After Ten Minutes With AI, Can People Still Keep Trying? A Study of 1,222 Participants
randomized controlled trial
A study that puts people into groups by chance and compares their results.
persistence
The ability to keep working when a task feels difficult.
unassisted performance
How well people work when the AI is no longer available.
What happened
Researchers from UC Berkeley (a U.S. university) and four other universities studied what happens after people use AI to solve problems. Brian Christian, a UC Berkeley researcher, was a co-author. The team ran three randomized controlled trials with 1,222 participants. The paper appeared as a COLM 2026 conference paper. More details appear in the Berkeley report and the paper.
The headline says ten minutes. The experiments actually measured short sessions lasting about 10 to 15 minutes.
How the experiments worked
Participants solved fraction problems or SAT-style reading questions. In the AI condition, ChatGPT (an AI that answers questions) sat beside the task. The assistant was GPT-5, with each problem and its solution already available. People could ask for the answer directly.
After several problems, researchers removed the AI without warning. Everyone then worked alone. A control group solved the same tasks without AI help.
AI users performed better while the tool was available. After removal, their independent results fell. They also skipped more questions.
In the first experiment, the AI group solved 57% of the final three fraction problems. The control group solved 73%. A larger replication still found a smaller accuracy gap. Its overall skip-rate gap was not statistically clear. In the reading experiment, the AI group solved 76% of the final questions. The control group solved 89%. Skip rates were 8% and 1%.
Why the result matters
AI can improve a task right now. Learning also requires trying, failing, and correcting mistakes. The researchers say instant answers may reset people’s sense of normal effort. Work without AI can then feel unusually slow.
The strongest decline appeared among people who asked AI for direct answers. People who mainly asked for hints did better afterward. This comparison was not randomized. It shows an association, not proof that direct answers caused the decline.
The paper describes this as a problem of persistence and unassisted performance. It does not say every person loses ability after using AI.
What is established
The study used experiments, not only surveys about people’s feelings. Researchers assigned people to conditions and tested them immediately after removing the tool. The pattern appeared in both math and reading tasks.
The Hacker News thread, a community technology discussion, had 54 points and 60 comments. That shows community attention. It does not prove the study is correct.
What remains unknown
The study measured only immediate effects after brief use. It did not test daily use over months. It did not show whether effects last for hours or days. Participants were U.S. online workers. The tasks were limited to fractions and SAT reading. A hint-based or Socratic AI might produce different results.
What to watch next
Future studies should follow the same people over longer periods. They should compare answer-giving tools with tools that offer hints. The main question is not whether AI helps. It clearly can. The harder question is whether AI can help people finish today’s task while protecting their ability to think tomorrow.
AI Helps First. What Happens When It Goes Away?
📰 Full story: After Ten Minutes With AI, Can People Still Keep Trying? A Study of 1,222 Participants
A study found a short-term drop after AI help disappeared.
randomized trial
A test that puts people into groups by chance.
persistence
The habit of continuing when work feels difficult.
Hacker News
A website where people discuss technology stories.
💡 The gist
- AI helped people answer questions at first.
- After AI disappeared, they solved fewer questions and skipped more.
- The study measured short-term effects, not a lifelong rule.
What did the researchers do?
Researchers from UC Berkeley (a U.S. university) and other schools ran a randomized trial. People were placed into groups by chance. The study included 1,222 people across three tests.
Some people used ChatGPT (an AI assistant that answers questions). Others did not. They solved fraction problems and SAT reading questions. The AI assistant used GPT-5. It could provide answers quickly.
After about 10 to 15 minutes, researchers removed the AI. Everyone then worked alone.
What happened?
AI users did better while the tool was available. After removal, their scores fell. They also skipped more questions.
In one math test, AI users solved 57% of the final questions. People without AI solved 73%. In a reading test, the numbers were 76% and 89%. The numbers are group averages. They do not describe every person.
The biggest drop appeared among people who asked for direct answers. People who asked for hints did better later. This was not a randomized comparison. It suggests a link, not a proven cause.
Why might this matter?
Learning needs more than quick answers. People need time to try, make mistakes, and fix them. Researchers think instant answers can make normal effort feel too slow. This may weaken persistence, or the habit of keeping going.
The paper does not say AI always harms people. It studied special tasks and short sessions.
What is still unknown?
