Study Finds 10 Minutes of AI Tool Use Weakens Persistence on Difficult Tasks

At the Conference on Language Modeling (COLM), researchers from UC Berkeley, Carnegie Mellon University, MIT, Oxford University and UCLA presented a peer-reviewed study titled "AI Assistance Lowers Persistence, Impairs Independent Performance." The study, based on a randomized controlled trial with 1,222 online participants, found that as little as 10 minutes of using an AI tool can significantly weaken people's ability to focus and persist on difficult tasks.

The team ran three experiments. In the first, 354 participants were split into two groups; one group solved basic fraction problems independently, while the other could consult a ChatGPT chatbot for help. The AI-assisted group initially performed more accurately, but when the tool was removed at question 12, their accuracy dropped sharply and their rate of skipping questions rose. A second experiment with 667 participants showed a similar pattern: without AI, those who had used it were more likely to answer incorrectly or give up, while the group that never used AI stayed on task and ultimately performed better. A third experiment with 201 participants using SAT reading comprehension questions confirmed the same result: persistence and accuracy both declined after AI was withdrawn.

The researchers suggest that over-reliance on AI-generated answers reduces cognitive adaptability. When AI provides answers within seconds, people become accustomed to that speed and lose what the authors call "productive struggle"—the process of grappling, testing and persisting that helps individuals discover their strengths and interests. The findings echo earlier work such as the "Google effect" published in Science in 2011 and the concept of "desirable difficulties," supporting the idea that ease often comes at the cost of learning.

An earlier version of the study drew attention earlier this year, with coverage from Wired, Psychology Today and The New York Times. The version presented at COLM is the updated, peer-reviewed iteration. The researchers stress that AI should act as a mentor rather than a substitute, and that the goal should be to enhance human capabilities rather than replace them.