About 10 minutes of relying on AI assistance for problem-solving can significantly impair a person’s ability to persist through difficult tasks, according to a study coauthored by Brian Christian, a research fellow at the Center for Human-Compatible Artificial Intelligence at UC Berkeley.
The researchers observed human-AI interactions through a series of tasks, including mathematical reasoning and reading comprehension problems from the SAT given to more than 1,200 participants.
They found the experimental group that was able to use AI assistance for part of the test were less accurate and less persistent in solving the remaining problems once it was taken away, than those who did not have access to AI in the first place.
According to the paper, the drop in accuracy and persistence emerged after only spending about 10 minutes interacting with AI.
The research began as a project with first author Grace Liu, a Ph.D. student at Carnegie Mellon University, and UCLA professor Rachit Dubey, the paper’s principal investigator. The two were inspired by the idea of “normalization to change.”
“We noticed in our own work that we were becoming normalized to increasing AI assistance in writing, coding, and research-related tasks and wanted to study this effect cognitively,” Liu said in an email.
The paper only examines the short-term effects during the time period when the test was administered. However, Christian stated in an email that he believes the long-term effects of AI use could be a “crucial area for future study both by ourselves and other researchers.”
Christian said he noticed a substantial risk in education and research when it comes to AI integration, one that is already unfolding.
“I think it’s vital to notice that the incentives of ‘maximize your output of published papers’ have become unglued from ‘develop your own mind and the minds of others into becoming better scientists and thinkers’” Christian said in the email. “This shift has happened quite suddenly over the last 1-2 years and needs to be taken very seriously by students and faculty alike.”
According to both Christian and Liu, there are ways in which AI can be used in order to complete tasks without losing out on important skills such as persistence in difficult tasks and problem-solving.
In the study, the researchers compared the group of people who used AI to get direct answers to the group who only asked AI for hints. They found that the subjects who asked AI for hints performed more similarly to the control group, who did not use AI at all.
“This is an important clue that the way we use AI seems to matter greatly, and this is something we are looking to investigate more closely,” Christian said in an email.