Em dashes don't prove AI wrote it

Em dashes don't prove AI writing. People have used the long dash for centuries, and newer AI models use fewer of them. Anyone can delete every one in a document in a few seconds. We learned this the hard way: an older version of this site leaned on em dashes, and it took the first chapter of Moby-Dick for AI writing.

Why em dashes became the best-known AI tell

There was something real behind the rumour. Language models picked up habits in training that showed up far more often than in human writing, and the dash was one of them. The authors of StoryScope, a study of AI fiction published at COLM 2026, open their paper by noting that AI-generated text "systematically overuses em-dashes", along with a handful of stock words and other surface patterns. They also note that cues like these have been strong enough for simple classifiers to pick up.

Word choice showed the same kind of skew. In a 2025 PNAS study, Alex Reinhart and colleagues compared GPT-4o with human writers responding to the same prompts. GPT-4o used "camaraderie" about 162 times as often as the humans, and "palpable" about 95 times as often. Findings like these made it feel as if AI writing came with a fingerprint you could read by eye.

The em dash spread fastest of all these tells for a simple reason: you can spot one without any software. That convenience is also its weakness.

Surface tells fade quickly

The same StoryScope introduction goes on to argue that these stylistic cues are fading. It reports that GPT-5.4 significantly reduced its use of em dashes, citing a public post on X. A habit that was common in one release can be trained out of the next.

Deliberate training can remove far more than punctuation: a 2026 study by Tuhin Chakrabarty, Jane Ginsburg and Paramveer Dhillon found that a model fine-tuned on individual authors' books slipped past leading detectors almost every time. We look at that study, and at what it means for detecting new releases, in can AI detectors catch the newest models?

Editing after the fact is even easier. Find-and-replace removes every dash in a document. Rewriting tools sold as "humanizers" exist to change exactly this kind of surface pattern, and the StoryScope authors note that detectors built on style are easily thrown by paraphrasing and light edits. Their conclusion is that surface signals tend to disappear, either in the next model release or after a quick edit, which is why they turned to the structure of stories instead.

Vocabulary moves in fashions too. Dmitry Kobak and colleagues tracked more than 15 million biomedical abstracts and found that in 2024 "underscores" appeared about 14 times more often than earlier years would predict, and "showcasing" about 11 times more often. A word like that is a real signal for a year or two, and then models, editors and writers move on.

What happened when this site counted em dashes

The first version of IsThisAiWritten scored a set of hand-picked style rules. When we audited it, it turned out to be, in effect, an em dash counter. Enough dashes, on their own, could push a text's score into AI territory. We replaced it.

The clearest casualty was Herman Melville. Chapter 1 of Moby-Dick, "Loomings", published in 1851, runs to about 2,200 words in the Project Gutenberg text and contains 27 em dashes. The novel's second sentence has two of them, wrapped around "never mind how long precisely". The old detector scored the chapter as AI.

It failed in the other direction as well. Modern AI writing often has few dashes, so it mostly didn't trip a dash counter. Nineteenth-century prose did.

So a rule built on the best-known AI tell managed to accuse a long-dead novelist while letting through machine-written text that happened to go easy on dashes. That is the strongest argument we know against judging writing by its punctuation.

Who an em dash rule hurts

Dashes are a favourite of writers who care about rhythm, from novelists and essayists to journalists and editors. Emily Dickinson's poems are full of them. Many people picked up the habit from the books they grew up reading, long before chatbots existed.

A rule that treats dashes as suspicious puts those writers under suspicion for a stylistic choice. Meanwhile someone who pastes in AI text and strips the dashes passes straight through. The rule punishes care and rewards a two-second edit.

The same logic applies to other single tells, from a fondness for semicolons to a habit of tidy closing paragraphs. Each one shifts the odds a little at most. None of them settles anything about who wrote a text.

How we use style signals now

The current detector makes its verdict with a trained classifier, TMR, an open RoBERTa-based model that runs in your browser. It reads the whole text in overlapping windows rather than counting features. Our report still shows style signals, such as em dashes and stock transitions, because people find them useful context, but they are explanations only and are not used in the verdict.

Like any detector, it can be wrong in both directions, so treat a result as a reason to look closer, not as proof. Our methodology page explains how it is tested, and our explainer on how AI detectors work shows how the verdict is made.

Hidden characters are a separate issue from dashes. If you want to check a text for zero-width characters or look-alike letters, our invisible character remover finds and cleans them.

Better questions than counting dashes

If you are trying to judge whether something was written with AI, these questions will tell you more than any punctuation mark:

• Is there a draft history, such as earlier versions, notes or tracked changes?
• Does the piece sound like the writer's other work, including work written in class or on the spot?
• For fiction, how is the story built? AI stories tend to state their themes and resolve neatly, as we explain in how AI fiction differs.
• What does a detector that scores the whole text say, and what are its published error rates?

Even taken together, these are evidence to weigh, not proof. Our guide to signs of AI writing covers the research on which patterns hold up and which don't.

Sources

1. Russell, Rajendhran, Pham, Iyyer and Wieting, StoryScope: Investigating idiosyncrasies in AI fiction (COLM 2026, arXiv 2604.03136), checked October 2026
2. Chakrabarty, Ginsburg and Dhillon, Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers (arXiv 2510.13939), checked October 2026
3. Reinhart et al., Do LLMs write like humans? Variation in grammatical and rhetorical styles (PNAS, 2025), checked October 2026
4. Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary (Science Advances, 2025), checked October 2026
5. Project Gutenberg: Moby Dick; Or, The Whale by Herman Melville (eBook #2701), checked October 2026