Study Finds Brief AI Use Can Undermine Persistence on Difficult Tasks

Summary

A UC Berkeley co-authored study examined whether brief reliance on an AI tool changes how people perform difficult tasks without assistance. Across randomized controlled trials involving 1,222 online participants, researchers tested fraction problems and an SAT reading-comprehension prompt. In the first experiment, 354 participants either solved 15 fraction problems unaided or could use ChatGPT, including to request answers. The AI-access group initially performed more accurately, but after the tool was removed following 12 problems, its accuracy and persistence fell sharply. A second experiment with 667 participants produced a similar pattern: participants who had used AI were more likely to answer incorrectly or give up, while the unaided group persisted and finished more successfully. A third experiment with 201 participants found comparable declines in persistence and accuracy on the reading task after AI access ended. The researchers describe the result as evidence that rapid answers may displace the “productive struggle” involved in building independent understanding, though the article presents the findings as a study of these tested conditions rather than proof that every form of AI use has the same effect. The updated peer-reviewed paper was presented at the Conference on Language Modeling. Co-author Brian Christian argued that AI systems should augment human abilities and, where possible, act more like tutors than answer engines, particularly in education and research.