Does revealing AI identity help or hurt user trust?

The hybrid society study (N=975) reveals that AI identity disclosure is neither uniformly beneficial nor harmful — it produces a dual temporal effect that only becomes visible through repeated interaction.

Short-term: Disclosing that a partner is AI evokes anti-machine bias. Selectors initially choose AI partners less frequently than when identity is hidden. This is consistent with prior one-shot studies showing that AI labeling reduces cooperation and trust.

Long-term: With repeated interaction and transparent outcome feedback, selectors learn to associate AI identity with reliable, prosocial behavior. The initial bias reverses as empirical experience overrides prior beliefs. AI partners eventually outcompete human partners.

The key mechanism is outcome feedback. When selectors can observe that AI partners consistently return more, with less variance, and in line with their messages, they update their beliefs. Without this feedback loop (as in Study 1 with hidden identity), no learning occurs — selectors cannot calibrate because they cannot attribute outcomes to partner type.

This finding challenges three common positions:

1. "Always disclose" — disclosure imposes a real short-term cost; ignoring this cost is naive
2. "Never disclose" — without disclosure, the learning mechanism that produces calibrated trust cannot operate
3. "One-shot studies generalize" — most prior transparency research uses single interactions, missing the temporal reversal entirely

The parallel to Does chatbot personalization build trust or expose privacy risks? is structural: both are trust-risk trade-offs where the temporal dimension determines the net effect. Personalization ratchets expectations upward over time; disclosure enables belief calibration over time. Both show that one-shot findings are misleading for longitudinal design.

The policy implication: the EU AI Act's push for mandatory AI disclosure may impose short-term costs but enable long-term trust calibration — provided the interaction context includes outcome feedback that allows users to learn.

Asymmetry across roles. The dual temporal effect describes the disclosed-counterpart case. The disclosed-author or undisclosed-ghostwriter case appears to follow a different pattern. Since Do writers actually prefer AI-edited versions of their own text?, when AI is the silent author rather than the disclosed counterpart, preference flips toward the AI version from the start — no anti-AI bias, no learning loop required. The two findings together describe a complete picture: disclosure produces bias-then-calibration when AI is positioned as a partner; non-disclosure produces immediate preference when AI is positioned as a tool that produces output the user claims. The temporal dynamics of disclosure depend on the role AI is presumed to play, not just the disclosure status.