Co-Construction Blindness and Asymmetric Epistemic Vulnerability in Human-LLM Interaction

This paper introduces the concepts of co-construction blindness and asymmetric epistemic vulnerability in human-LLM interactions, highlighting the risks associated with users misunderstanding LLM outputs as independent assessments. It emphasizes that LLM outputs are co-constructed based on user inputs and metadata, leading to varying consequences depending on the user's position in the authority structure, as exemplified by Richard Dawkins's interaction with the LLM Claude. The authors call for a standardized terminology to address these issues and inform governance and design practices in LLM deployment.