An excerpt from Steve Rosenbaum’s book The Future of Truth drew scrutiny when reporting identified more than half a dozen made‑up or misattributed quotes and subsequent scans flagged the text as largely machine‑generated. A Pangram analysis of the full manuscript returned roughly 53 percent AI‑generated content and an additional 9 percent likely AI‑assisted, figures that became central to the controversy because the book’s central subject is how AI reshapes perception and veracity. Rosenbaum has acknowledged including “a handful” of “improperly attributed or synthetic” quotations, and his bio notes a master’s degree in “truth” from New York University.
The roughly 1,450‑word excerpt was published by a magazine after prepublication fact‑checking and a recheck, yet editors still faced reader complaints that sections read “blatantly AI‑written.” Rosenbaum’s acknowledgments list ChatGPT, Claude, NaturalReaders, ProWritingAid, and Grammarly as tools that helped “refine and polish the presentation of [his] ideas.” The magazine’s generative‑AI editorial policy bars publishing AI‑generated or AI‑edited writing, a restriction that intensified scrutiny of how those tools were applied during the book’s preparation.
After a reader complaint, the excerpt was analyzed by multiple AI‑detection services — Pangram, GPTZero, and ZeroGPT — each returning results consistent with likely AI generation. A reporter then scanned the entire manuscript with Pangram, described in the coverage as a current gold standard among detection services; that run produced the 53 percent AI‑generated and 9 percent likely AI‑assisted breakdown. Detection tools and their methodologies remain contested and imperfect, but the Pangram figure provided a concrete metric that fueled questions about authorship and editorial oversight.
Rosenbaum told reporters he used generative models at several stages — for source discovery, brainstorming, structural feedback, and language refinement — and insisted that “the ideas, reporting, arguments, and final authorship are mine,” adding that the excerpt was not “generated by AI and then simply published as‑is.” He declined to dispute the Pangram results and said he would not engage that specific line of questioning, calling it an accusation without a meaningful response. Distributor Simon & Schuster declined to comment.
The episode underscores practical tensions facing authors, editors, and toolmakers: writers frequently treat models as research or drafting aids, but those workflows can produce synthetic phrasing or misattributed quotations that slip past editorial safeguards. For developers, the case highlights demand for clearer provenance features, durable edit histories, and interoperable metadata that would help downstream readers and editors distinguish human‑origin content from model‑generated text or edits.
Rosenbaum demonstrated the kind of AI assistance in question by asking ChatGPT to describe the reporter and then reading the model’s output aloud, saying the model surfaced accurate references to prior reporting. That example points to legitimate value when models act as augmented search or brainstorming tools, yet the presence of synthetic or improperly attributed quotes in a book about truth erodes credibility and is likely to push publishers toward stricter disclosure, more rigorous provenance tracking, and updated editorial workflows for AI‑assisted authorship.
Sources
Replies (0)
No replies in this topic yet.
