UC Berkeley Study: Short-Term AI Use Boosts Accuracy but Increases Task Abandonment
A new study finds that while AI helps solve problems immediately, users who rely on it are more likely to give up when solving tasks independently.
According to a report by ZDNet Korea, researchers from the University of California, Berkeley and other institutions have found that using artificial intelligence (AI) to help solve problems increases the likelihood of getting the right answer in the short term, but makes people more likely to give up when they have to solve problems on their own without AI. These changes were observed even after just about 10 minutes of AI use.
The research team conducted three randomized controlled trials with a total of 1,222 participants. The experiments divided participants into groups receiving AI help and those solving problems without help, comparing performance changes after stopping AI use. In the first experiment involving 354 participants solving basic fraction problems, those initially using ChatGPT had higher accuracy rates. However, after the 12th question, when AI use was stopped, their accuracy dropped rapidly. In contrast, participants who solved problems without AI from the start maintained relatively stable performance.
Similar results appeared in the second experiment (667 participants) and the third experiment (201 participants, SAT-style reading comprehension). When AI assistance was removed, these participants increasingly got questions wrong or skipped them entirely. This effect was particularly pronounced among participants who relied on AI to provide direct answers. The researchers suggest that becoming accustomed to an environment where AI provides immediate answers can make the process of thinking and trial-and-error feel relatively inefficient, leading to easier abandonment of tasks when working without AI.
The study analyzes short-term changes and does not prove that long-term AI use degrades intelligence or learning ability. The researchers emphasized that AI should evolve beyond simply providing answers, towards guiding users to find solutions themselves step-by-step.