The New York Times published an editors’ note on May 10, 2026 admitting that a quote attributed to Canadian Conservative leader Pierre Poilievre was never actually said by him. It was an AI-generated summary of his views, rendered by the tool as if it were a direct quotation, and the reporter ran it without checking. Simon Willison flagged the note in his blog, and it’s a small paragraph with very large implications.
Here’s what happened, in the Times’ own words. A reporter used an AI tool to pull together Poilievre’s positions on Canadian politics. The tool returned a passage formatted as a quote, including the word “turncoats” to describe politicians who switch parties. That phrasing landed in the published article as if Poilievre had said it. He didn’t. The Times has now replaced the fabricated line with actual quotes from a speech he gave in April.
Why This One Matters
This isn’t a quirky AI error story. It’s a newsroom of record getting caught publishing words a politician never spoke, because an AI assistant invented them and a reporter trusted the output. Willison has been documenting this exact failure mode for years. Large language models don’t distinguish between summarizing and quoting. Ask one to describe what someone thinks, and it will happily wrap the answer in quotation marks that look authoritative and are completely made up.
What stands out here is the chain of trust that broke. The reporter trusted the model. An editor trusted the reporter. A copy desk trusted the editor. The quote made it through every layer because it looked like a quote. That’s the whole problem with hallucinated quotations. They don’t announce themselves.
The Bigger Pattern
This is the third or fourth high-profile case in the past year of AI-fabricated quotes reaching publication. Earlier incidents hit smaller outlets and got brushed off as growing pains. The Times correction changes the conversation. If the paper of record can ship a hallucinated political quote, the assumption that “professional newsrooms catch this stuff” no longer holds.
A few things worth noting about how this kind of failure happens:
- AI tools format summaries in quotation marks when prompted loosely. Asking “what does X say about Y” can return fabricated direct speech.
- Reporters under deadline pressure use these tools as research shortcuts. The output looks polished enough to skip verification.
- Fact-checking workflows weren’t built around AI-generated content. There’s no standard step that says “verify any quote that came from a model.”
- The fabricated quote often sounds plausible, because the model is trained to produce plausible-sounding speech in the subject’s voice.
What Practitioners Should Take From This
If you’re using AI tools in any workflow that produces public-facing claims, you need a hard rule: no quotation marks from the model ever go out without a primary source check. Treat every quote a model produces as suspect by default. The Times correction is a free lesson in what happens when that rule isn’t enforced.
For newsroom leaders, the question now is policy. Most outlets have vague “use AI responsibly” guidance. That clearly isn’t enough. The next round of editorial standards needs explicit language about verifying any model-generated text against a primary source before publication, and probably a logging requirement so corrections can trace which tool produced what.
For everyone else, this is a useful reminder that AI-generated text is fluent, confident, and sometimes wrong in ways that are hard to spot. The fluency is the trap. A messy, obviously-machine summary triggers skepticism. A clean, well-formatted quotation doesn’t.
Expect more of these corrections to surface as outlets audit their use of AI tools. The Times set a marker by publishing the note plainly. Other publications will have to decide whether to be that transparent when their own AI-fabricated quotes get caught.
Full editors’ note and Willison’s commentary at the original source.