap in the first experiment was moderate in size (d = -0.45). In the second it was small (d = -0.20).
These figures come from the sixth version of the paper (October 2026).
Hints against answers
In the second experiment, people were asked how they had used the AI. The answers split the group in a useful way.
- Direct answers. 61% said they mostly asked for the answer. Afterwards they solved 65% of problems alone and skipped 13%.
- Hints. 27% said they mostly asked for hints. They solved 76% and skipped 5%.
- No AI. The comparison group solved 77% and skipped 7%.
So the people who asked for hints looked like the people who never had the AI. The people who asked for answers did not.
What it did not show
- No long-term effect. The AI help lasted 10 to 15 minutes and the test came straight after. Nobody was followed for days or weeks.
- No proof that hints protect you. People chose for themselves how to use the AI, and reported it themselves. People who ask for hints may differ in other ways. The authors say this part is not causal.
- Nothing about ability. The study measured what people did on three problems. It did not measure intelligence or memory.
- Nothing beyond these tasks. Fractions and reading questions, adults from one online panel in one country.
A fair one-line summary. "In a 2026 study with 1,222 adults, ten minutes of answer-giving AI went with fewer problems solved alone and more giving up straight afterwards."
Why it fits with other studies
The same pattern shows up in a larger and longer trial. In a peer-reviewed field trial with about 1,000 students, a plain chatbot that gave answers went with lower exam scores, and a version that gave hints largely avoided the drop. In a small trial with developers, those who asked the AI why something worked kept learning.
What to do with this
Try the problem before you ask. When you do ask, ask for a hint or a question, not the answer. You can set this once so every chat starts that way.
Questions
Does AI make you give up more easily?
In this study, people who had just had 10 to 15 minutes of AI help skipped more problems in one of the experiments: 20% against 11%. In the second experiment the difference was not clear. In the third it was 8% against 1%. It is a short-term result.
Is the persistence study peer reviewed?
Yes. It was accepted at the COLM 2026 conference. The arXiv copy is in its sixth version.
Is asking for hints safe, then?
Hint users did about as well as people with no AI. But they chose that style themselves, so the study cannot prove the hints were the reason. It is a promising sign, not a proof.
How long did the effect last?
Nobody knows. The test came right after the AI help. The study did not check again later.