A Semafor analysis using the AI-detection tool Pangram found that most guest columns published over the past month in The Wall Street Journal, The Washington Post and The New York Times were still human-written, but instances of substantial AI authorship are already appearing.
What the analysis found
Out of 310 guest submissions to the three publications over the past month, Pangram labeled 10 as at least 80% AI and found another 40 to be partially AI-generated. Pangram says it uses custom AI models to analyze text and claims a 0.5% false positive rate; it designates a piece as AI if more than 80% of the text is judged to be machine-generated.
Semafor’s informal testing found Pangram to be largely accurate in detecting AI-written material, though critics have complained about false positives.
High-profile examples and reactions
A debate intensified Tuesday after billionaire Stanley Druckenmiller acknowledged using AI to write an op‑ed in The Wall Street Journal critiquing Treasury interventions, and the paper’s opinion editor, Paul Gigot, defended publishing it. The piece included phrasing that many readers described as characteristically machine-like (for example: “This wasn’t liquidity management, it was price management.”) Pangram assigned the column a 100% AI score. Druckenmiller told NOTUS: “There’s a reason I moved from an English major to being an economics major. I write everything using AI now for the same reason I use a calculator when I do math problems.” Gigot said in a statement that AI is “a fact of modern life,” and that the important question is whether published contributions reflect an author’s original argument and whether the author has the standing and credibility to make it.
The New York Times and other flagged pieces
The New York Times prohibits the use of AI in developing and drafting guest essays. Still, Pangram flagged an Aug. 2 guest essay by Jen Easterly, who served as director of the Cybersecurity and Infrastructure Security Agency under President Joe Biden. One Pangram‑flagged sentence from that 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. Pangram also identified 11 other Times guest pieces over the past 30 days as partially AI-written.
Washington Post, Dartmouth provost and detector bias
Pangram identified an Aug. 9 guest column in The Washington Post by Dartmouth provost Santiago Schnell as 100% AI-generated. Schnell wrote that he composed the essay from his own thinking and then “used ChatGPT to refine a few arguments, check for grammatical errors, and submit it to the Washington Post,” adding that the piece “went through multiple rounds of edits by the Post.”
Research at Stanford has suggested detection tools can sometimes mislabel work by people whose first language is not English. Schnell, a native of Venezuela with a PhD in mathematical biology, noted there is evidence these detectors are biased against non-native English speakers.
A Washington Post spokesperson pointed to the paper’s policy requiring guest writers to confirm their submission was “not created or manipulated with artificial intelligence or editing software,” but the Post did not comment on the 17 guest articles Pangram found to be partially or totally AI-written.
Other media responses and policies
Earlier this month the Financial Times appended a note to a guest column after readers questioned whether AI had been used; the FT acknowledged using AI to shorten the piece even as it prohibits AI use in the writing process.
Analysis and research on preference and quality
Semafor’s writer expressed surprise at how little AI-generated content appeared in the sample: if Pangram is accurate, 260 of the 310 people who published guest pieces in the three papers relied entirely on human work. The writer hypothesizes the reasons are mostly quality concerns and a fear of embarrassment—current LLMs are still imperfect writers and generally don’t produce a high-quality article in one pass.
Michigan researchers found that many readers actually prefer AI-generated content to human content, but MFA-trained experts tended to favor human-generated writing when reading outputs from off-the-shelf chatbots. That changed after researchers fine‑tuned models on high-quality human-written text; the fine-tuned models were then preferred even by MFA-trained experts. (Fine-tuning steers a model toward a style without changing its underlying weights.)
Some observers also associate a particular, low-quality style — nicknamed “Claudish” after Anthropic’s chatbot — with computer-generated text. Many readers still find carefully crafted human writing more persuasive than machine-produced word salad.
Why it matters
The presence of AI in guest opinion pages of major national newspapers raises questions about editorial policy, transparency and credibility. Detection tools have limits and potential biases; publication practices and authors’ disclosures vary between outlets. As tools and newsroom standards evolve, debates over how to handle AI-assisted submissions are likely to continue.
