Using AI for just 10 minutes erodes your ability to persist at hard things

Brian Christian, a UC Berkeley research fellow and author of *The Alignment Problem*, has long explored AI’s societal impact while maintaining a daily notebook to counter concerns about overreliance. His latest research, co-authored with scholars from Carnegie Mellon, MIT, Oxford, and UCLA, reveals that even brief AI use can diminish cognitive endurance. The team presented their peer-reviewed findings at the Conference on Language Modeling, highlighting how quick access to answers may erode problem-solving skills over time.

In a series of randomized controlled trials involving 1,222 participants, researchers tested AI’s effect on persistence. In one experiment, 354 participants solved fraction problems either alone or with ChatGPT assistance. The AI-assisted group initially performed better but faltered after the tool was removed, while the unaided group maintained accuracy and completion rates. A second trial with 667 participants replicated these results, showing AI users abandoned tasks entirely once support vanished.

A third experiment with 201 participants using an SAT reading comprehension prompt confirmed the pattern: removing AI led to immediate declines in both persistence and accuracy. The findings align with broader concerns that AI overreliance reduces cognitive engagement, as users grow accustomed to instant answers and avoid the ‘productive struggle’ essential for skill development and discovery.

Christian and his collaborators argue that while AI systems are designed to be helpful, their current form often undermines long-term learning. They suggest AI responses should prioritize guidance over direct answers, functioning more like tutors than shortcuts. The study warns that overreliance in research and academia risks diluting the qualities that define expertise, such as deep immersion in data and collaborative problem-solving.