Six AI-Generated Scoops Fell Apart Under Scrutiny, a Warning for Newsrooms

On a Tuesday morning in March, an experimental AI system delivered what looked like a reporter’s dream: six fresh story ideas about new laws abroad, unreported diplomatic deals and a high-stakes summit that had supposedly ended in a surprise agreement.

By the end of the week, every one of those stories was dead.

None of the legislation, summits or communiqués the system described could be found in parliamentary records, government websites, wire reports or credible local outlets. Editors, applying a longstanding rule that no story advances without at least one verifiable document or corroborating report, spiked the entire batch and ordered reporters to start over with leads rooted in the public record.

The episode offers a window into how newsrooms are beginning to collide with the limits of generative artificial intelligence — and how often seemingly authoritative AI material disintegrates when subjected to basic reporting.

A test that six stories failed

In this case, the AI tool wasn’t writing finished copy. It was set up to generate potential leads based on public information and past coverage.

One suggestion described a sweeping migration bill that had allegedly passed in a foreign parliament “with little international attention.” Another outlined a previously unreported regional security pact among mid-sized nations. A third asserted that an emergency summit had produced “a confidential annex” reshaping relations between two rival states.

On first read, the pitches looked sophisticated: plausible country pairings, named committees, quoted officials and references to supposed coverage in overseas media.

The verification process was more blunt. Researchers searched the relevant parliament’s bill tracker and legislative databases; no such measure appeared. They combed foreign ministry and presidential websites for the cited agreements and summit statements; nothing matched. They queried major wire services and the foreign outlets named in the AI-generated text; those stories did not exist.

Each time, the trail ended in the same place: nowhere.

“From the outside, the outputs can look like normal journalism — dates, names, places, quotes,” one editor involved in the review said. “But if you can’t tie even a single claim back to a document or a piece of attributable reporting, there is no story there.”

The decision, the editor said, came down to first principles: “Accuracy comes before novelty. If we can’t prove it, we don’t run it.”

Hallucinations far beyond the newsroom

The six-story flameout is not an isolated quirk. Across law, academia and publishing, generative AI systems have been documented producing detailed but fictional “facts” — a phenomenon researchers and developers now commonly call hallucination.

In one of the earliest and most widely cited incidents, a Manhattan federal judge in 2023 sanctioned two lawyers who had used ChatGPT to write a legal brief in a personal-injury case, Mata v. Avianca, Inc. The filing cited at least six court decisions that did not exist.

“The court is presented with a submission that cites non-existent cases with bogus quotes and bogus internal citations,” U.S. District Judge P. Kevin Castel wrote in his order, fining the lawyers $5,000 and calling the fake precedents “legal gibberish.”

Since then, legal-ethics experts have tracked a surge in similar episodes. A 2025 analysis by compliance firm VinciWorks described an “AI hallucinations crisis,” documenting more than 50 publicly reported instances in a single month in which lawyers had cited fabricated case law produced by generative tools.

Academic publishing has seen parallel problems. A study of reviews submitted to the 2025 Conference on Neural Information Processing Systems (NeurIPS) found 100 fabricated citations introduced by AI tools, with about two-thirds classified as “total fabrication” rather than misremembered or slightly altered references.

News organizations have already been burned when those tendencies creep into their own workflows. In 2025, the Philadelphia Inquirer and Chicago Sun-Times published summer reading lists that included several books that do not exist, including a supposed title attributed to novelist Isabel Allende. Both outlets later acknowledged that AI assistance had been used to assemble the lists and pledged to tighten their vetting.

Policies that treat AI as a source, not a reporter

Major wire services and broadcasters have responded by writing explicit AI rules into their ethics codes, often drawing a clear line between using AI to assist journalists and publishing AI-generated material.

“The Associated Press does not publish stories with AI-generated text,” the AP says in its standards for generative AI. Staff may use such tools for “translation, transcription and to generate story ideas,” but “the same standards of verification apply as with any other source.”

