AI detectors are flagging executives' posts as AI-written, and not everyone agrees with the verdict

So far the evidence is flags, labels and a watermark, all described in one Business Insider reporter's first-person account ^[1][2][3]. The account does not say what happens to a post once a LinkedIn user flags it, or who can read the mark Anthropic puts in Claude's text. For now, the cost to a named author is reputational.

Software on the reader's side can now make that judgement. Pangram, which the reporter calls the leading detector on the market, claims to be 99.7% accurate ^[4]. Its Chrome extension put a label on every post and comment in her LinkedIn feed ^[5].

Peter Pang's workflow shows how those labels land on an executive. The Creao AI cofounder gives an agent ideas or an outline, and the agent pulls research and drafts the post for him to read ^[6][8]. Review of that kind can leave most of the model's words in place. Under Pangram's thresholds, a post that is up to a fifth human-written still gets the AI label ^[1].

"The final text carries my views," he said, "and I am the one who hits publish." ^[9] Pang's standard is accountability, and a company could reasonably adopt it. It answers a different question from the one the label asks.

The reporter turned to a detector because, she writes, study after study has shown that most people are bad at spotting machine-generated writing ^[12]. After a few weeks with the extension running, she had a named founder and a tally of occasional AI users in her own network: a well-known economist, a few journalists and a lot of marketing people ^[13].

The labels themselves are contested. According to the account, a backlash has formed among people who call efforts to identify and root out AI authorship antiquated, punitive and defamatory ^[11]. The 99.7% figure comes from Pangram itself, and even at face value it leaves an error rate of 0.3% ^[2]. A policy on AI drafting also needs a line on what the company does when a detector calls an executive's own writing machine-made.

This quarter's decision is which rule to write down. A rule that signed posts are human-written protects the executive's voice, but anyone with a detector can test that claim next quarter. A rule that permits AI drafts with human review and disclosure concedes the label in advance. Either rule closes off the unwritten middle, where a company learns its own standard from someone else's label. I'd expect the disclosure rule to be cheaper to hold, because it survives an accurate label and a false one alike.