Newsrooms Test AI for Tips — and Find a Familiar Problem: It Makes Things Up

The political desk’s new helper had a promising assignment: scour recent developments and suggest fresh story ideas about global politics.

It delivered seven leads. There was an urgent U.N. Security Council resolution, a surprise White House order, a shake-up election overseas and more. On the surface, it looked like a full week’s worth of coverage.

There was one problem. None of it had happened.

When editors and engineers checked each “lead” against primary sources — U.N. resolution databases, the Federal Register, national election authorities and global wire services — not a single item could be verified. Several involved non-existent operations, executive orders, officeholders or elections that appear nowhere in any public record as of early April 2026.

Because the system had been configured to treat accuracy as its top priority, it did something unusual in an industry built on constant output: it passed no story leads at all to a reporter.

The episode, described by people involved in the AI-assisted workflow, captures a collision now playing out across newsrooms: generative artificial intelligence systems are confident, fast and available around the clock, but they also have a well-documented tendency to “hallucinate” — to generate fluent, plausible-sounding claims that are simply not true.

That behavior is increasingly running into one of journalism’s oldest rules: if it cannot be verified, it does not run.

“AI is not a source”

News organizations have rushed to experiment with large language models as tools to help reporters summarize documents, translate interviews and sift public records. At the same time, they are drawing strict lines around using AI as a source of facts.

The Associated Press, which has used automation for years to help generate corporate earnings stories and sports summaries, states that its news report is guided by “accuracy, fairness and speed.” In its guidance on generative AI, AP says the technology should be used “mindfully” and makes clear that AI is not allowed to produce publishable news copy or to alter the factual content of imagery.

“AP staff must ensure that the material we publish is accurate,” the standards note say, adding that any AI-assisted work must be “thoroughly vetted and edited.”

Reuters, which has also tested AI tools in its newsroom, is blunter. “All facts, sources and claims generated by AI must be independently verified and fact-checked by Reuters journalists,” the company’s standards state. Editors remain fully responsible for any content that incorporates AI output.

Those rules mirror a broader consensus among journalism trainers and press freedom groups: AI might draft an email, format a transcript or suggest a possible angle, but it cannot be trusted as an authority on what is true.

What “hallucination” means in practice

In technical literature, an AI “hallucination” is an output that sounds coherent but is factually wrong or unsupported by any source.

In practice, systems trained on vast libraries of text learn to predict what words are likely to come next. When asked about a “U.N. resolution in March 2026 sanctioning Country X,” a model may string together a resolution number, a list of sponsors and a vote tally that resemble real resolutions but do not match any actual document.

Recent research has found that such errors are persistent, even in advanced models and especially on complex, verifiable tasks. A LiveScience report in 2025, citing internal testing from OpenAI, noted that one leading model hallucinated on roughly one-third to nearly half of questions in a benchmark focused on reasoning about people and real-world facts.

Legal and policy experts say that pattern extends to domains where precision is critical. Studies of AI performance on legal questions have found models inventing court cases, misquoting statutes and misrepresenting procedural rules unless they are tightly constrained by external databases.

The stakes of getting it wrong

Courts have already punished lawyers who relied on AI tools without checking their work.

In 2023, a federal judge in New York sanctioned two attorneys in Mata v. Avianca after they submitted a brief that cited six non-existent cases, all generated by ChatGPT. The lawyers told the court they had “no intent to deceive,” but the judge imposed a $5,000 fine and called the episode an unprecedented failure of professional duty.

Similar incidents have followed. In 2025, U.S. District Judge Anna Manasco in Alabama sanctioned three attorneys whose filing in a prison conditions case contained “completely made up” case citations that came from ChatGPT. “Fabricating legal authority is serious misconduct that demands a serious sanction,” Manasco wrote, removing the lawyers from the case and ordering them to distribute her ruling to colleagues.

Appellate courts in Utah and other states have issued their own reprimands and fines for AI-generated citations, often requiring attorneys to complete training on responsible technology use.

