The Future of Truth is a book about how AI is changing the nature of truth. It was published on May 12, 2026, with a foreword by Nobel Peace Prize–winning journalist Maria Ressa. Within a week of its release, The New York Times had found more than six quotations in it that were fabricated or misattributed. One was pinned on the tech journalist Kara Swisher—who said flatly that she had never said any such thing. The line had been invented by AI.
The author, Steven Rosenbaum, explained that this was an accident, not deliberate fabrication. Early in the writing, he said, he had used AI as a research tool, and he had flagged the AI-sourced material separately and sent it to his publisher's fact-checker. The fake quotes made it into the book anyway. A book meant to warn about the ways AI threatens the truth had been undone by exactly those means.
It's tempting to read the episode as the familiar warning that "AI makes things up." But peel back a layer and a different structure comes into view. AI didn't create a brand-new problem; it exposed a gap that publishing had left open for a long time.
A System Where No One Owns the Checking
In nonfiction publishing, fact-checking is not a contractual obligation. According to New York magazine, which reported on the case, hiring an outside fact-checker costs anywhere from $7,000 to more than $10,000 per book. Because of the expense, the step has long been skipped. In the same report, a senior nonfiction editor at a major publisher described finding errors in a book and raising the question of fact-checking, only to be told that "the publisher is not responsible for fact-checking or for hiring an outside verification firm, and that doing so would shift the liability onto the publisher."
Authors, for their part, are unlikely to spend heavily on verification out of their own pockets. One literary agent said there is "no established system." The issue comes up in every publishing contract, the agent noted, but no one ever resolves it. Few publishers have specific contract language governing the use of AI. One editor said she does send authors guidance not to use AI in certain ways, but that it "isn't legally binding, and once you're in the editing stage, it ultimately comes down to trust."
Responsibility for verification is pinned nowhere. Publishers say they have no obligation, authors can't absorb the cost, and agents say there's no system. Most of the time, the gap never surfaced as a problem. At the speed a human writes, errors crept in only rarely, and that much was tolerated. Once AI pushed up the pace of production, the responsibility no one had claimed turned into an accident.
A Forecast of the Publish-Before-You-Verify Age
This structure was diagnosed before AI ever arrived. In Too Big to Know, David Weinberger argued that the internet changed how knowledge works by driving the cost of publishing to nothing. A strategy of publish first, verify later took hold, he wrote, producing a vast heap of data that was made public—and accessible to anyone—before it had been checked. The way facts had once served as the foundation of knowledge was, in his telling, transformed.
What Weinberger saw is that when the cost of publishing approaches zero, fact-checking is the first thing to go. In the age of paper, the cost of printing and distribution was itself a barrier to publishing. Because putting out a single copy cost money, a publisher had to choose what to release—and review entered the process along the way. The internet erased that cost. Once publishing became free, the filtering step disappeared with it.
AI went a step further. If the internet made publishing cheap, AI made production cheap too. Now that one person can generate a book's worth of draft in a few days, only fact-checking remains expensive and slow. That work resists automation, because a human has to cross-check each claim against its original source, and the structure in which a single book runs into thousands of dollars hasn't changed. Production has sped up many times over, but verification can't keep pace. The wider the gap between the two speeds, the greater the chance that an unchecked sentence hardens into a printed book.
Rosenbaum's book is an accident produced by that gap. AI generated the sentences quickly, verification was pinned nowhere, and the result hardened into a book and went to market. Had it been written one line at a time by a human, six fake quotes would have been unlikely to pile up in a single volume. When the speed of production overtakes the speed of verification, error stops being a rare slip and becomes a product of the system.
The Solo Creator Has No One to Delegate Verification To
Here the bind facing solo operators and creators is even starker than the publisher's. A publishing house at least has multiple layers—editors, outside fact-checkers, legal review. For the solo creator, all of those layers collapse into one person. There is no one to offload the checking onto.
Someone who publishes their own e-books, writes their own newsletter, and produces their own video scripts controls every stage of production alone. Using AI makes that several times faster. Where a single piece once took days, from research to first draft, the process now shrinks to a few hours. The problem is that verification doesn't speed up the way production does. When you found and read the sources yourself, you wrote with the source in hand; when you write from an AI's summary, sentences that never passed through the original source go straight into the piece. The step a publisher skips to save money, the solo creator skips for lack of time and effort. The result is the same: an unverified sentence goes to market under your own name.
Rosenbaum had a publisher and a distributor behind him, and still the responsibility was entirely his. In his statement, he accepted full responsibility for the errors. The solo creator doesn't even have anyone to share it with. Even if a sentence was written by AI, the person who put their name on it is the one who answers for it. To put your name on something is to vouch not for where it came from, but for whether it's true.
So how far should verification go, and how? Sending every sentence to an outside expert isn't realistic for a solo creator. It's better to narrow the scope and focus.
Where to start? The very place Rosenbaum's accident began points to the answer. The fakes originated in direct quotations—words placed in Kara Swisher's mouth that she never said—because the quotation is what AI fabricates most convincingly. Anything inside quotation marks is safer if you confirm against the original source that the person actually said it. Next come numbers and dates. AI presents specific figures—"about 70 percent," "2023"—with confidence, but that confidence is separate from accuracy. Statistics, dollar amounts, and years should be re-checked against a primary source. Last are proper nouns. The names of people, companies, books, and events read smoothly even when AI gets them wrong, so they're easy to slip past. It's worth pausing once to confirm that they exist. Quotes, numbers, proper nouns: these three are where AI's errors leak out most often, and where a human can most reliably stop them before publishing.
There's a reason to focus on these three. All of them resolve to a single fact, so you can verify whether each is right or wrong on its own. Interpretation and opinion, by contrast, can't be settled by checking against a source. For a solo creator who can't add an infinite number of hands, it makes sense to pour the available time into the items that can actually be checked. It helps to flag the information you pulled from AI and to build into your workflow a step that, before publishing, revisits only the flagged sentences. Rosenbaum did go as far as flagging his AI-sourced material. He simply left out the step of checking it afterward.
Having Gained Speed, It's Time to Redesign Verification
AI has given the solo creator productive power on par with an entire publishing house. What it hasn't handed over are the verification stages a publishing house has. Production can be delegated; the responsibility of putting your name on something cannot.
What Rosenbaum's case leaves behind is not a lesson to avoid AI. He said he used AI only as a research tool—but he pinned the responsibility for verification on no one. The faster the production tool in your hands, the more you need to decide in advance who checks, when, and how. The gap a publisher leaves open to save money is one the solo creator can design back into their own work. If you can't place the verification stage outside yourself, the realistic alternative is to turn the final pre-publication check into a single fixed routine.
The publish-before-you-verify age has already arrived, just as Weinberger forecast. Kara Swisher learned that words she never spoke had been printed in a book only after the fact. Had someone checked that one sentence against its original source before publication, the book would not have collapsed. When you put your name on something and send it out into the world, deciding who runs that check, and when, has now become part of the work itself.
