Overview
- Researchers from Carnegie Mellon, MIT, Oxford and UCLA report in an arXiv paper that short chatbot use boosted speed and accuracy during tasks but hurt solo performance once access ended.
- Across three online experiments with about 1,222 participants, people worked on fraction arithmetic and reading comprehension with or without an assistant.
- When help was removed without warning, previously assisted users skipped more questions and answered fewer correctly than peers who never had AI.
- The decline was strongest when people asked for direct answers, while those who used the assistant for hints or explanations showed little to no drop-off.
- The team recommends AI that coaches and scaffolds rather than solves outright, and they note the study is a preprint from controlled settings that needs real-world follow-up.