A recent study published in *Radiology: Artificial Intelligence* reveals that large language models (LLMs) may be significantly impacted by cognitively biased inputs, affecting their accuracy when answering radiology board-style questions. Conducted by researchers at the University of Toronto, the study assessed 10 LLMs with varying reasoning capabilities using 400 questions. Results showed that biased prompts could lead to accuracy declines of up to 44.9%. The study emphasizes the importance of developing safeguards to mitigate these biases, suggesting that both adversarial testing and prompt-level safeguards are necessary for reliable clinical use. The findings highlight the need for further research on LLM robustness against misleading inputs in radiology.
Wed, 26 Aug 2026 21:00:18 GMT | Radiological Society of North America | RSNA