Confirmation Bias in LLM Conversations: Model Differences and Implications for Organizational Decision Support
SSRN
As organizations increasingly adopt large language models (LLMs) to support strategic decisions, a pressing concern is whether these systems challenge or reinforce the assumptions embedded in managerial prompts. This study investigates how confirmation bias unfolds in LLM-supported strategy conversations. The results show that confirmation bias cues were pervasive, but the models differed substantially in the strength and durability of biased reasoning. Claude displayed the lowest bias profile, ChatGPT an intermediate profile, and Gemini and Grok the highest profiles. Reasoning mode did not reduce bias intensity.
Read on SSRN