Poynter: AI learned to write from humans. Now humans are trying not to sound like AI.
Matthew Crowley at Poynter published this piece on September 28, 2026, documenting a shift that editorial teams are navigating in different ways: the concern that human-written content will be misidentified as AI-generated. Crowley coins the term “flagxiety” for this anxiety and traces its effect on editing decisions, particularly around the em dash.
The em dash debate became a flashpoint at the ACES VCON conference. Editor Morgan Strunsky argued that removing well-placed em dashes degrades writing quality in the name of avoiding AI detection. Merrill Perlman, former head of the New York Times copy desk, offered historical grounding: LLMs use em dashes because writers have used them for centuries, and a punctuation mark that appears in human writing since long before any AI model exists cannot be reasonably treated as an AI signal. The practical tension — editors choosing not to use a grammatically legitimate construction because a detection tool might flag it — is one concrete way AI detection anxiety has begun distorting editorial judgment.
The article is more broadly about which surface features have come to read as AI markers and why. Kory Stamper, a lexicographer, identifies a quality AI consistently struggles with: contingency. Meaning depends on who is writing, about what, in which culture, and in which era. AI output tends to produce language that is grammatically correct but contextually flat, because the model cannot fully weight the particulars that give a sentence its specific meaning for a specific audience. That flatness is recognizable even when no single word is obviously “AI.”
Other patterns Crowley identifies as common AI markers include overly arch or ironic tone, negative parallelism, stiff hedging language, and lists used where prose would carry the structure better. The Economist noted in July 2026 that Claude uses em dashes as frequently as human writers do; ChatGPT uses them less. The variation across models means no individual stylistic choice reliably identifies AI authorship — a point that makes detection tools inherently unreliable.
The article also places the em dash debate in a larger economic context. The Guardian reported in September 2026 that freelance job listings in automation-exposed categories dropped 21% after ChatGPT’s release. The AP Stylebook added new entries for “AI slop” and “AI agent.” These are parallel signs that AI has changed what professional writing is expected to look like, how it is valued, and how it is evaluated — and editors are working through those changes in real time, often without settled answers.
Useful for editors, staff writers, and content leads thinking about what guidelines to set on AI style markers, how to approach AI detection tools critically, and what it means to maintain editorial standards in an environment where the boundary between human and AI-generated text is genuinely difficult to locate.