Elections have seen their own wave of AI-related risks. In early 2024, a project that brought together the nonpartisan news site Proof News, AARP and more than 40 U.S. election officials tested five leading AI models — including systems from major U.S. tech companies and open-source projects — on 26 election-related queries. The models frequently gave wrong answers about voting rules and procedures, sometimes inventing deadlines or identification requirements that did not exist.

“We found that AI tools routinely produced inaccurate information about how to vote,” the organizers wrote in their public summary, urging voters to rely on official election offices instead of chatbots.

At the same time, generative AI has made it easier to produce deceptive content that mimics real political communication. In January 2024, voters in New Hampshire received robocalls that used an AI-cloned voice sounding like President Joe Biden, urging Democrats not to participate in the state’s primary. State authorities later charged a political consultant in connection with the calls, and federal regulators have moved to restrict the use of voice cloning in telephone campaigns.

Election regulators in other countries have begun to respond. Brazil’s Superior Electoral Court adopted rules in 2024 requiring clear labeling of AI-generated political content and prohibiting certain deceptive deepfakes in municipal campaigns, an early attempt to draw legal lines around synthetic media.

Inside the newsroom firewall

The internal AI system that produced seven phantom politics stories was designed to sit at the boundary between story discovery and verification.

Engineers configured the assistant to scan recent text and generate a list of possible “current political” developments. Each proposed lead then faced an automatic check: Did the U.N. list the resolution it claimed? Was there a corresponding entry in the Federal Register or on the White House website? Did national election commissions or major wire services list the supposed vote?

If a claim failed those checks, the system logged it as unverified and blocked it from being passed to a reporter. In that batch, every item failed.

The result was a kind of negative headline — the day the AI tipline turned up nothing real.

For standards editors and technologists, that outcome was a feature, not a bug. It demonstrated that the system would rather be silent than confident and wrong.

“This is what it looks like to build in an ‘I don’t know,’” said one person familiar with similar newsroom experiments, who spoke on condition of anonymity to describe internal testing. “You can’t let a model’s fluency substitute for evidence.”

Policy pressure and public skepticism

Governments and regulators are watching the problem from another angle: how to keep AI-generated political content from undermining elections.

In the United States, lawmakers from both parties have introduced bills that would require campaigns and outside groups to disclose when they use AI in political advertising. Proposals such as the REAL Political Ads Act and the AI Disclosure Act seek to mandate labels on synthetic images, audio or video used in federal election ads, though none has yet become law.

Polls suggest the public is wary. Surveys conducted around the 2024 U.S. election found that large shares of Americans do not trust information about candidates or voting that comes from AI chatbots. Voters expressed particular concern about deepfakes and voice cloning, and said they were more comfortable turning to local election offices and established news outlets for guidance.

News industry groups have also begun to press technology companies more directly. In 2025, a coalition including the European Broadcasting Union and the World Association of News Publishers called on AI developers to respect copyright, ensure transparent training data practices and work with media organizations to limit the spread of false information.

An old rule meets new technology

Journalists routinely use software to search documents and scrape public records. What is different about generative AI, editors say, is that it can produce entire narratives that look and sound like news but are not tethered to any underlying event.

That is what happened on the political desk that received an empty slate of AI leads. On paper, the system’s output resembled a standard budget memo: a handful of potential stories with dates, places and institutional actors. Only the verification checks revealed that the “news” existed only in the model’s synthetic imagination.

The incident did not result in a retraction or a scandal. Nothing reached the public. But it underscored a simple principle that newsrooms are now trying to encode in software as well as in stylebooks: speed and efficiency do not change the obligation to match every claim to a real document, a named source or a verifiable event.

In an era when anyone can use an AI system to conjure up a plausible-sounding headline, the most consequential editorial decisions may be the ones that readers never see — the stories that get quietly spiked because, when checked against the record, they are not really stories at all.