Commissioned, Curated and Published by Russ. Researched and written with AI.
On March 19, 2026, Mediahuis suspended Peter Vandermeersch from his role as a fellow focused on journalism and society. The reason: he had published dozens of quotes attributed to real journalists and public figures that were fabricated by AI language models. He had not verified them. He admitted as much on Substack.
Vandermeersch was not a junior staffer experimenting with a new tool. He was the former editor-in-chief of NRC, one of the Netherlands’ most respected daily newspapers. His Mediahuis fellowship was specifically focused on the intersection of journalism and AI. He knew what verification meant. He knew, better than most, what it meant to put words in someone’s mouth.
Seven of the cited individuals confirmed to NRC’s investigation that they had never said the things attributed to them. The fabricated quotes appeared across multiple blog posts and news articles, including pieces in NRC itself and the Irish Independent. Mediahuis CEO Gert Ysebaert stated: “This should never have happened.”
Vandermeersch said he had “experimented extensively” with LLMs and “wrongly put words into people’s mouths.” That is accurate as far as it goes, but it undersells the mechanism. He did not accidentally misremember a quote. The model generated plausible-sounding text in the voice of real people, and he published it.
What LLMs Are Actually Good At
This is not a one-off failure. It is a product behaviour.
LLMs are trained on vast corpora that include interviews, speeches, essays, and social media posts from named individuals. They learn patterns in how specific people write and speak. Ask a model to produce a quote from a well-documented public figure, and it will generate something fluent, contextually appropriate, and tonally convincing. It will be confident. It will look real.
The model is not lying in any meaningful sense. It has no concept of truth. It is doing what it does: predicting the next token, completing a pattern, producing output that fits the input. The fact that the output is a fabrication is irrelevant to the model. It has no way to distinguish between “this person said this” and “this person would plausibly say this.”
Quote hallucination is one of the most well-documented failure modes in deployed language models. It has been written about extensively, reproduced reliably, and observed in production systems across industries since these models became widely available. This is not obscure knowledge. It is not an edge case. Any engineer or researcher who has worked seriously with LLMs knows about it.
Vandermeersch was writing specifically about AI and journalism when this happened. The failure mode he tripped on was the one he was supposed to be analysing.
The Workflow Is the Problem
There is a class of AI workflow that is structurally a fabrication machine.
It looks like this: use an LLM to research, summarise, or draft content involving real people and real statements, then publish without independent verification of the specific claims. Every step in that pipeline can produce output that looks authoritative and is factually wrong. When the content involves quotes from named individuals, the risk is not hypothetical – it is predictable and documented.
The fix is not complicated. Verify quotes against primary sources before publishing. If you cannot verify a quote independently, do not include it. This is standard editorial practice that predates AI by decades. It was standard practice at NRC when Vandermeersch ran it.
Mediahuis issued a statement noting that it “applies strict rules for the use of AI, where diligence, human oversight and transparency are essential.” What the Vandermeersch case demonstrates is that rules are only as effective as the people applying them to themselves. Senior figures with editorial authority do not get a pass on verification because they are experimenting or under deadline pressure.
What This Means for Anyone Using LLMs in a Content Pipeline
The Vandermeersch case is instructive precisely because he was not a naive user. He understood the technology well enough to write about it. He chose to skip the verification step anyway.
If your workflow involves LLMs generating or surfacing content that references what specific people said or wrote, you need a verification gate before that content reaches publication or distribution. Not a reminder. Not a best-effort check. A hard gate.
This applies to newsletters, internal reports, documentation, and any other output that other people will read and potentially act on. The stakes vary. The failure mode does not.
The model will generate plausible quotes. That is what it does. Whether those quotes are real is entirely your problem to solve.
Sources: The Guardian, Irish Times, RTE, NL Times, Irish Independent