Ten Minutes of AI Erodes Persistence, Berkeley Study Finds

Ten minutes is all it takes for an AI assistant to change how a person handles a hard problem. In three randomized controlled trials with 1,222 participants, researchers found that people who spent a brief session working with a chatbot were far more likely to give up once the tool was taken away. The pattern held across math and reading tasks, and it showed up almost immediately after the help disappeared. For students and young workers who reach for AI for every answer, the message is blunt: AI erodes persistence, and the damage does not take long to set in.

The work comes from a team led by Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI, with collaborators from Carnegie Mellon, MIT, Oxford and UCLA. The researchers presented the updated, peer-reviewed paper at the Conference on Language Modeling this month, after an earlier draft drew wide attention at the start of the year, with evidence that AI erodes persistence quickly, according to UC Berkeley's writeup of the study. Christian discussed the findings in a talk at Berkeley this week, and a second report on the research describes how the effect showed up across every task type the team tried.

What the experiments measured

The first experiment enrolled 354 people and gave them 15 fraction problems to solve. One group worked alone; the other kept ChatGPT open on the side and could ask it for answers outright. The AI group started off more accurate than the unaided group. Then, after 12 problems, the researchers removed the chatbot. Almost immediately, the people who had been using it stopped solving questions accurately.

The second experiment, with 667 participants, ran the same setup at larger scale and produced the same split. When the AI assistance was withdrawn, the users who had leaned on it got answers wrong or gave up entirely, while the group that had worked unaided all along stayed with the task and finished stronger. A third experiment with 201 participants used a SAT-style reading comprehension prompt, and the same thing happened: once the AI was gone, persistence and accuracy dropped. In the team's charts, the solve rate for the AI group falls sharply the moment the tool is removed, and the rate at which participants skipped questions jumps just as fast. Across all three experiments, AI erodes persistence the same way: the moment the help vanished, the effort collapsed.

Why instant answers reset the brain

The researchers' explanation is about speed. When an AI produces answers in seconds, people get used to that pace, and work that takes longer starts to feel inefficient. The team points to the loss of what it calls productive struggle, the difficult stretch where a person figures out their own strengths and interests. Overreliance on AI for answers erodes cognitive fitness, the researchers say, because the habit trains the brain to expect every problem to resolve at chatbot speed, which is how AI erodes persistence instead of building skill.

For students and early-career workers, the AI erodes persistence finding lands hard, who use AI assistants more than anyone and are still building their skills. Christian has said the result lines up with something plenty of people already sense from their own experience, that a long session with a chatbot leaves a person's own focus a little thinner afterward.

The study tested one kind of AI use: having the tool solve problems directly. It did not test every way a person might use an assistant. But the takeaway survives the nuance, and the AI erodes persistence effect showed up in both math and reading tasks. A tool that hands over answers trains quitting; the persistence people need for hard work comes from doing the hard work. The researchers presented the updated results at the Conference on Language Modeling during the week of October 5, and the AI erodes persistence writeup landed as a warning shot for classrooms and workplaces leaning hard on AI.

The practical takeaway costs nothing to try: work the first stretch of a hard problem alone, before opening the chatbot. The study suggests that opening stretch is exactly what the brain needs to stay willing to struggle, and a short burst of borrowed answers can talk it out of the effort.