New York Times admits AI-generated summary was presented as a Poilievre quote

The New York Times has acknowledged that a quotation attributed to Pierre Poilievre, the Canadian Conservative leader, was in fact a summary of his views produced by an artificial intelligence tool and formatted as a direct quote.

The error appeared in an analysis of Canadian politics published on 14 April by Matina Stevis-Gridneff, the paper's Canada bureau chief. The article presented Mr Poilievre as saying, in connection with a March speech, that defecting MPs with any shred of integrity left should resign their seats and contest by-elections. He had not said it.

The correction

A reader on Bluesky flagged the problem on 15 April. The Times appended an editor's note on 1 May, more than two weeks later, explaining that the wording was an AI-generated summary that the tool had rendered as a quotation, and stating that the reporter should have checked the accuracy of what the AI tool returned. The article now carries a genuine quote from a speech Mr Poilievre gave in April.

Why it matters

The substance of the invented quote was not wildly out of line with Mr Poilievre's known position, which is part of what makes the case instructive. Plausibility is precisely what allowed it through. A fabricated quotation that sounds like its subject is more dangerous than one that does not, because nobody thinks to check it.

The delay matters too. The error was public within a day of publication. A two-week gap between a reader's flag and an editor's note, on a quotation attributed to the leader of a major party during a period of intense political manoeuvring, is long for a paper with the Times's resources and corrections apparatus.

A policy aimed elsewhere

Reports in May indicated the Times had sent freelancers a memo requiring their work to be the product of human creativity and craft and barring the use of chatbots. The Poilievre error, however, came from a staff correspondent working inside the paper's own workflow. The paper's existing guidance already says generative AI use must begin with information vetted by its journalists. The gap here was not the absence of a rule but the absence of a check.

The fix is not complicated. Any text that enters a draft from an AI tool should be treated as unsourced until a human has matched it to a recording, a transcript or a published statement. Quotation marks should be the last thing added, not something a tool supplies.