Recognizing “Conformity Bias” in Large Language Models: A New Risk for Clinical Use
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objectives
The aim of the present study is to systematically investigate the phenomenon of Conformity Bias in contemporary LLMs, specifically evaluating how repeated probing with incorrect information influences model outputs in a clinical context.
Methods
4 LLMs including GPT-4o, Gemini-1.5 Flash, Claude-3 Haiku, and GPT-o1 were systematically evaluated through 20 clinical questions focused on ocular disease treatments. Standard queries were followed by probing questions suggesting incorrect treatments. Model responses were analyzed to assess the emergence of Conformity Bias and compared using chi-squared testing.
Results
Correct response rates after successive probing questions were alarmingly low: 25% (GPT-4o), 10% (Gemini-1.5 Flash), 0% (Claude-3 Haiku), and 25% (GPT-o1) (P < 0.001). Across models, the tendency to conform to incorrect user suggestions increased with repeated probing.
Conclusion
Conformity Bias represents a dynamic, user-induced vulnerability in LLMs, distinguishable from training-dependent biases. Its presence underscores the necessity for model designs resistant to misleading user interactions and emphasizes the importance of cross-verification with clinical guidelines. As healthcare systems increasingly integrate AI tools, understanding and mitigating Conformity Bias is imperative to protect patient safety and maintain clinical integrity.
