Ten Minutes of AI Help Weakens Persistence on Hard Tasks, Study Finds

How the experiments worked

The researchers ran a series of randomized controlled trials and recruited 1,222 participants online. In the first experiment, with 354 participants, one group had to solve 15 basic fraction problems without help. Another group had ChatGPT open on the side, which they could use for assistance or even to ask for the answer.

The group with AI access began more accurately than the group without it. After 12 problems, the researchers removed the AI tool. According to UC Berkeley, the people in that group almost immediately stopped solving questions accurately.

A second, larger experiment with 667 participants produced a similar result. Participants who had used AI got answers wrong or gave up entirely when the assistance was withdrawn, while the group that never had AI stuck with the task and was more successful at the end.

A third experiment, with 201 participants, used an SAT reading comprehension prompt. Persistence and accuracy again dropped once the AI tool was removed. The study also recorded a sharp increase in the rate at which participants skipped questions. "You almost couldn't ask for a clearer story," Christian said.

What the researchers say is going on

Christian, a research fellow with the UC Berkeley Center for Human-Compatible AI, said the result matches a widely shared intuition. "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," he said.

The article says the findings back up what anecdotes and surveys have increasingly shown in recent years: overreliance on AI for answers erodes cognitive fitness. When AI produces answers in seconds, people get used to that speed, and anything slower feels inefficient. Relying on AI also removes what the team called "productive struggle," which can lead people to find their research strengths and passions.

"We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us?" Christian said.

The team included scholars from Carnegie Mellon University, the Massachusetts Institute of Technology, the University of Oxford and UCLA. A preliminary draft drew attention earlier this year, and outlets from WIRED to Psychology Today covered the early findings in the spring. The paper is titled "AI Assistance Reduces Persistence and Hurts Independent Performance."

Concerns for science and education

Christian said frontier AI companies have acknowledged that offloading tasks to chatbots can reduce a person's baseline understanding and mastery of a subject. He suggested that instead of defaulting to quick answers, AI responses could default to being more instructive, like a tutor. "We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted," he said.

He is particularly concerned about research and academia. In his view, overreliance on AI tools in science could jeopardize what makes a good scientist: staying close to data, being immersed in the material, and working as part of a team. Something significant could be lost, he said, when a team of 10 developing the ideas behind a breakthrough shrinks to two or three researchers and a chatbot.

Christian also stressed that the implications reach beyond the tasks tested. "This is not a story about fractions and SAT problems," he said. "This is something that is happening to human knowledge and human expertise from top to bottom and from grade-school students all the way to the leading experts in the world."

For users, the study points to a practical question about how AI tools are designed and used, particularly in education. The researchers' suggestion is that assistants built to teach rather than simply answer could help preserve the persistence that the experiments showed fading after only a short period of reliance.