Steven Rosenbaum says his book 'The Future of Truth: How AI Reshapes Reality' contained AI-hallucinated quotes, even after five editorial rounds and more than 4,000 tracked changes.
Rosenbaum said the book had been out five days when Ben Mullin, a media reporter at The New York Times, asked about one quote attributed to an anonymous DARPA researcher. He said the inquiry led to a front-page Times story headlined: 'The Future of Truth' Contains Quotes Made Up by A.I.
The author said the fallout was immediate. Events were canceled, a recorded podcast was spiked, and the book became a punchline before he could answer what happened. He said he had used AI to organize research, summarize documents, and refine language, and that this use was disclosed in the acknowledgments.
According to Rosenbaum, the book went through five editorial rounds and more than 4,000 tracked changes. He said that early AI errors were easier to catch because they were obvious, including false URLs, wrong DOI numbers, and quotes with no source. But he said cleaner errors survived because AI could replace facts with believable fictions.
Rosenbaum said a quote attributed to Kara Swisher never came from her. He said the new edition quotes her correctly and that he texted with her to confirm it. He added that 'the errors survived' because some of them 'read cleanly.'
He also said the book included more than 400 citations and that AI did not write the book. Still, he said the responsibility for the errors was his, writing that he should have caught them and that he owns it.
The larger lesson is bigger than one book. When AI can produce plausible falsehoods at scale, the cost is paid by readers, authors, and anyone who still thinks human review alone is enough. Capital is moving fast into tools that promise convenience, but truth is not a feature that can be outsourced without consequences.
The market has already voted for speed, scale, and automation. What this episode shows is that those incentives can collide with basic standards of verification. If the product is language, then the liability is trust. And once trust is damaged, no amount of productivity gains can fully buy it back.
