Addressing Social Misattributions of Large Language Models: An HCXAI-based Approach

LLMs can create dangerous mismatches between what designers intend and what users perceive: The paper identifies that Large Language Models' ability to simulate roles and personas can lead users to attribute social qualities (like trustworthiness or expertise) that designers never intended, particularly risking emotional manipulation in sensitive contexts like mental health support.

Current AI transparency frameworks are incomplete for social risks: The Social Transparency framework, while valuable, doesn't adequately address how users form social perceptions of AI systems. The authors argue this gap leaves users vulnerable to misplaced trust and emotional reliance on systems that aren't designed to provide genuine care or expertise.

A simple addition can improve responsible AI design: The paper proposes adding a fifth "W-question" to clarify what social attributes designers intentionally assign to LLMs. This straightforward addition to existing frameworks could help align designer intentions with user expectations and reduce harmful misattributions.

Social misattributions create multiple harms beyond technical failures: The research highlights that risks include not just emotional manipulation, but also "epistemic injustice" (unfair treatment in knowledge-sharing) and unwarranted trust—meaning users may follow harmful advice from systems they've incorrectly perceived as qualified experts.

Human-centered AI requires understanding social perception, not just technical capability: The paper emphasizes that responsible AI development must consider how humans actually interpret and relate to AI systems socially, not just whether the systems function technically as intended.