The Scoop
Humans are still writing the vast majority of guest columns published in the opinion pages of The Wall Street Journal, The Washington Post, and The New York Times, a Semafor analysis found — but large language models were creeping in well before a high-profile Journal op-ed provoked controversy this week.
asserted
models → write → controversy
Over the past month, 10 out of 310 guest submissions to the three publications were labeled as at least 80% “AI” by AI detector Pangram, suggesting AI is not yet making a meaningful dent in human contributions to the most prestigious publications.
asserted
AI → label → publications
Another 40 articles were partially AI-generated.
asserted
articles → generate → ?
Pangram uses custom AI models to analyze text and other content for AI and claims a 0.5% false positive rate.
uncertain
Pangram → use → rate
If Pangram finds more than 80% of the text is AI generated, it is labeled as AI.
asserted
it → find → AI
In informal testing by Semafor, Pangram was very accurate in detecting AI-written text, though some complain of false positives.
asserted
some → detect → positives
Semafor’s analysis follows a debate sparked Tuesday, when billionaire Stanley Druckenmiller — and later The Wall Street Journal’s opinion page editor, Paul Gigot — stood behind Druckenmiller’s use of AI to write an opinion piece critiquing Treasury Secretary Scott Bessent’s interventions in the bond markets.
asserted
Druckenmiller → follow → markets
The piece had the wooden hallmarks of AI text (“This wasn’t liquidity management, it was price management,“) but served its purpose as a blunt rebuke from a titan of finance to his protégé, and circulated widely in Washington and on Wall Street before — and after — its AI roots were confirmed.
asserted
roots → have → Street
Pangram gave it a 100% AI score.
asserted
Pangram → give → score
“There’s a reason I moved from an English major to being an economics major,” Druckenmiller told NOTUS.
asserted
Druckenmiller → ’ → NOTUS
“I write everything using AI now for the same reason I use a calculator when I do math problems.”
asserted
I → write → problems
Gigot said in a statement that 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.”
asserted
author → say → it
The New York Times prohibits the use of AI in “developing and drafting guest essays.”
asserted
Times → prohibit → essays
A Times spokesperson didn’t comment on Pangram’s finding that an Aug. 2 guest essay by Jen Easterly, who served as director of the Cybersecurity and Infrastructure Security Agency under President Joe Biden, was flagged as AI.
asserted
who → comment → AI
In one example of some of the text that was flagged by Pangram, Easterly wrote “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.”
asserted
water → flag → resilience
Easterly did not respond to a request for comment.
asserted
Easterly → respond → comment
The tool identified 11 other guest Times articles over the past 30 days that were partially AI-written, according to the analysis.
uncertain
that → identify → analysis
Pangram also identified an Aug. 9 guest column in The Washington Post, Universities are fighting AI cheating.
asserted
Universities → identify → cheating
“I wrote this essay based on my own thinking.
asserted
I → write → thinking
Then, I used ChatGPT to refine a few arguments, check for grammatical errors, and submit it to the Washington Post,” Schnell wrote, adding that the piece “went through multiple rounds of edits by the Post.”
asserted
piece → use → Post
Stanford research has also suggested that detection tools sometimes flag work by people whose first language isn’t English.
uncertain
language → suggest → people
Schnell, a native of Venezuela who has a PhD in mathematical biology, noted there is evidence that “they are biased against non-native English speakers, such as myself.”
asserted
they → have → myself
A spokesperson for The Washington Post referred to its policy that requires guest writers to confirm their submission was “not created or manipulated with artificial intelligence or editing software,” but did not comment on the 17 guest articles published in the paper and found by Pangram to have been partially or totally written by AI.
asserted
submission → refer → AI
Earlier this month, The Financial Times appended a note to a guest column after readers questioned whether it was written using AI.
asserted
it → append → AI
The FT, which conceded AI had been used to shorten the article, prohibits the use of AI in the writing process.
asserted
AI → concede → process
Perhaps I’ve grown a bit jaded, but I was surprised at how little AI-written content showed up in Semafor’s analysis.
asserted
content → grow → analysis
If Pangram’s tools are accurate, 260 out of 310 people who published guest articles in the NYT, WSJ, and Washington Post did it entirely with their own brains (or perhaps the brains of an underling).
asserted
who → publish → underling
My guess is it comes down mostly to quality and perhaps a slight fear of embarrassment.
asserted
it → come → embarrassment
Large language models just aren’t very good at writing.
asserted
models → write → writing
And even if you try to harness models to write in your own style, you’re not going to get a high-quality article in one shot.
asserted
you → try → shot
I don’t have a lot of confidence that LLMs will get much better at writing, though.
asserted
LLMs → have → writing
After ingesting the entirety of the web, this is where they ended up — a lowest common denominator of human discourse, and that is good enough for most things.
asserted
that → ingest → things
Frontier labs aren’t going to invest a lot of effort into making them better writers because there is no money in it.
asserted
them → go → it
The labs are concentrating on making models better at marketable skills, like legal work and accounting.
asserted
models → concentrate → work
They’re working to make models that behave reliably so that businesses can put them in charge of important things, not choose the perfect adjective.
asserted
businesses → work → adjective
What will make AI better at writing is its ability to learn on the job.
asserted
AI → make → job
Today, once a model is trained, it’s set in stone, but in the future, AI models might be able to update their weights on the fly — like humans do when they amass more vocabulary and develop their taste as they grow up, for example.
uncertain
they → train → example
So when AI models can learn over time how to write like individual people, we might have assistants that can pump out copy that is indistinguishable from us, hopefully at our best.
uncertain
that → learn → best
Until then, you have two choices: Use the 20-watt computer on your shoulders — or settle for a good-enough LLM.
asserted
you → have → LLM
Semafor’s methodology and reproducible analysis code can be found in a public GitHub repository here.
asserted
methodology → find → repository
…and 7 more, not listed.
The Aporia
