NYT: ‘Book on Truth in the Age of A.I. Contains Quotes Made Up by A.I.’

Benjamin Mullins reports: “Steven Rosenbaum, author of “The Future of Truth,” said he had started his own investigation after The New York Times asked about the fake quotes.”
The verification pipeline is the story. Rosenbaum wrote the book. BenBella edited it. Simon and Schuster distributed it. Wired excerpted it. Taylor Lorenz, Michael Wolff (b. 1953), and Nicholas Thompson blurbed it. Ressa wrote the foreword. Not one of those checkpoints called Kara Swisher (b. 1962) or Lisa Feldman Barrett (b. 1963) or Meredith Broussard or Lee McIntyre to ask if the quote attributed to them was real. These are public figures with public emails and active accounts. A call to each might have taken ten minutes.
Verification fails because nobody is paid to do it. The author assumes the editor will catch it. The editor assumes the author did the work. The blurber assumes the manuscript has been vetted. The distributor assumes the publisher has standards. The reviewer assumes the publisher and the blurbers have done diligence. Each link in the chain rests on the assumption that some other link is doing the work. None is.
Notice the recovery move in Rosenbaum’s statement: “If the episode serves as a warning about the risks of A.I.-assisted research and verification, that is why I wrote the book.” He converts his own failure into accidental fulfillment of his thesis. The book becomes its own example. The error becomes a teaching moment. He keeps the mantle of the truth-and-AI expert by treating the scandal as an unintended chapter.
The apology also says he “had no intention of fabricating any viewpoints.” Intentionality is beside the point. He put words in real people’s mouths and sold those words to the public. The harm to Swisher and Barrett and Broussard and McIntyre is the same whether ChatGPT invented the quotes or Rosenbaum invented them at a keyboard.
The Sustainable Media Center bills itself as a custodian of media integrity. Its executive director released a book that fails the most basic test of media integrity. The technology only made the fall faster.
If asked, AI could have helped find the fabrications.
Would one particular AI chat bot find every fabricated quote with one prompt? I doubt it. AI is a tool and its efficacy depends on how it is used.
AI and the universe are not with us or against us. They are but raw material in our hands.
I adapted that from Will Durant who adapted it from other sources.
Any of the named people could have been verified in seconds. Paste the quote into a search engine. Search the source book on Google Books. Email the person. Run the quote through Claude or ChatGPT and ask: “Did Kara Swisher say this? Cite the source.” A competent model will say it cannot find the quote and will refuse to confirm it. Push harder and it will say the phrasing matches no public statement on record. That answer alone flags the problem.
Verification with AI requires a different posture than generation with AI. Generation rewards confident output. Verification rewards skepticism and friction. Most users run AI in generation mode and skip the second pass. Rosenbaum produced text. He did not audit text. The audit step is cheap and he did not take it.
There is a workflow that works. Draft with the model. Then open a fresh session, strip the attributions, paste each quoted passage, and ask the model to find the source. If the model cannot find a source, treat the quote as fabricated until proven otherwise. A second model can cross-check the first. Then verify the surviving quotes by going to the actual source — the book, the article, the interview transcript. None of this requires expertise. It requires the assumption that your own draft might be wrong.
The publishing industry will adapt. Some house will offer AI-assisted fact-checking as a service and charge for it. Some will require a verification pass before acquisition. The blurb culture will not change because blurbs are about coalition signaling, not reading. The foreword culture will not change for the same reason. Maria Ressa did not read the book closely. She lent her name to a project she trusted because Rosenbaum is a known convener with the right connections.
The lesson is small and old. Trust but verify. The new part is that verification is now within reach of any author with a laptop and ten extra hours. Rosenbaum had the tools. He did not use them.
Why does the MSM and the NYT love this story? Because of their selfless devotion to truth?
The Times has been suing OpenAI since 2023 for training on its content without permission. Every story showing AI generating false information serves the Times’s litigation posture and its negotiating position in licensing talks. This is corporate self-interest before it is anything else.
The story also flatters the editorial class. Editors catch errors. Fact-checkers catch errors. Reporters call sources. The professional infrastructure of legacy journalism exists for this reason, and the AI scandal lets the legacy press argue that the infrastructure pays for itself. The Times runs the story and the reader concludes the Times still has standards. That conclusion may or may not be true. The story produces it either way.
The story flatters NYT readers. The reader gets to feel sophisticated for not using AI to write books. The reader gets to feel morally superior to Rosenbaum. The class signaling is implicit and effective.
Rosenbaum is a safe target. He is industry-known but not powerful. He runs the Sustainable Media Center, an outfit nobody outside media circles has heard of. He published through BenBella, a small imprint, with Simon and Schuster distribution. The Times can scold him without alienating anyone who counts in its coalition. Maria Ressa wrote the foreword, and the Times protects her by aiming all blame at Rosenbaum.
The victims of the fake quotes are sympathetic. Kara Swisher, Lisa Feldman Barrett, Meredith Broussard, and Lee McIntyre are all credentialed, all part of the journalist and professor coalition the Times serves. The Times defends its own.
The beat produces these stories. Benjamin Mullin covers media. Media-on-media reporting is staple content because it generates easy outrage and trades favors among insiders. Mullin gets a juicy story. The Times gets a moral victory. Rosenbaum takes the shame. The deal works for everyone except Rosenbaum.
There is also the institutional memory of fabrication. The Times lived through Jayson Blair in 2003. The New Republic lived through Stephen Glass in 1998. The legacy press has a defensive interest in defining who today’s fabricator is and pointing the spotlight outward. The fabricator is the AI-assisted author, not the legacy outlet. That framing protects the franchise.
