Ten minutes. That’s about how long it took for an AI assistant to change how hard people were willing to work on a problem once the assistant was taken away.
That is the headline finding of a new set of experiments by researchers from Carnegie Mellon University, MIT, the University of Oxford, UCLA and UC Berkeley. In three randomized trials with 1,222 online participants, people who had a chatbot on hand did better at first, then solved fewer problems and gave up more often when the help disappeared.
The paper, “AI Assistance Reduces Persistence and Hurts Independent Performance,” was presented at the Conference on Language Modeling (COLM), and UC Berkeley publicized it on Oct. 9. Here’s what the researchers did, what they found, and what it might mean for anyone who leans on AI at work or school.
How the Experiments Worked
All three experiments followed the same basic setup. Participants, all based in the US and recruited through the research platform Prolific, worked through a short series of problems. Half were randomly assigned to have an AI assistant in a sidebar. The other half worked alone.
The assistant was GPT-5, and it had been given each problem and its solution in advance. Participants could ask it for hints, explanations or simply the answer.
Then came the twist. After the 12th problem, the AI was removed without warning, and everyone finished the remaining problems on their own. Those final problems were the real test.
The three trials
- Experiment 1: 354 people solved basic fraction problems.
- Experiment 2: 667 people did a larger version of the fraction test, this time with a pretest to measure each person’s starting skill.
- Experiment 3: 201 people answered SAT-style reading comprehension questions.
Some people dropped out along the way, so the final analyses covered 307, 585 and 168 participants.
What the Researchers Found
The pattern was strikingly consistent. While the AI was available, the assisted group did well. As soon as it disappeared, their accuracy fell and, in two of the three experiments, they skipped more problems. The unassisted group held steady.
The numbers
- Fractions, first trial: after the AI was removed, the assisted group solved 57% of the test problems, compared with 73% for people who never had AI. They also skipped 20% of problems versus 11%.
- Fractions, larger replication: the gap narrowed but held, with a 71% solve rate for the AI group versus 77% for the control group. The difference in skipping was not statistically significant in this trial.
- Reading comprehension: 76% versus 89% solved, and 8% versus 1% skipped.
The researchers measured these effects after roughly 10 minutes of AI use, which is what makes the result stand out. This was not months of dependence. It was a single short session.
“It’s a striking finding, but in a way it supplies ammunition for a story that a lot of us kind of feel in our gut,” said Brian Christian, a co-author and research fellow at UC Berkeley’s Center for Human-Compatible AI.
How You Use It May Matter Most
The most practical finding came from Experiment 2, where participants were asked how they had used the assistant.
- 61% said they mainly used it to get direct answers.
- 27% said they used it mostly for hints or clarification.
- 12% said they barely used it at all.
The answer-seekers fared worst. Their solve rate on the final problems was 65%, compared with 77% for the no-AI group, and they skipped problems nearly twice as often. Their performance also fell compared with their own pretest scores, while the no-AI group improved slightly.
The hint-seekers looked very different. Their skip rate (5%) was no higher than the control group’s (7%), and their solve rate (76%) was close to it.
The authors are careful here: these comparisons are based on what people reported about themselves, and they show a correlation, not proof of cause. People who ask for answers may differ in other ways from people who ask for hints. Still, the pattern lines up with what teachers have long said about the value of working through a problem.
Why Would a Chatbot Make People Quit?
The researchers suggest a few explanations. Getting instant answers may reset what feels like a reasonable amount of effort, so slow, manual work suddenly feels inefficient. Leaning on AI also removes what educators call “productive struggle,” the frustrating but useful stretch where learning actually happens.
Christian put the irony bluntly: “We showed, ironically, that they are often helping us in ways that are kind of unhelpful.”
He also argued that the findings reach well beyond math worksheets. “This is not a story about fractions and SAT problems,” he said. Persistence, the paper notes, is central to building skills over time.
The Catch
This is a careful study, but it has real limits worth keeping in mind.
- Short, simple tasks. The tests involved basic fractions and reading questions, not complex real-world work. The authors explicitly warn against concluding that AI lowers persistence and ability on every kind of task.
- Online volunteers. Participants were paid a few dollars to complete the tasks online, which isn’t the same as a student studying for a final or a worker on deadline.
- Short-term effects. The study measured what happened minutes after the AI was removed. It doesn’t show whether the effect lasts, fades or builds over weeks and months.
- Mixed results on giving up. In the largest trial, the AI group didn’t skip significantly more problems overall, which the authors say may reflect differences in how people used the tool.
- A dropout imbalance. In the first experiment, more people dropped out of the control group, which the authors acknowledge may have widened the gap.
What This Means for You
None of this means you should stop using AI tools. It does suggest being deliberate about how you use them, especially when the goal is to learn something rather than just finish it.
- Ask for hints, not answers, when you’re trying to build a skill. Prompts like “give me a hint” or “explain the first step only” keep you doing the thinking.
- Try first, then check. Give a problem an honest attempt before opening the chatbot, and use the AI to review your work.
- Watch for the urge to quit. If a problem feels unusually tedious right after a session of AI help, that may be the effect this study describes.
- For parents and teachers: many AI tools now offer study or tutor modes that guide rather than solve. They may be worth turning on for homework.
What the Researchers Want From AI Companies
The authors argue that AI systems should be designed to support people’s long-term abilities, not just complete the task in front of them. That could mean an assistant that sometimes holds back the answer and acts more like a tutor by default.
“Just as the best human collaborators know when not to help, so too should AI,” they write. Christian added: “We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted.”
Some AI developers have started studying the same question. Berkeley’s writeup points to research from Anthropic on how AI assistance affects people’s coding skills, a sign the issue is getting attention inside the industry as well.
More on Contoh
- AI Chatbots Give a Narrower Range of Information Than a Google Search, Study Finds
- Common Sense Media Rates ChatGPT for Teens an “Unacceptable Risk.” OpenAI Disputes the Findings
- Opinion: Stop Asking Whether AI Will Take Your Job. Ask Which Parts of It Should Go