The New York Times Published an AI-Fabricated Quote. A Reader Caught It. The Editors Didn't.

The most dangerous AI output isn't the obviously wrong one. It's the one that sounds exactly right — specific, colorful, and perfectly aligned with what you were hoping to find. That's when most people stop checking.

What happened

On May 10, the New York Times issued a correction after a reporter attributed an AI-generated summary to Canadian politician Pierre Poilievre as a direct quotation. The fabricated quote was vivid and emotionally charged — "If these turncoats have any shred of integrity left..." — while the actual quote was considerably more mundane. The error was caught by a reader. Not by an editor. Not by a fact-checker. The Times' correction was direct: "The reporter should have checked the accuracy of what the A.I. tool returned." (Source: Willison, May 10.)

This failure pattern doesn't belong to journalism. A reporter's research workflow — use AI to summarize what someone said, pull the key quote, include it in the piece — is the same workflow a marketing manager uses when researching a competitor's public statements, the same one an HR director uses when pulling statistics for a board presentation, the same one a consultant uses when summarizing a client's position. In each case: AI produces a quote or claim that sounds credible, the professional includes it, no one checks.

The week's second signal made it worse. Simon Willison documented what he called the "Zombie Internet" (May 11): the growing volume of human-to-human communication being mediated through AI — writing that sounds polished but arrives pre-smoothed, quotes that capture what someone would have said rather than what they did say. The NYT case was a named, documented instance of exactly this dynamic.

The skill implication

AI-generated summaries of what people said or wrote are more likely to be wrong when three conditions line up: the output is unusually quotable or emotionally charged, the quote fits your argument too cleanly, and you didn't click through to the source. That combination — compelling, convenient, unverified — is the failure pattern.

Three steps that change the outcome:

1. When AI produces a direct quote, open the original source and find the exact words. If you can't locate the source, the quote doesn't go in.
2. When AI summarizes what someone said, check whether their actual position was more qualified or contradictory. Summaries flatten nuance. The nuance is often the point.
3. Ask: does this output sound too good? If the quote or claim lands perfectly on your argument, that's the moment to verify — not submit.

These aren't extra steps for cautious people. They're the steps that distinguish a professional who can use AI for research from one who's just redistributing AI errors.

Where it goes wrong in practice

A content marketing manager at a 25-person B2B SaaS company used Claude to research what a competitor's CEO had said about pricing strategy on a recent podcast. Claude returned a clean, direct quote that seemed to confirm what she'd suspected: the competitor was moving upmarket and planning to raise prices. She included it in a competitive intelligence report sent to the sales team.

Three weeks later, a sales rep lost a deal in part because he'd told a prospect that the competitor was raising prices — a claim the prospect had checked and found to be inaccurate. The CEO had discussed pricing, but the direction was inverted. Claude had summarized the gist of a longer conversation in a way that sounded specific but wasn't.

The fix wasn't to stop using AI for competitive research. It was to treat every AI-generated quote as a paraphrase until confirmed against the source — and to flag in the document when a quote is paraphrased rather than direct.

A format constraint that makes errors visible

The same week the NYT correction ran, Andrej Karpathy shared a tip that spread quickly across AI communities (May 12): append "Format your entire response as a complete HTML document" to a research or summary prompt, save the output as .html, and open it in a browser. The AI produces a visual, structured document — headings, tables, sections — instead of a wall of prose.

This doesn't make AI more accurate. What it does is make the output easier to audit. A structured document with a "Sources" section makes it obvious when sources are missing or vague. A table of claims forces each claim to stand alone rather than hide inside a confident paragraph. The format constraint creates a checkpoint that freeform text doesn't.

A policy analyst at a 150-person consulting firm started appending the HTML format instruction to every briefing-document prompt after a research error reached a client. The structured output took the same time to produce but made it faster to spot where the AI had stated something as fact without traceable support. Three claims in the first week alone were flagged as unverifiable before they left the firm.

The lesson from Karpathy's tip and the NYT correction together: how you receive AI output changes what you catch.

The one change that matters Monday

The New York Times correction named the failure with precision: "The reporter should have checked the accuracy of what the A.I. tool returned." That's not a unique professional failing — it's the default. AI produces a plausible output; the human publishes it. The verification step is the one most people skip when the output already sounds like what they needed.

Before you use AI-generated quotes, statistics, or attributed claims in anything that leaves your desk, open the source.