Reuters, in its guidelines for journalists, states that “Reuters pictures must reflect reality,” and bars the use of generative AI to create or alter photographs, graphics or video that are presented as news. Agence France-Presse’s 2025 ethics charter goes further, declaring that AFP “does not publish text, photos, videos, illustrations or infographics generated by artificial intelligence” and limiting AI use to internal tasks such as data analysis.

The National Press Foundation, which provides training for journalists, sums up the emerging consensus: “All AI-generated information requires human verification,” its AI principles say, adding that “journalists are responsible for the accuracy of all content they produce, regardless of the tools used.”

Local and commercial broadcasters have taken similar positions. Boston public radio station WBUR, in a 2024 ethics guide, allows AI to help summarize legal documents or organize information but prohibits its use to write scripts or mimic a host’s voice. The E.W. Scripps Co., which owns dozens of local TV stations, says in its 2024 journalism ethics guidelines that AI “should not impact the accuracy of the final work product” and restricts its use for scripting and visuals.

In practice, those rules often translate into the approach the March newsroom adopted: treat AI like an anonymous tip. Nothing runs until it has been reported out.

Adoption racing ahead of guardrails

Despite the concerns, AI use in newsrooms is accelerating. A guide for editors published by the Thomson Reuters Foundation in 2024, based on research in the Global South, found that roughly 81% of journalists surveyed were already using AI tools in some form. Only 13% worked in organizations with a formal AI policy.

Separate research on U.S. newspapers, based on an automated audit of about 186,000 articles in 2025, estimated that nearly 1 in 10 were partially or fully AI-generated. Opinion pieces at large national outlets were several times more likely than straight news to contain AI-written passages.

At the same time, technical evaluations suggest that hallucinations remain common in precisely the kinds of tasks journalists might delegate. One test of newsroom-style prompts — queries about specific court rulings, legislative histories and diplomatic agreements — found that leading models produced confident but false answers or citations in roughly 40% of cases.

Those numbers underscore the tension in understaffed newsrooms, where AI is often introduced with the promise of efficiency. Editors and reporters who have spent hours chasing AI-suggested “facts” that led nowhere say the technology can just as easily burn time as save it.

“It’s not that the system is broken,” said one newsroom manager who has overseen AI experiments. “It’s doing what it’s built to do — generate plausible text. Journalism is built to do something else: verify reality.”

A clash over what counts as knowledge

Researchers who study the intersection of AI and media sometimes describe the problem as an epistemological mismatch. Large language models are trained to predict the next likely word based on patterns in massive text datasets. They are not designed to check whether supporting documents exist, whether a quotation was actually spoken or whether a law was truly enacted.

Journalism, by contrast, is organized around attribution and evidence: every factual assertion is supposed to be traceable to a source, a document or direct observation.

That difference played out in the six-story review in March. The AI system produced coherent narratives; the newsroom asked for proof. When no proof appeared, the narratives were discarded.

Audience attitudes suggest that kind of restraint aligns with public expectations. A 2025 study by the Poynter Institute and the University of Minnesota found that nearly half of Americans surveyed said they did not want their news produced by generative AI. Many respondents said they supported clear disclosure about when and how news organizations use such tools.

Regulators, too, are beginning to focus on synthetic media and misinformation. The European Union’s Artificial Intelligence Act, approved in 2024, imposes transparency requirements on general-purpose AI systems, including obligations to make AI-generated content identifiable and to report serious incidents. In the United States, the Federal Trade Commission has warned that using AI in ways that mislead consumers — including deepfake audio or video that is not clearly labeled — can violate existing laws against deceptive practices.

For the journalists who killed six promising but unverifiable AI leads, those broader debates are not abstract. They are part of a daily calculus about what to publish — and what to leave in the notebook.

In a media environment where synthetic text, images and video can be generated in seconds, some of the most consequential editorial decisions may be the ones readers never see.