Overview
- An international team from Oxford, MIT, UCLA, and Carnegie Mellon ran controlled tests on fraction arithmetic and reading comprehension to measure the effects of AI help.
- Participants who used AI and then lost access quit more often and solved fewer problems than peers who never used it, showing reduced persistence and weaker solo performance.
- The drop-off appeared after only 10 minutes of AI-assisted work and showed up across multiple trials with hundreds of people, including samples of about 350 and 670.
- The authors cautioned that small short-term gains can mask a longer slide in effort and focus, likening the risk to a slow “boiling frog” effect that becomes hard to reverse.
- The team said AI can remove “desirable difficulties” that build skill over time, prompting a call to design tools more carefully for classrooms and training programs.