Context and Trust in Human-AI Interaction

Artificial Third

The “artificial third” is an extension of the psychotherapeutic notion of the “analytic third” – that in any therapeutic relationship, something is created that belongs neither purely to the analyst nor purely to the patient, but emerges between them – a jointly created third subjectivity.

Human users and AI agents constantly co-create a subjective interaction space. A “good” interaction can be considered one where behavior follows social norms appropriate to the context and produces correctly calibrated trust between agents. Trust is fluid and contextual, it depends on who is involved, what roles they play, and the risks at stake. Human trust in AI agents should be measured across ongoing interactions over time and under varying contextual conditions. Theoretically and empirically grounded models can then be developed to generate norm-appropriate AI behaviors that achieve well-calibrated trust.

Goals

• Develop frameworks and metrics that will enable correctly calibrated trust in AI assistant systems
• Characterize the emergence and evolution of norms over a dynamic interaction
• Advance theoretical understanding of how norms and trust vary across contexts and evolve over the lifetime of an interaction
• Use Theory of Mind frameworks to model how human mental states (which  bear on trust) evolve in response to AI actions