New Study Shows Brief AI Use Quickly Erodes Persistence and Independent Problem-Solving

A peer-reviewed study presented at the 2026 Conference on Language Modeling (COLM) provides causal evidence that even short exposure to AI assistance can reduce human persistence and impair performance on tasks once the tool is removed. Titled 'AI Assistance Reduces Persistence and Hurts Independent Performance,' the paper—authored by researchers from Carnegie Mellon, Oxford, UC Berkeley, MIT, and UCLA—details randomized controlled trials with 1,222 participants across math reasoning and reading comprehension tasks. In experiments, groups using AI for initial problems outperformed controls but then showed sharply lower accuracy and higher rates of giving up when assistance was withdrawn after roughly 10-15 minutes. For instance, in one fraction-solving task, the AI group solved only 57% of unassisted test problems versus 73-74% for controls, with elevated skip rates. Effects were strongest among those seeking direct answers rather than hints. UC Berkeley's coverage and the arXiv preprint (2604.04721) underscore the findings' alignment with broader concerns about cognitive offloading and 'deskilling.' Commentators like Cal Newport have noted the unusually rapid onset compared to prior tech impacts. The study calls for AI designs that better scaffold long-term competence. This research adds empirical weight to anecdotal reports of AI fostering dependency, while highlighting that not all uses (e.g., hints vs. solutions) carry equal risks.