In November 2024, journalist Jack Shafer and others noticed something odd about Steven Rosenbaum and Robin Raskin's book The Future of Truth. It contained quotes and citations that appear to have been fabricated by AI. Publisher Mango pulled the book. Rosenbaum, a veteran documentarian, acknowledged AI was used in the drafting process. If you write with AI, including inside tools like YouWrite's Refine, the failure mode this case exposes is one every writer is one lazy afternoon away from repeating.
What actually happened, at the sentence level
Coverage from Vanity Fair (Charlotte Klein, November 21, 2024) and follow-ups elsewhere flagged specific problems: quotes attributed to real people who said no such thing, references to work that either didn't exist or didn't say what the book claimed. Textbook large-language-model behavior. When an LLM is asked to support an argument with a quote from, say, a well-known technologist, it does not search a database of things that person actually said. It generates a plausible sentence in that person's rhetorical style and attaches their name.
The output looks correct. It reads correct. It is often almost correct, which is worse than being obviously wrong, because near-truth passes the eye test.
Why hallucinations cluster around proper nouns
Hallucination rates rise sharply in three zones:
- Direct quotations attributed to a named person.
- Citations with a title, author, year, and publisher.
- Statistics with a specific number and source.
These are exactly the elements that give nonfiction its authority. They are also the elements a model is most tempted to invent, because the training objective rewards fluent, specific-sounding output. A vague paraphrase is safer for the model to get right and less impressive on the page. A crisp fake quote is riskier and more persuasive. Guess which one shows up more.
This is not just a professional author problem
Read the Rosenbaum story as a cautionary tale for name-brand nonfiction writers with book deals and you miss the larger risk.
Consider who is actually using AI to help write long-form text right now:
- A retiree drafting a family memoir who asks a chatbot to "add a paragraph about what life was like in Youngstown steel mills in 1962."
- A worldbuilder assembling a game lore bible who asks for "a plausible medieval siege tactic from the 1300s" to lend realism.
- A founder writing a personal essay who asks the model to "find a good statistic about first-time entrepreneurs."
- A parent writing a eulogy who asks for "a fitting quote from Maya Angelou about grief."
Every one of those prompts will get an answer. Some answers will be correct. Some will be confidently wrong. The memoir reader has no way to tell which is which, and neither does the writer, unless they check.
A fabricated Angelou quote in a eulogy that gets shared online is its own small pollution of the record. A false detail in a family memoir becomes, for the next generation, family history.
The real mistake was not using AI
Rosenbaum's error, based on public reporting, wasn't that he used an LLM. He treated the output as a draft to be lightly edited rather than as a prompt to be independently verified. Process failure, not technology failure, and it is the exact failure most first-time AI users make.
AI output is not a rough draft. It is a research lead. A rough draft, in the old sense, was something you wrote from sources you gathered. You knew where every fact came from because you put it there. LLM output arrives with no such provenance. Treating it as a draft imports a lie into your process, the lie that the citations inside it have been checked by someone. They haven't. Not by the model, not by anyone.
A verification habit that takes ten minutes
You do not need a fact-checking department. You need a rule and a highlighter.
The rule: before publishing anything AI helped you write, highlight every proper noun, every quotation mark, and every number. Then verify each highlighted item against a primary source. Not a summary. Not another AI. The primary source.
For a quote, find the interview, book, or transcript where the person actually said it. If you cannot find it in under five minutes of searching, cut the quote or paraphrase without attribution. For a citation, open the actual paper or book and confirm the author, year, and claim. For a statistic, find the original report, not a news article that references it. The Pew Research Center, the Bureau of Labor Statistics, the WHO, and academic publishers all publish their underlying data. Use it.
This is boring. That is the point. Verification is boring in the way that flossing is boring. The people who do it are not smarter than the people who don't. They just do it.
What to do when a source cannot be found
If a quote or statistic that your AI draft produced cannot be located in primary sources, do not massage the wording and keep it. Assume it was fabricated. In the Rosenbaum case, several of the flagged quotes were attributed to real, prominent people who could have been contacted directly. They weren't. That is the gap where accountable writing lives.
Where YouWrite falls short here
Honesty first: no AI writing tool, including ours, can guarantee that generated citations are real. Refine can tighten prose, restructure arguments, and catch inconsistencies in tone, but it cannot verify that a quote from a 2019 conference talk was actually delivered. That responsibility sits with the writer. We can, and should, do more to remind users of that at the point of generation. Competitors that promise "researched" output while quietly relying on the same base models are being less than candid. If a tool cannot show you the URL of the source it is citing, treat the citation as unverified by default.
Competitor tools, briefly and fairly
Jasper and Copy.ai are optimized for marketing copy, where hallucinated quotes are less catastrophic and often absent by design. Sudowrite is aimed at fiction, where invention is the feature. Perplexity and tools with retrieval-augmented generation do link to sources, which is a real advantage for research-heavy work, though the linked source still needs to actually support the claim, which it sometimes doesn't. None of these tools remove the verification step. They shift where in the process it happens.
The writer who reads The Future of Truth mess and concludes "AI can't be trusted for serious work" is drawing the wrong lesson. The right one is smaller and more useful: highlight the proper nouns, open the primary source, and don't publish until every mark on the page can be traced back to something real. If you use YouWrite's Refine on an AI-assisted draft, run that highlighter pass first. The tool sharpens what you give it. It cannot audit what a different model invented three prompts ago.