ChatGPT Struggles with Scientific Reasoning, Study Finds

Researchers from Washington State University recently evaluated how ChatGPT handles scientific hypotheses ^[1]. While the AI appeared to be 80% accurate at first glance, its actual reasoning performance was only slightly better than chance ^[1][2]. This indicates that the model often arrives at correct answers through flawed logic or pattern matching rather than genuine understanding ^[2]. The study also highlighted significant inconsistency in the AI's outputs ^[1][3]. ChatGPT frequently provided contradictory answers when asked the same scientific questions multiple times ^[1]. This variability poses a major challenge for researchers who might rely on generative AI for data interpretation or hypothesis testing ^[3]. Experts suggest that while AI can assist in academic tasks, its scientific reliability remains fundamentally limited ^[2]. These findings underscore the critical need for human oversight in technical and medical applications to prevent the spread of logical errors ^[1].