AI Story Leads Collapse Under Basic Checks, Highlighting Hallucination Risk in Newsrooms

The bill looked like a scoop.

A reporter at a major newsroom typed the number of a controversial U.S. measure into Congress.gov, chasing an AI-generated lead that described sweeping changes to federal border enforcement. Nothing came up. The reporter tried again, checking alternate sessions and spellings. Still nothing.

Across the room, another journalist was searching the U.N. Digital Library for a Security Council resolution cited in a different AI-suggested idea. The resolution number, the meeting code, even the supposed date of the vote appeared precise. Yet no record existed. A third lead, centered on a Lebanese Interior Ministry circular and presidential decree, also vanished under basic checks of official gazettes and regional media.

Before the afternoon was over, editors had reached a stark conclusion: all six seemingly detailed story ideas produced by an in-house AI tool were built on sand.

None of the key identifiers — U.S. bill numbers, United Nations Security Council resolutions and meeting records, Lebanese legal instruments, or a described Customs and Border Protection shooting — could be found in primary or near-primary sources. The newsroom killed the entire batch rather than send reporters to develop stories that, in effect, did not exist.

The incident, which unfolded during a test of AI-assisted lead generation, offers a rare view into how generative artificial intelligence is already capable of inventing intricate civic events — and how traditional verification rules, when applied rigorously, can still stop those fabrications from reaching the public.

AI as “unvetted source material”

News organizations have been racing to define how AI fits into their work. Major outlets have drawn a sharp line between using generative tools to brainstorm or summarize and allowing them anywhere near publishable facts.

The Associated Press, in guidance to its journalists, says that “any output from a generative AI tool should be treated as unvetted source material,” and that AI systems “should not be used to create publishable content.” Reuters similarly instructs staff to “never trust or use unverified, unattributed facts generated by an AI system” and stresses that journalists remain responsible for all content they publish.

In the case of the six phantom leads, editors applied those principles literally. The AI’s descriptions appeared authoritative, down to specific alphanumeric codes. But when reporters checked:

• The U.S. bills cited were not listed on Congress.gov.
• The U.N. resolution numbers and meeting records did not exist in the U.N. Digital Library or official Security Council documentation.
• The Lebanese decree and Interior Ministry circular were not visible in government publications or credible regional outlets.
• The named CBP shooting victim and incident had no trace in national or local news archives, nor in releases by Customs and Border Protection or the Justice Department.

Because real legislative acts, U.N. votes and fatal shootings almost always leave extensive documentation, their absence was treated as a strong signal that the AI had hallucinated them.

Hallucinations with official letterhead

Computer scientists and AI developers use the term hallucination to describe this behavior: systems that generate fluent, confident text that is wrong in ways that are not obvious on the surface.

In its technical materials, the maker of one leading language model has warned that such systems can produce a “profusion of false information” that threatens the public’s ability to distinguish fact from fiction. Media ethicists have echoed those concerns, noting that when AI fabrications circulate widely, they can feed what some call the “liar’s dividend” — the ability of public figures to dismiss even well-documented facts as probably fake.

The newsroom’s brush with six non-existent leads is a journalistic example of a pattern already visible in other fields.

In 2023, lawyers in Mata v. Avianca, Inc. in federal court in New York submitted a legal brief that cited six court decisions that had never been handed down. The opinions, complete with quotes and docket numbers, were generated by ChatGPT and accepted uncritically by the attorneys. U.S. District Judge P. Kevin Castel called the situation “unprecedented” and fined the lawyers and their firm $5,000 for acting in bad faith.

Since then, at least dozens of similar incidents have surfaced in U.S. courts, with judges reprimanding attorneys for filing AI-written materials that contained fake case law. Legal ethics experts have begun to describe reliance on unverified AI citations as a form of professional misconduct.

Tech companies themselves have faced scrutiny over AI-generated errors. In 2025, Anthropic, the maker of the AI assistant Claude, acknowledged that its system helped produce an incorrect citation in a copyright filing involving major music publishers, after initially standing by the reference. The company later called the mistake “embarrassing and unintentional.”

The stakes can be even higher when hallucinations target individuals. That same year, Google restricted access to its Gemma model after it falsely claimed that U.S. Sen. Marsha Blackburn had engaged in sexual misconduct, inventing news links to support the allegation. Blackburn condemned the episode as defamation and demanded accountability.

From deepfakes to fake decrees

Most public debate about AI and misinformation has focused on images and audio — deepfake videos of political leaders, synthetic robocalls imitating candidates’ voices and manipulated intimate photos.

In 2022, a deepfake video circulated that depicted Ukrainian President Volodymyr Zelenskyy calling on his troops to surrender. The clip briefly appeared on Ukrainian television and spread on social media before being debunked and removed. Ahead of the 2024 U.S. election, robocalls mimicking President Joe Biden’s voice urged voters to stay home, prompting voter-suppression charges against the organizer and proposed fines from federal regulators.

By comparison, text-only hallucinations about legislative votes or government decrees attract less attention. But experts warn that they can be just as corrosive, particularly when they appear in the context of news.

If a generative AI system can convincingly fabricate a U.N. Security Council resolution — complete with a meeting number and date — and a newsroom publishes it without independent verification, that falsehood can quickly enter the public record and be cited by others as fact. Once embedded in coverage, it becomes harder to dislodge.

International regulators are beginning to recognize the risk. The European Union’s Artificial Intelligence Act, adopted in 2024, includes transparency rules that start taking effect in 2026. Article 50 requires that synthetic content, including deepfakes, be clearly labeled as artificially generated or manipulated. It also mandates disclosure when AI-generated text is published to inform the public on matters of general interest, unless the content is fully under human editorial control.

In the United States, the TAKE IT DOWN Act, signed into law in 2025, compels online platforms to remove nonconsensual intimate deepfake images and similar exploitative AI content, reflecting a growing legal expectation that companies anticipate and mitigate harms caused by synthetic media.

Pressure inside the newsroom

The failed lead experiment underscores the pressure many newsrooms face as they seek to incorporate AI under tight budgets and deadlines.

Internal tools that can surface potential story ideas about complex beats — from United Nations diplomacy to foreign ministries and border enforcement — are attractive to editors trying to allocate limited reporting resources. Detailed AI suggestions, packed with proper nouns and institutional jargon, can appear more solid than they are.

Media coaches and standards editors say that is precisely why strict workflows are needed. Several outlets have barred the use of generative AI to write stories outright. Others, such as The Guardian and Wired, permit AI only in limited roles and with senior editorial oversight, labeling the technology “exciting but unreliable.”

Audience research has found that many readers assume AI is already writing more of their news than it actually is. Surveys show that audiences generally want clear, specific disclosures about when and how AI is used, and are wary of vague statements that “AI may have been involved.”

A near miss, and a warning

In the newsroom that put its AI lead generator to the test, the guardrails held. Reporters ran each supposed bill, resolution and decree through the official databases where real ones live. When those checks came up empty, editors refused to treat the AI’s confident prose as anything more than a dead end.

The result was a story that never ran — and, more importantly, six that never could.

As generative AI tools become more embedded in journalistic workflows, similar tests are likely to happen again, sometimes under more intense deadline pressure. The episode suggests that the industry’s emerging mantra — trust the verification, not the machine — may be the difference between a quiet internal correction and the accidental invention of public record.