AI-mediated video communication reduces perceived trust and confidence in judgments. People watching AI-filtered video feel less certain about what they are seeing. They report lower trust in their interlocutors. The psychological response is one of wariness.
But their actual judgment accuracy — their ability to detect lies — remains unchanged.
The dissociation is structural. Trust and accuracy are decoupled by the AI mediation layer. The feeling of unreliability is a social and emotional response that does not correspond to an epistemic degradation. People are worse at trusting while being equally good at judging.
The effect intensifies when avatar users and non-users interact together. Mixed environments — where some participants use AI filters and others do not — produce the strongest trust reduction. The asymmetry matters more than the technology itself.
This creates a design problem that accuracy metrics cannot address. If a communication system preserves judgment accuracy but erodes trust, it has failed at something measurement-adjacent but distinct. Trust is not a noisy version of accuracy. It is a separate quantity that operates by different rules, responds to different stimuli, and degrades under conditions that leave performance intact.
The implication for AI-mediated communication: the system works but people don't believe it works. Building trustworthy tools is necessary even when — perhaps especially when — the tools are already accurate. The epistemic and relational dimensions of communication are independent, and AI mediation selectively damages the relational one.