A nice clean fake quote

The New York Times's April 14, 2026 piece on the Canadian election was framed as an analysis: how did Mark Carney's Liberals walk back from likely defeat and end up forming a majority government, and why was Pierre Poilievre's Conservative camp so sour about it. The Times's Canada bureau chief, Matina Stevis-Gridneff, wrote the story. It was the kind of piece the paper publishes after every foreign election where the result was unexpected, and it leaned heavily on direct quotation from political figures on both sides.

One of the quotes attributed to Poilievre read, in part, that turncoats with any "shred of integrity left" should resign their seats and run in by-elections. It was the kind of line that fits cleanly into a "sore loser" framing: short, punchy, easily lifted into a social-media graphic. The problem was that Poilievre never said it.

The original wording was eventually traced to an AI tool that had been used somewhere in the drafting workflow. The model produced a summary of Poilievre's reaction to the election, and that summary was rendered with quotation marks around it, which made it look like a direct quote from the man rather than a paraphrase by the model. The text moved from the drafting tool into the published piece without anyone matching it against an actual recording, transcript, or wire-service quote.

A reader caught it the next day

The error did not survive contact with the audience for long. On April 15, a Bluesky user replying to Stevis-Gridneff's post about the article flagged the quote as unfamiliar. Canadian political reporters have spent a lot of time listening to Poilievre, who is not exactly a stranger to memorable phrasing, and the cadence of the "quote" sounded off. Once readers started looking for it, the quote could not be located in any of his April speeches or interviews.

Bluesky is a small platform. The original flag was not a viral moment. It did, however, make the existence of the error a matter of public record almost immediately. From that point forward, every day the article stayed unchanged was a day the Times had been told there was a fabricated quote in its copy and had chosen not to fix it.

The correction did not run until May 1, more than two weeks after publication and the initial reader flag. The editor's note acknowledged that the wording attributed to Poilievre was an AI-generated summary of his views about Canadian politics that AI had rendered as a quotation. The article body was updated with a less colorful, real quote from a Poilievre speech given in April. Simon Willison's blog and The Walrus both reproduced the editor's note and tracked the timeline, with Karyn Pugliese, a journalism professor at Carleton, writing up the chain of events from a working-press perspective.

The new AI policy that does not apply to the actual error

The story stayed in the Canadian press cycle through early May, and at some point in the middle of that cycle, the Times's Standards and Trust desk circulated new guidance about AI use in newsroom drafting. La Voce di New York, picking up reporting on internal memos, summarized the upshot: the new restrictions on AI use apply to freelancers, not to staff reporters.

Stevis-Gridneff is a staff reporter. The error path that led to the Poilievre "quote" was inside the staff workflow. The new rule does not change anything about that path.

That choice is worth pausing on. The most charitable read is that staff reporters at the Times already have detailed internal guidelines and that the freelancer rule is filling a separate gap. The less charitable read is that an institution caught publishing a fabricated quote attributed to the leader of Canada's official opposition responded by tightening rules on the part of the workforce that was not responsible for the incident, while leaving the actual responsible workflow untouched.

Press Gazette's live tracker of AI-related journalism scandals had already logged a thick run of similar incidents at other outlets during 2025 and 2026. The pattern in most of those cases is similar: AI used somewhere in the drafting chain, fabricated material survives review, the outlet publishes, the public catches it, the outlet issues a correction and a policy update. The Times case is on the higher-end of that pattern because of the prestige and reach of the publication, and because of how long the corrected text stayed live.

How a "quote" becomes a quote

The technical mechanism here is worth describing plainly because it is going to keep happening unless newsrooms understand it. A generative model asked to summarize a politician's public reaction to an event will, depending on the prompt, produce text that looks like a paraphrase or text that looks like a direct quotation. The model has no internal flag for "this is paraphrase" versus "this is direct quotation." It is producing tokens that fit the request and the surrounding context. If the prompt or the surrounding draft contains other quoted material, the model will often produce more quoted material.

Once that output is in a draft, the visual cues that normally separate paraphrase from quotation - the quotation marks themselves - have already been generated. A human editor reading the draft to check for clarity sees what looks like a quote and treats it as one, because that is what a quote looks like. Unless there is a step in the workflow that specifically pairs every quotation with a source artifact (a recording timestamp, a transcript line, a wire-service tag), there is no friction in the editing process that would catch the synthesis.

At a newsroom with a strong copy desk, the friction usually exists informally. A copy editor pulls up the source, confirms the wording, and moves on. The institutional habit is what protects against fake quotes more than any specific policy. When AI tools are introduced into the drafting step without a corresponding strengthening of that habit, the result is exactly the Poilievre case: a "quote" with no source artifact attached, sitting in a published piece for over two weeks.

The graveyard lesson

The lesson is not that AI cannot be used responsibly in a newsroom. Plenty of outlets use AI for transcript cleanup, translation, and summary work without producing fabricated quotes attributed to named political figures. The lesson is that quotation is the place in journalism where the cost of a small workflow gap is highest. Quotes are the unit of attribution. Once a politician sees a direct quote attributed to them that they did not say, the institutional trust between them, the paper, and the readers takes a hit that the next polite correction note does not fully fix.

And once a paper's response to a fabricated quote story is a new AI policy that explicitly carves out the part of the workforce that produced the fabricated quote, the underlying point about newsroom verification is being deliberately misread. The freelancers were not the problem. Whoever was running an AI tool in the staff drafting chain, without anyone matching the quotation against a real source, was. The Times's correction does not pretend otherwise. The policy that followed it might as well have.