In November 2025, media commentator Steven Rosenbaum's book The Pirate Within was pulled from sale after readers noticed fabricated quotes, invented sources, and misattributed passages throughout. Rosenbaum's public explanation, quoted in The Guardian and 404 Media, boiled down to a line that has already become a meme: ChatGPT fucked up the book.
At YouWrite we build Refine to sit between a draft and a final manuscript, so the story landed with a particular dread. Not because the tool was blamed, but because the reasoning behind that blame is so seductive that most writers would reach for it under pressure.
The seduction of blaming the tool
When a fabricated Susan Sontag quote appears in your manuscript under your name, there are two possible framings. First: you inserted a Susan Sontag quote without checking whether Sontag ever said it. Second: ChatGPT gave you a Sontag quote, and you assumed it was real.
Both describe the same act. Only one puts you in the frame as the responsible adult.
The second framing feels technically accurate. The words did come from the model. The attribution error was, in some literal sense, generated. But the framing quietly relocates authorship of the decision to publish from the writer to the machine, and that is where it becomes a lie. The model did not decide to include the quote. It did not decide to skip verification. It did not sign the contract, cash the advance, or approve the galley.
A writer under deadline, working with a system that produces fluent, confident, plausible prose, is placed in a cognitive situation that human brains handle poorly. Fluent output triggers trust. The processing fluency literature has documented this since Rolf Reber and Norbert Schwarz's work in the early 2000s. Text that reads smoothly is judged as more likely to be true. ChatGPT's default register (calm, structured, mildly authoritative) is essentially weaponized fluency.
So the failure mode is not carelessness in the moral sense. It is a predictable interaction between a specific tool behavior and a specific human bias. Which is exactly why the writer, not the tool, has to design against it.
What the model structurally cannot do
Be precise about the machine's actual capabilities. Vagueness here is what lets writers off the hook.
A large language model, at inference time, produces statistically probable next tokens given its training and your prompt. It is not:
- Retrieving a quote from a verified source database.
- Checking whether a person named in its output exists.
- Distinguishing between a fact it was trained on and a fact it interpolated.
- Aware of the boundary between what it knows and what it is confabulating.
When ChatGPT produces the sentence As Joan Didion wrote in her 1979 essay for Harper's, "...", the model is not looking anything up. It is generating a shape of sentence that essays about Joan Didion tend to contain. The quote may be real, misattributed, paraphrased, or entirely invented, and the model has no internal signal that distinguishes these cases. Retrieval-augmented systems and tools with browsing narrow this gap. They do not close it. Even a cited URL can be hallucinated, or real but misquoted.
This is a structural property, not a bug that will be patched next quarter. Any workflow that treats model output as research is broken by design.
The line of responsibility
Once you accept the structural limit, the division of labor becomes obvious. The model can draft, restructure, tighten, suggest. The writer owns every factual claim, every attribution, every name, date, statistic, and quotation that appears under their byline. Not because of an ethical abstraction. Because no one else in the pipeline can.
This is where The Atlantic's coverage of the Rosenbaum affair, focused on ethics and scandal, leaves the practical question unanswered. Yes, it was dishonest. Now what does a working writer actually do differently on Monday morning?
The review checklist you have to own personally
Any time an AI system touches factual or attributed content in your draft, run this pass yourself. Not your assistant. Not another AI.
1. Extract every factual claim
Go through the AI-touched sections and mark every sentence that asserts something about the world: a name, date, statistic, event, study, or quote. If a sentence would be wrong if the fact were wrong, it is on the list.
2. Verify quotes against a primary source
For every quotation, find the original book, article, interview, or transcript. Not a quote aggregator site. Not another AI. If you cannot locate the primary source in under ten minutes, cut the quote or rewrite the passage without it. Quote-hallucination is the most common failure mode and the one most likely to end your career.
3. Confirm every named person exists and did the thing attributed to them
Search for the person. Confirm the affiliation, the year, the paper, the role. Models frequently invent plausible-sounding academics with plausible-sounding papers at real universities.
4. Check statistics against the cited source, not a paraphrase
If the draft says a 2021 Pew study found 63 percent, open the Pew study. Look at the actual number. Models routinely shift percentages by five to fifteen points and invent study years.
5. Read one full pass aloud, cold
Read the final manuscript aloud with the AI tool closed. Fluency masks nonsense. Your ear catches what your eye slides past.
6. Log what the AI touched
Keep a simple record of which sections were AI-generated, AI-edited, or human-only. If something breaks later, you know where to look. This is boring and it will save you.
A word about our own tools
Refine, the YouWrite service that polishes and restructures drafts, does not verify facts. It cannot. We tell users this in the product, but we could tell them more often and more loudly, and we probably should. Any editing layer built on an LLM inherits the same structural limits described above. If we implied otherwise in marketing copy, that would be our version of the same dodge Rosenbaum made.
Competitor tools handle this disclosure with varying honesty. Sudowrite is upfront that it is a fiction tool, not a research assistant. Jasper leans into enterprise polish and tends to underplay the verification burden. Grammarly, now branded as an AI writing partner, is quieter than it should be about hallucination risk in its generative features. Perplexity does the best job of surfacing sources, but its citations still require the same manual check as any other output.
The writer's job did not change when the models arrived. The surface area of possible errors did. Rosenbaum's book was pulled because someone downstream did the checking he did not. Next time, the reader might just quietly close the book, and tell three friends not to buy it.