Steve Rosenbaum's AI-Truth Book Had AI-Hallucinated Quotes

Steve Rosenbaum published a book in May 2026 arguing that AI erodes our collective grip on truth. The book itself, it turned out, contained AI-hallucinated quotations attributed to real people who never said them. Wired, which had run a 1,450-word excerpt, retracted it after an investigation.

The New York Times reported the book contained over a half-dozen made-up or misattributed quotes. Among them: a quotation falsely attributed to tech journalist Kara Swisher describing AI language models as 'like a mirror' that 'reflects our own morality back at us, polished and articulate, but ultimately empty behind the surface.' Swisher told the Times she never said that. Researchers Lisa Feldman Barrett, Meredith Broussard, and Lee McIntyre were also named as sources of fabricated or distorted quotations.

Wired ran the excerpt and the full book through multiple AI detection services, including Pangram, GPTZero, and ZeroGPT. Each flagged the material as AI-generated or likely AI-generated. Pangram's analysis placed 53 percent of the full book as AI-generated and 9 percent as likely AI-assisted. In the resulting interview, Rosenbaum acknowledged using ChatGPT, Claude, NaturalReaders, ProWritingAid, and Grammarly during research and editorial development. He said 'false positives can occur' on the detection results but declined to engage further. He also criticized Wired's generative AI policy as restrictive and suggested the publication's writers probably use AI in secret.

The honest caveat is that AI detection tools are not definitive instruments, and Rosenbaum's point about false positives is fair in the abstract. What the reporting does establish is that he used multiple AI tools throughout the process, that multiple quotes in the final text were fabricated or misattributed, and that he acknowledged both. Wired retracted the excerpt Friday afternoon. Rosenbaum argued to the Times that if the episode 'serves as a warning about the risks of A.I.-assisted research and verification, that is why I wrote the book.' As Techdirt noted, one critic called that defense circular logic.

The cleaner takeaway for publishers and editors is structural. The gap between an author's theoretical understanding of AI risk and their actual verification workflow is exactly where unverified hallucinations get through. The book reportedly contains 380 citations in total; how many others have been checked is an open question nobody has yet answered.