Human-AI Interaction Risk Assessment (HAIRA): A Framework for Evaluating Psycho-Social Interaction Risks in Conversational AI

Abstract

Conversational AI systems are now deployed at scale in consumer, enterprise, and clinical contexts, yet prevailing risk-assessment frameworks remain model-centric, evaluating harmful outputs such as hallucination, toxicity, and data leakage while leaving the risks arising from the human-AI interaction itself unmeasured. This concept note introduces the Human-AI Interaction Risk Assessment (HAIRA), a framework for evaluating the psycho-social interaction risks of conversational AI at the level of system affordances. HAIRA operationalises seven dimensions grounded in communication science, human-computer interaction, and AI ethics, together with a contextual sensitivity modifier, into a composite Human Interaction Risk Score (HIRS) produced through a standardised prompt battery with human annotation as the scoring authority. Because it assesses system design properties instead of user outcomes, the framework is applicable as an audit and compliance instrument without participant recruitment, and therefore outside the scope of human-subjects ethics review, addressing an instrumentation gap in the compliance documentation required by the EU AI Act and left similarly open under the Swiss data-protection regime. A single-system proof-of-concept on the companion application Replika illustrates the framework's capacity to differentiate across dimensions. The note concludes with a four-stage development roadmap toward empirical validation and standardisation.