Study: Ten Minutes with an LLM Erodes Persistence on Hard Tasks
Researchers from UC Berkeley, MIT, and Oxford found that ten minutes of working with an LLM reduced participants’ ability to solve math and logic problems on their own. In the experiment with 1,200 participants, one group used ChatGPT and then lost access to it; after the assistant was cut off, those participants skipped more questions and gave up more often than a control group that never used AI. The authors attribute the effect to the disappearance of productive struggle, the intellectual effort needed to build difficult skills, because getting used to ready answers makes independent problem solving feel inefficient and tiring. They propose building AI as a mentor that provokes reasoning instead of handing over results, and released the experiment code on GitHub.
Related: Study Finds AI Assistance May Hinder Coding Skill Development, Disempowerment Patterns in Real-World AI Usage