September 18, 2026, (Inside AI) — A new editing ritual has taken hold among writers: after finishing a draft, they scan it for em dashes, colons, and paired commas, then strip them out. The goal is not clarity. It is camouflage. The result is a safer paragraph, not a better one.
The em dash, a punctuation mark in English prose for more than 200 years, now functions as a confession. Writers remove it because AI detectors flag it as a machine tell. Colons and parenthetical commas face similar suspicion. Shared lists of these giveaways circulate in writing groups and newsrooms, turning punctuation into a liability.
This is not a minor stylistic shift. It is a structural change in how writing gets made. A second stage has been added to the writing process: after drafting, the writer must now prove humanity. That proof often means inserting flaws on purpose. Leave the typo. Choose the weaker word. The better word appears so often in published work that it no longer feels safe.
The cost falls hardest on non-native English speakers. A Stanford study ran seven popular detection tools over TOEFL essays. More than 60 per cent were flagged as AI. Every essay was written by a human. The same tools were almost never wrong about native English writers. Writing in a second language tends to be plainer and steadier, which these tools read as artificial. The tidy exam answer, with its clean opening and conclusion, resembles the shape a detector distrusts.
Read: The Uncanny Valley Now Swallows Human Identity, Not Just Robots
A second bill comes from the tools themselves. Cornell University researchers ran an experiment last year with writers from India and the United States. When Indian participants wrote with AI suggestions, their sentence structure and word choices moved toward American style. Writers are squeezed from both sides. The machine pulls prose toward someone else's norm, and suspicion punishes them when they do not get there alone.
The suspicion has reached book publishing. In May, the detection tool Pangram was run over the five regional winners of the Commonwealth Short Story Prize and flagged three, including Mehendi Nights by Indian winner Sharon Aruparayil, which it scored as 88 per cent machine written. Organisers asked the writers for working drafts and time-stamped files, and concluded no AI had been used. In August, Nigerian author Jerry Falade's crime novel was withdrawn from sale by his own agents. Falade denies using AI and says the scrutiny was racial. A detector says a text resembles machine writing. That is not the same as saying who wrote it. The burden lands on the author to prove something no tool can prove either way.
Some universities have switched detectors off, deciding they are too unreliable to serve as proof. The tools can be disabled. The habit will not be. A generation is being taught to read its own fluency as evidence against itself.
The deeper loss is not punctuation. It is that writers are learning to write badly on purpose. AI detectors are built around the idea that machine writing has predictable patterns. So does careful human writing. A controlled vocabulary and a sentence that lands where you aimed it are marks of someone who edited well. They are also marks a detector counts against you.
Ten newsletters and 10 posts on a feed share one cadence. Some of that is the machine writing. Some of it is people writing to prove they are not the machine. A few now describe themselves as writers from before ChatGPT, as though the year they learned the craft is a credential. Two opposite fears, and you cannot tell them apart on the page.
Read: No, AI Doesn’t Mean the End of Mathematics, at Least Not Yet
Writing is how we find out what we believe. An instrument you are afraid to use properly stops being an instrument. The detectors can be switched off. The habit will not be. Stop doing the second job. Write the sentence you meant. Keep the em dash if you want to. Let the suspicion be somebody else's problem.