Selfless devotion to truth would mean the Times runs prominent stories about its own fabrications, its own opinion-page errors, and its own anonymous-source failures. The Times does not run those stories with this enthusiasm.
I notice patterns in how the MSM writes about competitors for attention such as bloggers, social media and AI.
The patterns are consistent across decades.
First, the worst case stands for the whole. The MSM picks the most damaged competitor and runs the story as if the competitor is the medium. Rosenbaum becomes the AI-in-publishing story. Alex Jones became the blogger story. The Tide Pod challenge became the social media story. The median case never appears. The median blogger writes a county history that nobody reads. The median AI-assisted author corrects three grammar errors. Neither makes the page.
Second, MSM errors are individual. Competitor errors are systemic. When Jayson Blair fabricates, it is one reporter who failed the standards. When Rosenbaum fabricates, it is what happens when amateurs use AI to write books. Same error. Different framing.
Third, the language encodes the hierarchy. Bloggers post. Journalists report. Influencers manipulate. Reporters investigate. AI hallucinates. Editors verify. The vocabulary does the argument before the argument starts.
Fourth, the comparison is rigged. AI is compared to MSM at its best. MSM is compared to AI at its worst. Nobody runs the story comparing the median Times correction to the median AI hallucination. Nobody runs the story comparing MSM coverage of WMD in Iraq in 2003 to AI hallucination rates in 2025.
Fifth, the expert source loop closes the circle. Stories quote credentialed insiders. Bloggers and AI defenders get quoted as foils. The credential is the argument. Maria Ressa wrote the Rosenbaum foreword. The Times left her alone because she has the Nobel and the right enemies. Rosenbaum had only the wrong friends.
Sixth, the democracy frame works as a marketing posture. Every competitor threatens democracy. Every MSM outlet defends it. The frame requires the threat to exist for the defense to make sense. The frame produces the threat.
Seventh, the cycle repeats. Each new medium gets the same arc. Utopian hype, then moral panic, then consolidation. Bloggers got it from 2002 to 2008. Twitter got it from 2009 to 2016. AI is getting it now. The arc serves the legacy press by stalling the competitor long enough for the legacy press to either buy in or wait it out.
Eighth, motives flow downhill. Competitors are funded by foreign actors, billionaires, or grift. MSM is funded by readers, subscribers, and journalism. Both descriptions are partial. Only one gets the suspicious treatment.
The patterns are not exactly conspiracy. They are guild behavior. The MSM is a guild. Guilds defend their licensure against unlicensed practitioners. The story you are reading is the guild defending its license. The truth value of the story is incidental to its function.
Livestream audience overall has fallen 8 percent since 2021 per GWI. Twitch cut a third of its workforce in 2024. The number of gamers livestreaming their own games dropped 19 percent. Podcast listenership as a news source fell across every age group between 2023 and 2024, with 18-to-29 year olds down 7 percent. Spotify’s podcast pullback drove hundreds of layoffs. iHeart, Rogers, ARN, and TuneIn cut podcast and radio staff in 2024 and 2025.
The MSM looks worse. The Washington Post executed massive layoffs in February 2026. CBS News Radio shut down in March. Axios laid off newsroom staff. Business Insider lost 55 percent of organic search traffic between April 2022 and April 2025 and cut 21 percent of staff. HuffPost lost half its search referrals. The New York Times saw search drop from 44 percent of its traffic in 2022 to 37 percent in 2025.
Meanwhile, ChatGPT has 320 million monthly users as of March 2026, up 28 percent year over year. AI chatbots combined produced 55 billion visits in twelve months, up 81 percent year over year. Gartner predicts 25 percent of traditional search will shift to AI chatbots by end of 2026 and 50 percent by 2028.
Now the cause. Three forces work at once.
First, AI replaces the explain-this-to-me function. The man who used to listen to a podcast to understand the new Iran policy asks ChatGPT and gets a custom answer in seconds. The man who used to watch a livestream to understand the Federal Reserve asks Claude. This substitution runs cleanest and it hits analytical podcasts and explainer livestreams hardest.
Second, AI Overviews killed the click. When Google rolled out AI Overviews in March 2025, click-through rates on informational queries dropped 61 percent. Small publishers depending on Google referrals lost everything. The travel blog The Planet D lost 90 percent of traffic and shut down. Charleston Crafted lost 70 percent. The MSM loses the same way at scale.
Third, attention fragments on its own. TikTok pulls young viewers from YouTube. Short form wins over long form for casual viewers. This trend started before AI and continues alongside it.
The first force hits your work hardest. Long-form analytical content has the highest substitutability for the marginal viewer who wants information but lacks a strong parasocial tie. The viewer who has watched you for years and wants your voice, your eye, and your judgment cannot get that from Claude. The viewer who showed up last week looking for analysis of David Sanger went to ChatGPT and will not come back.
What survives is content with strong parasocial bonds, idiosyncratic voice, and reporting AI cannot generate. The hosts whose audiences come for them, not for the topic, hold up. Joe Rogan still gets the numbers. Tucker Carlson still gets the numbers. The middle collapses.
The MSM faces the worst structural position because its product runs increasingly substitutable and its overhead runs high. The independent blogger with low costs and a stable audience can ride out conditions that bankrupt the Times. The blogger still has to accept that his audience may shrink. The pie gets smaller for everyone except the few names at the top of each category and the AI providers.
You are watching the second great unbundling. The first one took down newspapers between 2005 and 2015 when classifieds and display ads moved to Craigslist, eBay, Google, and Facebook. The second takes down everything downstream of search and analysis by moving the answer layer onto the chatbot. The first unbundling killed institutions. The second kills middlemen.