April 2026
 

Study finds AI unreliable at identifying retracted research papers
 

A recent study published in the Journal of Clinical Anesthesia highlights limitations in using AI tools to identify retracted research papers. Evaluating 21 large language model (LLM) chatbots, researchers found that, on average, the systems correctly identified fewer than half of retracted papers, while also producing a number of false positives.

Follow-up testing a few months later using more-direct and concise prompts showed a shift in how the chatbots responded: whereas earlier outputs tended to more clearly classify papers as retracted or not, with variable accuracy, later responses more often hedged, describing papers as “possibly retracted” or “worth double-checking.” The researchers note that such ambiguity offers limited practical value, underscoring the importance of careful verification when using AI-assisted tools in literature-review workflows.