On Monday, The Wall Street Journal published a sentence I wish I’d written. Of a Treasury yield the Treasury would rather not see, it said: that isn’t a crisis, it is an invoice.
Read it out loud. It has a pulse. The first half takes something away and the second replaces it with a slap. All of it fits in eight words. I have spent years trying to make sentences like that. Cicero wrote them for the Roman Forum.
Cicero was killed for his sentences, while Stanley Druckenmiller was given away by one.
The column ran under Druckenmiller’s name, against Scott Bessent—his old colleague at Soros and, by most accounts, his protégé—for holding down Treasury yields. On Tuesday, NOTUS ran it through Pangram, an AI detector, and Druckenmiller confirmed the result without much interest in the matter. Of course he used AI. He uses it the way one uses a calculator. He isn’t embarrassed. Nor the paper. The editorial page editor backed him: what matters is that the argument belongs to the author, and that the author has the standing to make it.
He is right that nothing was invented this week. Columns under large names have been drafted for decades by aides, ghostwriters and communications shops, most of them uncredited and undisclosed, sanded down until they could offend nobody. The taboo was never about the machinery. It was about saying out loud that the byline came loose from the labor a long time ago.
FOUND COIN
A signature does not certify who wrote it, but who answers for it.
What is new is stranger. We have begun to judge an argument by how it was made. The detector found no errors in that column. Instead, it identified a rhetorical pattern. The two-beat rhythm, the em dash, a certain regularity—the signs now circulating as proof of machinery. Adam Kirsch, in The Atlantic, gave the practice its right name: Lombroso, the criminal known by the shape of his jaw. Reading the inside off the outside fails now for the reason it failed then. There are writers dropping deliberate misspellings into their sentences to prove they are alive. I took four of those constructions out of this essay before publishing it. The cost is small. It irritates me anyway.
The suspicion does not fall on everyone alike. That is what the enthusiasts and the inquisitors miss. Druckenmiller can shrug: forty years of record, a reputation in the market he writes about, and a page that defended him in a statement the same day. The detectors won’t prevent the AI slop. They point first at the careful—the writer who learned the language late and still hears its joints, the dyslexic, or anyone who builds a sentence with effort. In May, the internet’s language police accused the Caribbean winner of the Commonwealth Short Story Prize of having written his story with AI. He was cleared, but only after handing over his drafts, timestamps, and notes. The publisher that had supported the prize for years withdrew before the verdict. Proving his innocence cost him his privacy. The accusation cost nothing.
I spent part of the summer writing a paper about this for the Aspen Institute in the mornings, and my novel by hand in the afternoons. Neither habit is a virtue. One of them is just slower. Detection is the wrong instrument. The question is not so much whether it can fool me—today the answer is yes—as whether the author could regret it. Regret requires a self with something to lose. It is a test the machine cannot pass, and it is what the reader is after when he wants to know whether anyone was behind the page. By that measure, Druckenmiller passes. Whatever else that column was, it was aimed.
But the Journal’s defense only half works, and the half that fails matters. Accountability is what the author owes the subject. Disclosure is what he owes the reader. Two debts. Settling the first does not cancel the second.
For five centuries, polish was evidence: a well-made sentence showed that somebody had worked on it. Polish is free now, and the look of a page no longer says whether anyone was behind it.
FOUND COIN
For five centuries, polish proved somebody had been there. Now it proves the opposite.
In May, a book about how artificial intelligence damages our relation to the truth turned out to be full of quotations invented by its own tools. It had blurbs, it cleared the checks, it was well made. That was the problem.
What the reader can no longer deduce, he has to be told. There is no third option, and no program is going to create one. So the work is dull and institutional: standard clauses on training rights, so that an author and a platform start from the same place. A disclosure standard a publisher can apply and a librarian can catalog by. Collective licensing cheap enough that following the rules isn’t a luxury. None of this is a law. All of it is a table: publishers, guilds, translators, labs and librarians arguing in public instead of finding out afterward. I proposed that Aspen, or any institution able to build dialogue and agreement, convene it. I prefer that to another round of detector scores used as evidence.
There is an Editor’s line under my name at the foot of this essay, like in the Aspen paper, both in fine print: an AI tool helped copyedit it. It looks like a confession. It isn’t.
I’ll keep putting it there. Not because anyone is checking.