Researchers do not know if the effect lasts for days. They did not study months of daily use. Participants were U.S. online workers. Hacker News, a technology discussion site, gave the story 54 points and 60 comments. That shows attention, not proof.
Future studies should test longer periods. They should also test AI that gives hints instead of full answers.
💬 Does AI make persistence weaker? HN comments give both sides
Some commenters say AI lets them do more. Others say work without AI feels harder for a while. Many also say the study needed fairer comparisons. Speed and performance claims here are personal reports, while the participant numbers come from the study description.
- The article and comments describe a study in which 354 people solved 15 fraction problems. The AI group did better at first, but after AI was taken away after 12 problems, its accuracy quickly dropped.[UC Berkeley article](https://vcresearch.berkeley.edu/news/using-ai-just-10-minutes-erodes-your-ability-persist-hard-things)
- The study also involved reading tasks, so stories about faster coding are not the same test. Critics say sudden removal may have tested people’s expectation that AI would be there. They wanted groups with a calculator, books, human help, no help, or another activity.
- One concern is that AI may improve the final result while making the same task harder without AI. A coder also reported needing recovery time after an AI-assisted task got stuck, though they did not describe a permanent loss.
- Other commenters reported the opposite: AI helped them pass technical obstacles and work faster on bigger, harder projects. They still said good work needs focus, design, debugging, and checking.
- A tool analogy says losing a calculator or saw makes manual work harder, but that is not automatically brain damage.
- One commenter reported that giving the AI less context prevents long detours; in their experience, it sometimes used 50,000–100,000 tokens chasing the wrong path. Narrow inputs and human review were offered as safeguards.
initial digest at 60 comments (revision 1). We fetched 82 comments and sampled 82 across the thread. These are HN users’ reports, not independently verified facts.
AI Helps First. What Happens When It Goes Away?
📰 Full story: After Ten Minutes With AI, Can People Still Keep Trying? A Study of 1,222 Participants
A study asked whether quick answers make hard work feel harder later.
AI
A computer helper that can answer questions.
ChatGPT
A computer helper that answers questions.
Hacker News
A place where people talk about technology.
What happened?
AI means a computer helper. It can answer questions quickly.
People at UC Berkeley (a U.S. university) ran a test with 1,222 people. Some people used ChatGPT (a computer helper that answers questions). Some people did not.
They did math and reading questions. After about ten minutes, researchers took the AI away. Everyone worked alone.
The AI group did well at first. Later, that group got fewer answers right. More people skipped questions.
The researchers think instant answers can make waiting feel hard. People also get less practice thinking alone.
This study did not test years of AI use. It tested what happened soon after. Researchers still need longer tests.
Hacker News is a place for technology talk. It gave the story 54 points and 60 comments. Many people noticed the story. Those numbers do not prove the study is true.
AI that gives hints may work differently. Researchers want to test that next.
💬 Does AI make it harder to keep trying?
Some people say AI helps them do big jobs. Some say going back to work without AI feels hard for a little while. Others say the test may not have been fair.
- The article and comments describe a study of 354 people solving 15 fraction problems. The AI group did better first, but its answers fell quickly when AI was hidden after 12 problems.[UC Berkeley article](https://vcresearch.berkeley.edu/news/using-ai-just-10-minutes-erodes-your-ability-persist-hard-things)
- The study also used reading tasks, so it was not a coding test. People said taking AI away so suddenly may be the problem. They wanted to compare calculators, books, people helping, and no help too.
- People reported different experiences. AI helped some people work faster on harder projects. Others needed time to get their focus back when AI work got stuck. A missing tool can make hand work harder, but that does not prove the brain is broken.
- One person reported that AI can wander if it sees too much information, sometimes using 50,000–100,000 tokens on the wrong path. Keeping its information small and checking its work can help.
initial digest at 60 comments (revision 1). We fetched 82 comments and sampled 82 across the thread. These are HN users’ reports, not independently verified facts.
💬 Does AI weaken persistence? HN commenters split over the study and its controls
The comments split between people who say AI expands what they can accomplish and people who feel temporarily less focused or persistent without it. Many also question the study’s abrupt AI removal and comparison conditions. Claims about speed, performance, recovery time, and token use are commenter self-reports unless attributed to the study description.
initial digest at 60 comments (revision 1). We fetched 82 comments and sampled 82 across the thread. These are HN users’ reports, not independently verified facts.