Publishers and editors police how writing is made. They can't verify it, and readers care only whether the work is good.
In August Stanley Druckenmiller admitted that he had written an op-ed for this page with the help of artificial intelligence. He uses AI, he explained, for the same reason he uses a calculator. Paul Gigot, the Journal's editorial-page editor, declined to apologize. No one doubted the opinion was Mr. Druckenmiller's own, he wrote, and his only sin against the "media pharisees" was admitting it.
Few other gatekeepers agree. Nieman Lab asked editors at 24 prominent newspapers whether they would follow Mr. Gigot's lead. None who answered would. The New York Times, USA Today and the Boston Globe bar nearly all AI help from outside contributors.
Book publishing is stricter still. Of the publishers and agents in the Duotrope database whose policies are known, 91% forbid AI. In March Hachette pulled Mia Ballard's novel "Shy Girl" over evidence of AI use. In July a debut thriller that had drawn 14 bidders and offers above $2 million was withdrawn after a journalist ran it through detection software. In September the Académie Goncourt dropped a best-selling novel from its prize list after an anonymous account posted detector scores. None of the three authors admits using AI.
All these rules test inputs. They ask how a work was made, not whether it is any good. The Luddites complained of shoddy stockings as well as lost wages, so the grievance is old. So is the economic error behind it.
How a manuscript was made is a sunk cost. Whether the author used a quill, a ghostwriter or a chatbot, the effort is spent. The gatekeeper's only question is whether taking the work on promises more than it costs. Method matters only if it changes that answer.
It can, through three channels. Method may signal quality when reading is expensive, and an agent buried in queries may sensibly discard those that smell of a machine. Method affects property: in March the Supreme Court declined to disturb the rule that copyright protects only human authorship, so a publisher buying machine-made prose may be buying less than it thinks. And some readers may prefer human work, as some shoppers prefer organic carrots.
Each channel is narrower than the gatekeepers suppose. The first shrinks as they use AI to read. The second touches machine-made text, not human work an AI helped improve. The third is a taste, which a market can price.
Where AI wins first depends on how cheaply output can be tested.
Mathematics has the cheapest test. A proof written in Lean, a programming language for proofs, checks or fails by computer, whoever wrote it. On Oct. 6 OpenAI posted 722 papers produced by an unreleased model, claiming progress on hundreds of open problems. About a fifth have their main results already checked in Lean. Of one, Rutgers mathematician Alex Kontorovich wrote that a human author would have earned "an instant Fields Medal." Mathematicians have protested since, but chiefly about credit, access to the model and the pace of release, not the soundness of the checked proofs.
Books have a test too: sales. But it is costly and comes late. A publisher must pay an advance and edit, print and market a book before learning anything. So agents ration the slots with cheap rules about inputs. AI attacks the rationing from both ends. It makes manuscripts cheaper to read, and self-publishing lets an author take the market test without anyone's permission.
Op-eds are hardest, because readers buy the paper, not the piece. With no price on any one essay, an editor guards the brand with a cheap rule. Yet an op-ed reader pays for the author's opinion, not his sentences. That was Mr. Gigot's point.
Input rules have a deeper flaw: they cannot be enforced.
Anthropic now watermarks the text its Claude models write, by biasing word choices in a pattern only its key can detect. The company concedes that the mark cannot show a human wrote anything, cannot detect rival models and vanishes when every word is rewritten. Detection software does worse. Newsrooms that tested the Goncourt novel got verdicts ranging from "very likely" human to 100% machine. In the same affair, a poet changed about 20 words in a passage and flipped its score from 100% machine to 100% human.
Statisticians settled the disputed Federalist papers in the 1960s by counting words such as Hamilton's "upon" and Madison's "whilst." They had three suspects, ample samples and no one trying to fool them. AI detection faces countless suspects, all of them mimics, and authors who edit.
So gatekeepers fall back on oaths. The Authors Guild's Human Authored mark, now on some 6,500 titles, rests on a signed promise backed by the threat of a lawsuit. The Guild even permits AI research, brainstorming and outlining, so long as no machine sentences appear. Human authorship has become what economists call a credence good, one that buyers cannot verify even after consuming it. So is "organic," and its premium draws fraud. In 2019 a Missouri farmer got more than 10 years in federal prison for passing off over $120 million of grain as organic.
An honor system screens out the honest. The author who discloses is rejected; the one who lies gets through. Mr. Druckenmiller's offense was candor.
The output is the one thing anyone can check. Mathematicians check it with Lean, readers with their wallets, editors with their own judgment. For the price of a pizza a month, any writer can now hire a tireless research assistant and critic. Gatekeepers who keep asking how a book was made will find fewer authors asking for their blessing. Everyone else will judge the book.
No comments:
Post a Comment