The AI Book With Synthetic Quotes Shows Citation Is the Product Boundary

Kara Swisher told the New York Times she never said what a book attributed to her. The quote — a crack about having "a stick up my butt" — was not hers. It was ChatGPT's.

Steven Rosenbaum wrote The Future of Truth, a book about how artificial intelligence erodes reality. He used ChatGPT and Claude during research, writing, and editing. He found the AI answered correctly about eight times out of ten, he told Ars Technica — and confabulated the rest. The book's publisher ran it past a fact-checker and two copy editors. All three missed the fabrications.

Of the 285 outside citations Rosenbaum included, the Times identified six as problematic, including three entirely synthetic quotes with no apparent source. Lisa Feldman Barrett's name appeared twice in passages that do not appear in her actual book, and which she says are also wrong. The New York Times contacted Rosenbaum on May 17 and he confirmed the hallucinations the following evening. The book was published on May 12, 2026.

The failure is not the scandal. The scandal is the gap between what the technology promises and what three separate human gatekeepers failed to catch. Rosenbaum is not a naive user — he is a veteran documentarian and author who understood the risks and used the tools anyway. The book is not a self-published pamphlet. It came from Matt Holt Books, an imprint of Simon & Schuster. Three checkpoints, one failure mode.

What makes this a structural story rather than a single-author cautionary tale is the timeline. Three days after the Times expose, the preprint server ArXiv announced a one-year ban for researchers who submit papers with LLM-generated content that hasn't been checked. The policy is an implicit admission: the repository's existing editorial process assumed human authorship. That assumption no longer holds.

The mechanism that made this possible is cheap to replicate. Large language models generate fluent text at near-zero marginal cost. They do not retrieve facts — they generate plausible sequences of tokens that trained on facts, and they do so with the confident register of a reference librarian. The hallucination rate Rosenbaum describes — roughly two in ten queries producing confident errors — is not a bugfix target in most commercial deployments. It is the output. Gatekeepers built for a world where fabrication required human effort are now auditing text where the marginal cost of fabrication approaches zero.

Citation is the accountability mechanism that survived. In a world where AI-generated text is abundant, the scarce thing is a traceable link back to a human source who actually said or wrote what the text claims. Verification infrastructure — the process of confirming that a claim traces to a person, a document, a datum — is what the Rosenbaum failure exposed as valuable. Citation used to be an editorial footnote. It is becoming the only accountability mechanism that survives when the text itself cannot be trusted.

The counterargument is the one that should accompany every story about a single instance of a systemic risk: this might be an outlier. Most AI-assisted books probably do not contain high-profile fabrications from named journalists. We see the ones that get caught — the Kara Swishers, the Lisa Feldman Barretts — because they have the standing to notice and object. The fabrications that landed in uncontroversial text, the invented citations in chapters nobody reviewed, the misattributed quotes that matched what a field expected to hear: those are invisible.

Whether this is an isolated failure or a predictable symptom of a publishing industry that has not yet rebuilt its editorial stack for AI-assisted authorship is the open question. The book is still being sold. Matt Holt Books has not recalled it. Rosenbaum's position is that the tool is still useful and the error rate is manageable. The New York Times is not persuaded. What to watch next is whether other named sources in the book come forward, and whether the publisher's post-publication review finds more than six problematic citations.

What to watch: ArXiv's ban rolls out and other preprint servers signal whether they will follow. If they do not, the gap between academic norms and commercial AI usage grows wider. The citation infrastructure that made academic publishing accountable — peer review, reference checking, editorial oversight — was not designed for a world where the author is a language model and the human is a manager of outputs.