Cognitive Bias in Large Language Models: Implications for Research and Practice

AbstractThe use of large language models (LLMs) such as ChatGPT in clinical settings is growing, but concerns about their susceptibility to cognitive biases persist. Wang and Redelmeier's study reveals that LLMs are prone to biases, raising important questions about their role in medical decision-making. To prevent errors in decision-making with LLMs, it is recommended that clinicians aim to critically engage with LLMs (e.g. refuting their hypotheses rather than looking for confirmation) researchers should focus on identifying and evaluating collaborative strategies between AI and human decision-making. Furthermore, research on context-specific implementation is important. We need to ensure that AI complements, rather than replicates, human cognitive processes. (Funded by the Netherlands Organisation for Health Research and Development.).