Machine-written prose has started appearing on the opinion pages of the three most prestigious newspapers in the United States, but not yet in the volumes the panic suggests. An analysis of 310 guest submissions published over one month by The Wall Street Journal, The Washington Post and The New York Times found 10 that an AI detector labelled at least 80% machine-generated, with another 40 flagged as partially AI-written.
Put the other way around: if the tool is accurate, 260 of 310 guest columns were written by people using nothing but their own judgment. The interesting part of the finding is not the count. It is that the three papers have three different rules, and that at least one of them has decided the rule does not much matter.
How the count was made
The analysis, published by Semafor, used a detector called Pangram, which runs custom models over submitted text and labels a piece “AI” when more than 80% of it appears machine-generated. The company claims a false positive rate of 0.01%.
That claim deserves the scrutiny any detector claim gets. Stanford research has found that detection tools disproportionately flag writing by people whose first language is not English – a limitation with direct bearing on at least one of the flagged pieces, and a reason to treat individual results as an indication rather than a verdict.
The op-ed that started the argument
The analysis followed a public fight over a Journal column by billionaire investor Stanley Druckenmiller, critiquing Treasury Secretary Scott Bessent’s interventions in the bond markets. Pangram scored it 100% AI. It read like it – “This wasn’t liquidity management, it was price management,” ran one line – and it circulated widely in Washington and on Wall Street anyway, both before and after its origins were confirmed.
Druckenmiller did not retreat. “There’s a reason I moved from an English major to being an economics major,” he told NOTUS. “I write everything using AI now for the same reason I use a calculator when I do math problems.”
His editor backed him. Journal opinion page editor Paul Gigot said AI is “a fact of modern life,” and that what matters is “whether what we publish from contributors reflects an author’s original argument, and if the author has the standing and credibility to make it.” That is a coherent position, and it is worth stating plainly what it concedes: the value of a guest column lies in whose name is on it, not in who typed it.
Where the rules say otherwise
The other two papers have written the opposite rule and are living with awkward results. The New York Times prohibits AI in “developing and drafting guest essays.” Pangram nonetheless flagged an August 2 guest essay by Jen Easterly, the former director of the Cybersecurity and Infrastructure Security Agency, along with 11 other partially AI-written Times guest pieces over 30 days. One flagged passage in the Easterly essay read: “In an era of nation-state cyberconflict, the ultimate measure of resilience is brutally simple: When attackers get in, clean water must still come out.” Easterly did not respond to a request for comment, and the Times declined to address the finding.
The Washington Post requires guest writers to confirm that a submission was “not created or manipulated with artificial intelligence or editing software.” Pangram identified 17 Post guest articles as partially or wholly AI-written, including an August 9 column by Dartmouth provost Santiago Schnell that it scored 100%.
Schnell’s response is the most useful in the whole story, because it describes what most flagged authors are probably doing: “I wrote this essay based on my own thinking. Then, I used ChatGPT to refine a few arguments, check for grammatical errors, and submit it to the Washington Post,” he wrote, adding that the piece “went through multiple rounds of edits by the Post.” A native of Venezuela with a doctorate in mathematical biology, he also noted evidence that detectors “are biased against non-native English speakers, such as myself.”
The line nobody has drawn
Every policy in this story collapses at the same point: none of them distinguishes between generating an argument and polishing one. A non-native speaker who runs a finished draft through a model for grammar trips the same detector as a fund manager who described his view in three sentences and published what came back.
Those are different acts. One is a spell-checker with better manners; the other outsources the reasoning that a bylined opinion is supposed to represent. Papers that ban “AI” without saying which they mean will keep flagging the first and missing the second, since a lightly edited machine draft defeats detection more easily than an honest polish does.
Why it matters more than the numbers suggest
Opinion pages sell a specific product: the reasoning of a named person with standing. Readers accept ghostwriting from politicians and executives because a human ghostwriter still interviews the principal and produces the principal’s argument. A language model does not have access to what the author thinks – only to what people who sound like the author usually say.
The Financial Times appended a note to a guest column this month after readers questioned it, conceding AI had been used to shorten the piece. That is roughly where the industry stands: prohibitions on paper, disclosure after the fact, and no agreed test. NarwhalTV has covered institutional responses to the same problem in New York City’s school AI ban.
