A study involving researchers from Carnegie Mellon University, the University of California, Berkeley, the Massachusetts Institute of Technology, the University of Oxford, and the University of California, Los Angeles shows that after receiving approximately 10 minutes of AI assistance, participants were more likely to skip questions or give up when solving problems independently. The research team conducted three randomized controlled trials with a total of 1,222 participants.
The first trial asked 354 participants to complete 15 fraction problems, with one group working independently and the other allowed to use ChatGPT, though the latter could not continue using AI after question 12. The AI group had higher accuracy initially, but performance declined after AI access ended. Among the 307 participants included in the analysis, the AI group's subsequent accuracy was 57%, compared to 73% for the control group; the skip rates were 20% and 11%, respectively.
AI assistance may reduce learners' independent thinking ability
The second trial involved 667 participants and added a pre-test, while providing the control group with an assistance interface of the same format. Results showed the AI group's accuracy was 71%, lower than the control group's 77%. In the AI group, 61% of participants directly asked the AI for answers, achieving 65% accuracy; those who used AI for hints scored 76%, close to the control group.
The third trial had 201 participants complete SAT reading questions. The AI group scored 76% accuracy versus 89% for the control group; skip rates were 8% and 1%, respectively. The research paper, titled "AI Assistance Reduces Persistence and Impairs Independent Performance," has been published in a revised peer-reviewed version at a language model conference. The researchers suggest that directly obtaining answers may reduce learners' process of attempting and understanding through difficulty. The team recommends that AI systems adopt more tutor-style prompts to guide users through reasoning rather than defaulting to providing answers directly, thereby assisting rather than replacing human capability.