When the Machine Writes the Quote: What 2026's AI Corrections Reveal

Five months into 2026, a new category has entered the corrections columns. It is not the misspelt name or the transposed figure. It is the quotation that nobody said.

In February Ars Technica retracted a story within two hours of publication after it emerged that quotes attributed to the developer Scott Shambaugh had been generated by an AI tool. In May The New York Times appended an editor's note to an analysis of Canadian politics after a "quote" from Pierre Poilievre turned out to be an AI-generated summary of his views, formatted as direct speech. Lower down the media food chain, outlets have begun correcting AI-written summary bullet points that misstated the stories they sat above.

The tool did its job

It is tempting to blame the technology. That would be a mistake. A language model asked to summarise a politician's position will produce fluent, plausible prose, and if it is asked for something quotable it will produce something quotable. It has no concept of a recording, a transcript or a press release. It does not know the difference between what someone said and what someone might have said. That is not a bug that will be patched; it is how these systems work.

The failure in each case was editorial. Quotation marks are the most explicit promise journalism makes: these words came from this person. In both the Ars Technica and Times cases, text produced by a tool crossed into a draft and acquired that promise without anyone checking it against a source. The Times's own note made the point bluntly: the reporter should have checked what the tool returned.

Policies aimed at the wrong target

Several outlets have responded by restricting AI use by freelancers or by publishing broad principles about human oversight. Principles are fine, but the Times error came from a staff correspondent working under guidance that already required AI use to begin with vetted information. The rule existed. The check did not.

What would actually work is narrower and duller. Any text that originates with an AI tool should be treated as unsourced copy until a human has matched every factual claim to a primary source. Quotations should be pasted from transcripts or recordings, never retyped from a summary. And corrections for AI-originated errors should say so, as both Ars Technica and the Times to their credit did, because readers are entitled to know how the error arose.

Why this is different

Ordinary errors are distributed randomly. AI fabrications are not. They cluster around what a person is expected to say, which makes them persuasive and makes them harder to spot. A fabricated quote that sounds wrong is caught by the first sub-editor. One that sounds exactly right can run for weeks.

That is the real lesson of this year's corrections. The danger is not that machines will make journalism obviously worse. It is that they will make it subtly, plausibly wrong, in ways that confirm what readers and editors already believe.