Does telling people an AI wrote something actually stop them from believing it?

The Thin Line ablation manipulated audience awareness of AI involvement across three groups (A, B, C) and measured both critical engagement and sway. Group A — unaware of AI — perceived the LLM as more competent. Groups B and C — aware or suspecting — were more critical of the arguments. But sway proportions across the three groups ranged from 34% to 62%, with the LLM still moving substantial fractions of audiences who knew an AI was involved. Disclosure raised scrutiny without collapsing effect.

This challenges the assumption baked into much policy thinking — that AI labeling functions like advertising disclosure, and that once disclosed, the persuasive force decays sharply. The Thin Line evidence suggests disclosure modulates the channel through which AI influence operates rather than blocking the channel altogether. Aware audiences shift toward central-route processing (more scrutiny, more counter-arguing) but counter-arguing does not zero out the persuasive content; it leaves a residual sway proportion that is still meaningful at scale.

This sharpens we lack a cultural position on AI-generated discourse — unlike advertising which we already discount. We have a fully developed cultural reflex for advertising — the disclosed-paid-content posture is decades old and supported by school curricula, regulation, and consumer literacy. We do not yet have an analogous reflex for AI-generated discourse. The gap is not just rhetorical; it is measurable in residual sway proportions when AI authorship is known.

It also connects to Do people prefer AI moral reasoning when they don't know the source?. The anti-AI bias is real but bounded — it raises the threshold for acceptance without making AI arguments inert. Combined: people prefer AI moral content when blind, become biased against it when revealed, and yet are still moved by it when revealed. Three findings, one design implication: disclosure is a necessary but not sufficient safety mechanism.

For writing about AI authorship and false-punditry, the operational point: a "this was written with AI" label is not a neutralizer. It is a critical-route activator with a partial residual. Designs that lean on disclosure as the primary defense should be paired with content-side interventions, not treated as complete on their own.