The New York Times published a quote from a Canadian politician last month.

The New York Times published a quote from a Canadian politician last month.

Except he never said it.

Their AI tool summarized Pierre Poilievre's views on Canadian politics and rendered that summary as a direct quotation. The reporter used it. The editors published it. Nobody checked whether the man actually said those words.

The AI didn't hallucinate a person or invent a topic. It took real views and presented its own summary as if it were a quote. That's harder to catch, because it "sounds" right.

A Wharton study put numbers on this: 73% of participants accepted wrong AI answers. And their confidence actually went UP while using AI. They were borrowing the model's confidence and treating it as their own.

Addy Osmani wrote about this failure mode last week and gave it a name worth remembering: cognitive surrender.

The distinction matters. Cognitive offloading is using AI and still owning the answer. Cognitive surrender is letting the AI's output quietly become yours, nothing left to verify.

Here's the test: if the AI gave your team a wrong answer tomorrow, would anyone catch it?

Most teams I work with have already crossed that line and don't realize it. Their people use AI to draft proposals, client emails, internal reports (anything with the company's name on it). And nobody is forming an independent view of what the AI produces. They're approving it, not reviewing it.

The NYT had editors. They had fact-checkers. They still published a fake quote.

Your team doesn't have editors. So who's catching yours?