Stanley Druckenmiller, the former fund manager who founded Duquesne Capital, published a column in the Wall Street Journal's opinion section on Monday under the headline "Let the Bond Market Speak," criticizing Treasury Secretary Scott Bessent's leadership. Readers were drawn to it at first by Druckenmiller's personal connection to Bessent, then started arguing about the prose instead. They posted results from the AI detection service Pangram, pointed at "not X, but Y" constructions and other phrasings characteristic of Claude, and concluded a model had drafted it. Druckenmiller soon confirmed he had used AI to formulate the argument. The Journal, which ran the piece with no AI disclosure, says it did not need one.
Druckenmiller's own defense was brisk. He said he moved from studying English to economics and now writes everything with AI for the same reason he uses a calculator in mathematics, and that he does not see why it should matter as long as the column carries the message he intended. His name is on the piece, he said, and the position published is his.
The response from the paper is the part that travels further. In a statement to Semafor's Max Tani, opinion editor Paul Gigot called AI "a fact of modern life" and said the column could run without disclosing AI use because it reflects Druckenmiller's "genuine opinion." People will use AI in their work and writing, Gigot said, for research, grammar checking, editing and other tasks. What matters to the Journal is whether the piece expresses the author's original argument and whether he has the standing and competence to make it. The paper has worked with Druckenmiller for years, he added, and nobody can doubt the column reflects his real position.
Most news organizations, the Journal included, have internal rules on staff use of generative AI. The newer development is that newsrooms are now having the argument in public about the outsiders who write for them, and they are not landing in the same place.
Last week the Financial Times issued a prominent retroactive correction to a column by Harvard professor Ricardo Hausmann, stating that the author had not told the paper AI was used to cut a longer draft down before submission and before the FT's own editing. The FT's editorial code forbids exactly that. In May, Futurism reported that the New York Times had imposed broad new restrictions on guest contributors after a run of AI incidents, among them a "Modern Love" column many readers took for AI slop and a book review largely produced by AI that turned out to be full of plagiarism. In its letter to freelancers, the Times said all text and visuals must be the product of human creativity and work, and that every submission must consist solely of the author's own original reporting, writing and other material.
Tani called this a precedent-setting moment: the editor of the second most important opinion page in the country has permitted AI drafting without mandatory disclosure. Axios made the historical objection that billionaires have long used ghostwriters and had staff prepare speeches and written arguments — but noted that drafting is part of assembling a coherent position, which leaves the question of how much the model shapes the author's reasoning and how much of that reasoning is being handed to a chatbot.
That distinction — whether AI moves an already-formed thought onto the page or participates in developing the position — is the one the Journal has declined to adjudicate, and I think the way it declined is more consequential than the column that prompted it. Gigot has replaced a question about process with a question about standing. The test is no longer how a piece was made but whether the byline is credible enough that nobody needs to ask. That is a workable instinct and an unworkable rule, because it cannot be written down without specifying whose name earns the exemption. Druckenmiller gets it because the paper has known him for years. Every contributor the paper has not known for years is on the other side of a line the Journal has not drawn in public.
The calculator analogy does not survive contact with the thing being described either. A calculator returns the answer to a question you already posed. A language model returns a draft containing sentences you did not write, including arguments you had not made until you read them — and Druckenmiller's own account is that he used it to formulate the argument, not to check the grammar.
What nobody has explained is how the standard could be verified by anyone, including Gigot. "Expresses the author's original argument" is a claim about a process that leaves no artifact: no draft history, no prompt log, nothing an editor could inspect afterward. The Journal's assurance rests entirely on knowing the author. And the paper has not said whether it will add a note now that the author has confirmed AI use, nor whether the same latitude extends to a contributor it has not worked with for years. Readers, not editors, are the ones who told the public how the column was made, and they did it with a detector.
Expect more of these constructions on opinion pages, in the Journal and elsewhere; the stylistic tells are the only evidence readers have. The awkward part is what the page is for. Journal opinion moves markets and policy, and it does so because readers price it as one identifiable investor's reasoning about the Treasury. Gigot's standard asks them to trust the byline rather than the process — and the byline is the one part of the arrangement that has always been easy to lend.