Just 10 minutes of AI use may reduce persistence on difficult tasks

Researchers warn that relying on instant answers could undermine learning, while hint-based AI assistance shows more encouraging results.

Just 10 to 15 minutes of using an artificial intelligence chatbot may make people less willing to tackle difficult problems independently, according to a peer-reviewed study involving more than 1,200 participants.

The research found that participants who used AI to solve simple mathematics and reading comprehension exercises performed worse and skipped more questions after access to the chatbot was removed.

The findings were presented at the Conference on Language Modeling (COLM), held October 6-9 in San Francisco, California.

The study was led by Grace Liu of Carnegie Mellon University, with co-author Brian Christian of the University of California, Berkeley.

Three experiments reveal declining persistence

The researchers conducted three randomized controlled experiments involving 1,222 participants, recruited through the online research platform Prolific. The final analysis included 1,058 participants.

In the first experiment, 354 participants were divided into two groups. One group had access to an AI chatbot while solving 12 fraction problems, while the other completed the same exercises without technological assistance.

The AI users could request direct answers or ask for hints. Researchers then unexpectedly removed access to the chatbot and gave both groups three additional problems.

The results showed a noticeable difference. Participants who had used AI solved 57% of the final questions and skipped 20%. Those who had never used the chatbot solved 74% and skipped only 11%.

A second experiment involving 667 participants produced similar results. After access to AI was removed, the chatbot group solved 71% of the final problems, compared with 77% among participants who had worked independently.

Although the difference in skipping rates was not statistically significant across the full sample, the researchers identified an important distinction in how participants used the technology.

Among AI users, 61% relied on direct answers, while 27% used the chatbot primarily for hints or clarification.

Participants who requested direct answers solved 65% of the final problems, compared with 77% in the control group.

The researchers cautioned that comparisons between different patterns of AI use were observational and did not necessarily establish a direct causal relationship.

Reading comprehension also affected

The third experiment examined whether the same pattern could be observed in reading comprehension.

A total of 201 participants answered five questions before chatbot access was disabled for a further three questions.

Those who had used AI solved 76% of the final questions and skipped 8%.

By comparison, participants who had worked without AI solved 89% and skipped just 1%.

The findings suggest that reduced persistence may not be limited to mathematical exercises but could also affect tasks requiring reading and independent reasoning.

Researchers warn against replacing the learning process

Liu told Live Science that one of the study's most important findings was the relationship between reliance on AI and declining persistence.

"People who offload more have higher declines in persistence," she said.

The researchers linked the findings to the concept of desirable difficulties, a term used in cognitive science to describe the productive effort and struggle through which people develop knowledge and skills.

Although tools such as calculators also allow people to outsource certain tasks, Liu argued that AI is different because it can take over reasoning across a much wider range of activities.

One possible explanation is that users become accustomed to receiving immediate answers and consequently become less willing to spend time working through problems independently.

AI tutors could offer an alternative

The researchers stressed that their findings should not be interpreted as proof that AI inevitably damages learning or long-term intellectual abilities.

The experiments lasted only 13 to 15 minutes, and participants were assessed immediately after chatbot access was removed.

The researchers said longer-term studies are needed to determine whether repeated reliance on AI has lasting effects on knowledge, competence and problem-solving abilities.

They also suggested that AI systems should be designed more like educational tutors, guiding users towards solutions rather than simply providing answers.

The second experiment offered some support for this approach, as participants who used AI for hints did not show the same decline in persistence compared with the control group.

In a statement published by the University of California, Berkeley, Christian said the goal should be to develop technology that strengthens human abilities rather than replaces them.

The findings add to the growing debate over the role of artificial intelligence in education and everyday learning, highlighting the importance of maintaining independent thinking and problem-solving skills as AI tools become increasingly accessible.

Source: Live Science; University of California, Berkeley. Study presented at the Conference on Language Modeling (COLM).