Fast Company's Pete Pachal spent about a year and a half deliberately taking em dashes out of his drafts, on the grounds that chatbots overuse them and their presence had become a tell [12]. A study he cites found that among the major models, only Claude still uses em dashes more often than human writers do, because ChatGPT, Gemini and others now avoid them on purpose [9]. His conclusion is that the defensive edit backfires: strip the dashes and you read more like ChatGPT, not less [13]. Any signal cheap enough for a reviewer to spot is cheap enough for a model release to sand off.
That is the ground a content team stands on when it drops a classifier into the workflow. The detector scores the artefact, while every question in the room is about the author. Gigot's test, that the published piece reflects the author's original argument and that the author has the standing and credibility to make it [6], has the useful property of being answerable about a column that was dictated or drafted by a model, and of not expiring when the next model ships.
Both numbers in this story are worth holding at once. Graphite's analysis puts AI-generated articles at roughly half of everything published online, level with the human-written share [1]. Semafor's sample of contributors to the Journal, the Times and the Post found about 3% coming back as entirely AI-written, the Journal a little higher at 5% [7]. These are different populations counted by different means, so treat the ratio loosely, but the pooled op-ed rate is about a sixteenth of the web-wide share (3 divided by 50) [10], and the Journal sits roughly two thirds above the three-paper figure (5 minus 3, divided by 3) [11]. Pages that already carry a named author and an editor are where detectors find the least to flag.
The reaction to Druckenmiller was not about accuracy. Nobody argued the column was wrong; NOTUS quoted a journalism professor saying the AI use raised questions about how much time and thought actually went into the piece [8], and the evidence people pointed at was a single transition, "There is a quieter cost, too." [15] Pachal's summary is that style stood in as a proxy for effort, and effort as a proxy for credibility [14]. That judgment plays out in the reader's head, beyond any classifier's reach, which is the second of the three failures he lists: the tell registers where detection software gets no vote [2].
The practical test, before anyone buys seats, is the message a desk would actually send to a writer it flagged. A tool's score that the draft came back as machine-written amounts to a number the freelancer will deny and the reader will never see, not a policy. A brief whose promised interview notes never arrived, or an argument the author cannot extend for two minutes on a call, points instead to where the real work sits: provenance asked for at commissioning time (drafts, notes, sources, a disclosure line), and a standard published in advance, so Friday's answer matches the one given on Monday.
