Whose thinking am I reading?

Machines are making fluent prose cheap. When everyone can produce a clean paragraph, clean paragraphs stop impressing anyone, and what stays scarce is what was doing the work all along, which is judgment, experience, and a claim specific enough to be wrong. PICS/ISTOCK

Since February, a video made with AI and posted on an Indian platform is supposed to carry a label. The amended IT Rules require synthetically generated content to be marked clearly and prominently, with provenance metadata underneath where the technology allows, and users must declare that what they are uploading is machine-made. India now has one of the strictest such regimes anywhere.

Last month an opinion column drafted with heavy AI help ran in one of the world's most influential newspapers with no label at all, and nobody thought it odd until a detection tool said so.

The column was Stanley Druckenmiller's, in The Wall Street Journal. He objected to Treasury Secretary Scott Bessent doubling the size of the American government's bond buybacks, which he read as propping up the price of long-dated Treasuries rather than keeping the market working. Druckenmiller spent four decades in markets and once worked alongside Bessent at Soros Fund Management, so the criticism read as a family quarrel held in public.

For about a day the story was about bonds. Then the economist Claudia Sahm ran the piece through Pangram, an AI-detection service, and posted the verdict: the text scored as entirely machine-written. Asked about it, Druckenmiller declined to be embarrassed. Of course he had used AI; he uses it the way he uses a calculator. He denied that the machine had written the whole thing and said he had rejected plenty of what it offered. Then he made the point that actually matters: "My name is on the piece. It's my message."

Paul Gigot, who runs the Journal's editorial page, backed him. Gigot's test is whether a column shows the contributor's own argument and whether the contributor has the standing to make it. Druckenmiller passed on both counts, and no note was appended.

Set that beside the Financial Times, which weeks earlier ran a column by the Harvard economist Ricardo Hausmann on Trump's tariffs. When that prose drew the same suspicion, the FT appended an editor's note: AI had been used to condense a longer draft before submission, and the paper's code forbids the use of AI in writing altogether.

Two serious newspapers, the same technology, the same month, opposite verdicts. That is not hypocrisy. It is what an industry looks like before it has decided what the problem is. India has not had this argument yet, and it is coming, because Indian newsrooms are already using these tools.

We are unusually well prepared to get it wrong, because our instinct is to regulate the artefact rather than the responsibility. A parliamentary standing committee examining the Press Council Act has already floated the idea that people producing news or news-like content with AI should be licensed, registered and made to carry disclosure labels. Set aside what licensing does to a free press. It would not even work. A label tells you a machine was in the room. It does not tell you the only thing a reader needs to know, which is who is answerable for the idea.

A byline has never meant that the person named typed every word. Politicians have speechwriters, chief executives have communications staff, and nobody has ever argued that the sub-editor who fixed your sentence now owns your argument. What the byline claims is intellectual responsibility: that the person named understands the reasoning, believes it, and will defend it when someone comes at it. By that standard Druckenmiller's case is easy. He has been saying this for years and can be questioned about it on television tomorrow.

Take two columnists. The first works out her position alone, reads the data, writes a draft, then asks a model to tighten it. The second asks the model for the strongest case against the government's policy, picks the best of what comes back, and builds a column around it. Both used AI, and both could say so honestly. But one has outsourced her expression and the other a share of her judgment, and a label describing them in the same words tells the reader nothing.

Think of it as a ladder: spelling and grammar at the bottom, then style, then structure, where the piece is rearranged and the argument staged more effectively, then substance, where the machine supplies interpretations and conclusions, and at the top generation, where a person contributes a prompt and a name. The question was never whether AI was in the room. It is how far up that ladder it was allowed to climb.

My own field solved a version of this long ago. Physics journals require contribution statements, so readers know who designed the study, who ran the analysis, and who wrote the manuscript. Nobody finds it humiliating; credit is worth something, and things worth something get specified. Opinion journalism never needed the equivalent, because until recently the only way to produce a plausible argument was to have one.

Detection cannot substitute for this. Pangram called the column entirely artificial; Druckenmiller says he rejected much of what he was given. Both may be true, because a detector reads cadence, not provenance, and cadence is exactly what a human editing a machine draft leaves behind. A rule enforced by such tools catches the honest and misses everyone else.

There is an upside. For centuries the ability to write well in English was an unearned advantage, and in this country a spectacularly unequal one: people with something to say went unheard because they could not phrase it in the right register, while people with nothing to say were published because they could. Machines are making fluent prose cheap. When everyone can produce a clean paragraph, clean paragraphs stop impressing anyone, and what stays scarce is what was doing the work all along, which is judgment, experience, and a claim specific enough to be wrong.

So a byline used to mean: I wrote this. It should come to mean: I will answer for every idea in it, including the ones I did not think of first. That is the harder standard, and the one worth writing into a code of conduct, because what a reader needs to know was never whether a machine touched the text. It is whose thinking they are reading, and who is prepared to stand behind it.

It would be absurd to write this and not say so. An AI system checked my facts, cut my sentences, and disagreed with me twice, usefully. That is structural help, not just stylistic. What survived is mine.

Nishant Sahdev is a physicist at the University of North Carolina at Chapel Hill. He writes for Sunday mid-day and is a contributor to The Wall Street Journal.