The Trust Filter

When video communication is mediated by AI — retouched, background-replaced, or rendered as avatars — viewers trust the speaker less and feel less confident in their judgments. But their actual ability to detect lies doesn’t change.

Two experiments, 2,000 participants, three conditions of increasing AI mediation. The results separate trust from accuracy: AI mediation degrades the former without affecting the latter. People who watch AI-mediated video are no worse at distinguishing truth from lies than people who watch unmediated video. They’re just less sure of themselves.

This cuts against two common concerns. First, the fear that AI mediation will help liars: it doesn’t, because lie detection accuracy is unchanged. Second, the theory that people detect lies from visual cues that AI might alter: they don’t, which is consistent with decades of research showing that visual cues are unreliable indicators of deception in any medium.

The damage is to confidence and trust, not to judgment quality. When some participants use avatars and others don’t, trust drops most sharply — the asymmetry itself is the problem. A call where everyone uses avatars is less threatening than one where some people are visually present and others are AI-rendered. The uncanny valley operates at the social level: not within a single face, but between faces that play by different rules.

The practical implication is a design constraint. AI-mediated communication tools need to maintain trust and confidence, not just accuracy. A tool that preserves detection accuracy while destroying trust is not neutral — it degrades the social infrastructure that communication depends on.