New study: Heavy AI use weakens logical reasoning skills

Researchers Shang Wu, Hongyu Yao, Catarina Belem, Shuyuan Fu, Mark Steyvers, and Padhraic Smyth from UC Irvine conducted a controlled study on how both the frequency and quality of AI assistance affect human skill development in logical reasoning. Their paper, accepted at the Hybrid Human Artificial Intelligence (HHAI) 2026 conference, presents two key findings. First, the amount of AI usage matters: participants who relied heavily on AI underperformed relative to comparable peers when solving problems without assistance. In contrast, light AI users performed similarly to those who never used AI. Second, the informativeness of the AI’s output is critical. Low-information AI (providing minimal or unhelpful hints) did not improve immediate task performance and actually reduced learning, as measured by performance after AI access was removed. High-information AI (detailed, guided hints) boosted short-run performance but showed heterogeneous effects on learning: some users still failed to retain skills once the AI was taken away.

The study suggests that AI can act either as a complement or a substitute for human reasoning, depending on context. When AI provides rich, explanatory feedback and users engage actively, it may strengthen logical reasoning. But when users passively accept low-quality or overly simple answers, the AI undermines skill development. The authors emphasize that regulating AI access and usage design—such as limiting reliance on low-information outputs—will be crucial for educational and professional environments. The findings add empirical weight to concerns about “AI deskilling” and highlight the need for interfaces that encourage independent reasoning rather than quick answers.