In 1759, an English novelist named Laurence Sterne published the first volumes of Tristram Shandy, a comic novel packed with em dashes. In 2026, one em dash can get your email accused of being written by a robot.

Here is the charge sheet. Use an em dash — like this one — and you are a suspect. List three things in a row, and you are a suspect. Put a comma before the "and" at the end of a list, and you are a suspect. Lay out a page with neat, balanced sections, and the accusation arrives before anyone reads your point.

The logic runs like this. ChatGPT, the AI assistant made by a US company called OpenAI, uses these habits a lot. So anyone who uses them must be ChatGPT. By the same logic, anyone who owns a broom can fly.

For the record, every em dash in this article was typed by a human. That human is me. I put them in on purpose — and I am keeping them.

The machine learned to write from us.

Exhibit A: the em dash has a long record

So let us open the case files. The em dash was a working part of English novels by the 18th century. Sterne used it to copy the stop-start rhythm of someone thinking out loud. Jane Austen used dashes to blank out names, such as the army regiments in Pride and Prejudice. Readers felt they were reading about real people.

Then came Emily Dickinson, the American poet. She wrote nearly 1,800 poems and filled them with thousands of dashes. She died in 1886, and her editors stripped most of the dashes out before her poems went to print. According to the design podcast 99% Invisible, her first collection, published in 1890, went from 1,151 dashes to 52.

So in 1890, editors were already scrubbing em dashes out of good writing. The panic is old. It simply has new staff.

The story now has a sequel. In November 2025, Sam Altman, the chief executive of OpenAI, announced a fix. ChatGPT would finally drop the em dash if users asked it to in their settings. He called it a "small-but-happy win". So the robot can now be talked out of the em dash. Emily Dickinson remains under investigation.

Exhibits B and C: a US$5 million comma and a Roman general

The prosecution moves on to the Oxford comma. It takes its name from Oxford University Press, the publishing house of the University of Oxford. Horace Hart, the university's printer, kept a book of house rules for his typesetters. In July 1905, a new edition added an example with the comma in place: "Peter was a wise, holy, and energetic man." The comma had no official name until 1978, when a history of the press called it the Oxford comma.

It also has a price tag. In 2014, three delivery drivers sued Oakhurst Dairy, a dairy company in Maine in the US, for unpaid overtime. Maine's overtime law exempted a list of jobs ending in "packing for shipment or distribution". With no comma after "shipment", it was unclear whether distribution counted as a separate exempt job. The drivers' job was distribution.

On 13 March 2017, a US federal appeals court ruled that the law was unclear, and sided with the drivers. Judge David Barron opened his ruling with a line every editor now owns: "For want of a comma, we have this case."

Here is the twist. The law had followed Maine's own drafting manual, which told lawmakers to leave the Oxford comma out. In February 2018, the dairy agreed to pay US$5 million to settle. Maine has since rewritten the list with semicolons, the punctuation version of a restraining order.

The rule of three goes back to ancient Rome. Around 47 BC, Julius Caesar reported his quick win at the Battle of Zela in three words: "Veni, vidi, vici." Aristotle's Rhetoric, the ancient Greek handbook on persuasion, discusses balanced phrasing of this kind. ChatGPT was released in November 2022, a little late to the party.

Symmetrical layouts and rhetorical framing face the same charges. A balanced page and a well-built argument are exactly what writing teachers ask for. By this standard, every wedding invitation you have ever received is under suspicion, and so is the guest list.

Why the robot writes like your English teacher

Here is the part the accusers skip. An AI model like ChatGPT learns to write by studying huge amounts of human writing: books, articles, speeches and legal documents. It picks up the habits that show up again and again in good writing, then repeats them. The em dash appears in its writing because it appeared in ours first.

AI detectors run the same idea in reverse. Many of them score how predictable a piece of text is. Writing that follows familiar patterns scores as "AI". That is bad news for anyone who writes clearly, formally or in a well-known style.

My take on this? Every witch hunt needs a test the accused cannot pass, and the AI detector is this era's version.

The comic high point came in 2023. Screenshots shared online showed an AI detector called ZeroGPT rating the US Constitution as 92.15% AI-generated. The Constitution was written in 1787. Edward Tian created a rival detector called GPTZero. He told the tech site Ars Technica why it happens: the Constitution appears again and again in the text these AI models learn from. To a detector, the Constitution reads like machine writing, because the machines learned from it.

The joke stops being funny with the next finding. In 2023, researchers at Stanford University tested seven popular AI detectors on 91 essays. The essays were written for the TOEFL, an English test for people who speak English as a second language. The detectors flagged more than half of those essays as AI. Every one of them was written by a person. Essays by US school students were judged correctly almost every time.

Here is why that matters to you. The researchers linked the false flags to plain, predictable word choice, which is common when people write in their second language. So the detectors come down heavily on second-language writers. In a workplace, those are often the staff and freelancers with the least power to argue back.

Put the torches down: what to do this week

If you run a business, you judge writing all the time. It might be a freelancer's blog post, a new hire's email or an agency's proposal. Here is how to do it fairly.

Judge the thinking and ignore the punctuation. Weak AI copy gives itself away through its content. Look for vague claims, no specific examples, no figures you can check, and nothing only your business could say. A dash tells you nothing about any of that. Ask one question of every draft: could a competitor put their logo on this and publish it unchanged? If yes, send it back. If no, the dashes can stay.

Stop pasting people's work into AI detectors. The Stanford study showed these tools flag human writing, and second-language writers bear the brunt. Say you reject a freelancer or question a staff member because of a detector score. You are accusing a real person of cheating, on evidence that failed on the US Constitution. Talk to them instead. A writer who did the work can walk you through every choice they made.

Write a one-page house style guide. Decide once whether your business uses the Oxford comma, the em dash and lists of three. Write it down and give it to everyone who writes for you, human or machine. Then every punctuation choice in your copy is a decision you made, and nobody can call it a tell. A one-page guide is cheap. Ask Oakhurst Dairy what one missing comma cost.

Good writing was here first — and it is staying.