Berlingske has been publishing since 1749, which makes it older than the United States and considerably older than the idea that you should fact-check a robot before printing what it tells you. In May 2026 the Danish daily collided with that second idea at speed.

A report, three experts, and a problem

The article in question covered a Danish economic council's assessment of a proposed pesticide ban intended to protect clean drinking water. Reasonable subject, the kind of policy reporting a serious paper does all the time. To give the piece authority, it quoted three experts.

Two of them did not exist. One was a "technology professor Anne Kjær Nielsen" from the Technical University of Denmark, DTU, presented with a direct quote. There is no such professor. A second cited expert was likewise fabricated. The third was a real person who had been misquoted, saddled with words he did not say. So the article's expert scaffolding was, in plain terms, two ghosts and one misrepresentation.

How a summarizer ends up inventing people

The mechanism matters, because it is the kind of failure that feels impossible until you understand how these tools work. A journalist used Berlingske's internal AI system to summarize expert quotes from earlier material. But the journalist did not actually give the tool the source material to summarize. Asked to produce expert quotes about a report it had never been shown, the model did not respond with "I have nothing to work from." It did what language models do when handed a gap: it filled the gap with the most plausible-sounding content it could generate. Plausible-sounding expert quotes require plausible-sounding experts, so it produced those too, complete with a name, an institution, and a credential.

This is the part people consistently misunderstand about AI summarizers. They are not retrieving facts from a database. They are predicting text. Give a summarizer the actual document and it will mostly stay tethered to it. Ask it to summarize a document you never provided and you have not asked for a summary at all; you have asked for fiction in the register of a summary. It will oblige, fluently, and the output will look exactly like real reporting because looking like real reporting is the one thing it is genuinely good at.

Caught by the competition

Berlingske did not catch this itself. The errors were flagged by a reporter at Dagbladet Information, a rival Danish paper, who did the radical thing of questioning whether the cited sources were real and discovered they were not. That detail is its own small indictment. The fabricated professor made it through whatever passed for editorial review at Berlingske and was caught only when someone outside the building went looking. The adversarial system of journalism, competitors checking each other's work, ended up doing the quality assurance that should have happened before publication.

The response

Once the story broke, Berlingske moved through the now-standard sequence. The responsible journalist was suspended pending an external review meant to determine whether this was an isolated lapse or a symptom of something broader. A multi-paragraph section of the article was removed and the misattributed quotes were corrected. Editor-in-chief Tom Jensen called the episode deeply regrettable and was clear about the boundary the paper had crossed, saying the outlet does not allow AI to write its articles and that when AI is used as a research tool, the output must be verified because everyone knows it can make mistakes. Danish commentators were not especially charitable about it; an editorial in the rival paper Information called the affair an alarming reminder of how dangerous AI can be for public debate when fabricated experts slip into the record under a trusted masthead.

And then the part that makes this a 2026 story rather than a cautionary fable with a tidy ending: Berlingske confirmed it will keep using AI, presumably with refreshed guidelines. The lesson the paper drew was not "stop," but "we have to actually check," which is the correct lesson and also the lesson every newsroom claims to have already learned right up until the moment a fabricated professor appears in print under their masthead.

The recurring shape of newsroom slop

This is not an isolated Danish embarrassment. It is the same failure that has already produced suspensions and corrections at other European outlets, where a journalist leaned on a model to summarize or generate quotes and trusted the fluent result without confirming the humans in it were real. The common thread is never that AI "went rogue." It is that a person used a plausibility engine in a role that demanded accuracy, skipped the verification step, and let the output inherit the credibility of the publication.

The fix is unglamorous and entirely within human control. If you ask a model to summarize sources, give it the sources, and then confirm that every name, quote, and institution in the output corresponds to a real person who really said it. A quote is a claim that someone spoke specific words. A model cannot make that claim true; it can only make it look true, which is precisely the property that turns a time-saving tool into a libel risk wearing a press badge. Berlingske survived 277 years before learning this one in public. The internal AI tool needed about an afternoon.