Why LLMs love the em dash

Large language models do not love the em dash. They inherited a dash-heavy habit from older print books, and later training made that habit cheap to keep.

People now treat the long dash as an AI fingerprint. Some human writers have stopped using it so they do not get accused of pasting a chatbot. The mark still does real work in English. The panic around it does not.

A hyphen and an em dash do different jobs. A hyphen joins words. user-centric is a hyphen doing hyphen work. The long dash interrupts a sentence, inserts an aside, or replaces a colon or a semicolon. Models reach for that interruptor because it is a general-purpose joint. A single hyphen cannot do the same work without looking like a typo. If I write "the deploy failed - we rolled back", the ASCII hyphen is standing in. Readers forgive it on a site that banned the real mark. In a novel, that same glyph looks broken.

I do not think we have a clean causal story. Sean Goedecke says we still do not really know. What follows is a stack of theories.

Inherited from books, then rewarded

Maria Sukhareva generated stories on GPT-3.5-turbo against GPT-4o and GPT-4.1. Later models used about 10x more em dashes. Every GPT-4o and GPT-4.1 story in her set had at least one. GPT-3.5 sometimes produced a story with none. That is a usage shift inside one lab's line. It is still the best before-and-after we have.

She and Goedecke both point at older print books entering the mix. Dash use in English peaked around 1860. Goedecke is citing a punctuation-frequency paper for that peak, not a number he invented. Anthropic's book-scanning showed up in court filings around February 2024. Goedecke thinks newly digitized older books are the strongest explanation. He also notes that Moby-Dick has 1728 em dashes. Writers in that pile like to stop a sentence in the middle and keep talking. If those books are back in the mix, the long dash becomes a normal next token again.

The second theory is versatility at next-token time. One mark can stand in for a parenthetical, a colon, or a punchy comma. That is a linguistic fact. It is not proof that the model wants anything. Goedecke is skeptical that "playing it safe" is a good idiom for sampling. I share that skepticism. A sampler is not being careful with your feelings. It is emitting a likely piece.

The third theory is tokens and RLHF. Sukhareva argues that a leading space plus an em dash is often one token, while ", and" costs more. She also points at PPO's per-token KL penalty from Ouyang et al. 2022, which favors shorter completions. When she asked models to rewrite dash sentences without the dash, and filtered the set heavily, token counts rose about 1.74% on GPT-3.5 and about 2.23% on GPT-4o. A cheap joint that also reads like edited prose is easy to leave in.

Goedecke does not buy this as the main cause. Many of those dashes could have been a comma. If the model wanted fewer tokens, it could waffle less. I am not picking a winner here. The token account is tidy. The objection is also tidy. Nobody held the books fixed and trained two otherwise identical models.

The fourth theory is the RLHF dialect story, and it is weak for this mark. Goedecke measured Nigerian English at 0.022% em dashes, against roughly 0.25% to 0.275% in general English corpora. That gap may help explain why models lean on delve. Juzek and Ward made that case. It does not explain dashes. If contractor dialect were driving this mark, you would expect the mark to be rare in the dialect and then rare in the model. The model does the opposite.

A loud, bad AI detector

Wikipedia's Signs of AI writing says an em dash alone is a bad detector. I think that page is right. A mark this common in older books will show up in model text. It will also show up in humans who read those books.

The Washington Post analyzed ChatGPT style around this tell in November 2025. Sam Altman publicly celebrated ChatGPT starting to follow em-dash formatting rules that same month. Once the lab is talking about the formatting rule in public, the tell is no longer a private nerd observation.

The tell is now so loud that people who used the mark for years get accused. I banned it on this site anyway. The rule lives in src/design/system.ts on the design system page. Never an em dash. A hyphen with spaces when a dash is needed. React, TypeScript, Node - shipping user-centric apps. The line is ugly on purpose. A reader can still parse it, and a detector that only counts long dashes cannot.

If you want to sound like a person, pick a period. Training data and a short token made the long dash easy to emit. We put the habit in the mix, then started hunting people who still write that way.