E. M. Freeburg’s March preprint measured em dash frequency across twelve instruction-tuned models from five providers (~240,000 words) against a human baseline of 3.23 per 1,000 words. GPT-4.1 hit 10.62 unconstrained, the 3.28x figure now circulating. The result that actually matters is suppression resistance: told outright “do not use em dashes”, GPT-4.1 still produced 3.86 per 1,000 words, while Claude Opus 4.6 went to zero and Meta’s Llama models never produced any at all. The em dash is a fingerprint of a lab’s post-training pipeline. Treat it as a weak prior and never as evidence: human writers in the same study ranged up to 17.12 per 1,000.
Windows: Alt+0151. macOS: Option+Shift+Hyphen. Unicode: U+2014.
Every one of those is a small deliberate act. You look it up once, and then you choose to remember it. The comma is a key; the em dash is a ritual, the one punctuation mark with an onboarding cost. Which makes it strange that the most prolific producer of em dashes on Earth has never pressed a key in its life.
The number, and where it came from
The figure people quote comes from “The Last Fingerprint: How Markdown Training Shapes LLM Prose”, submitted to arXiv on 27 March by E. M. Freeburg, an independent researcher. Twelve instruction-tuned models from Anthropic, OpenAI, Meta, Google and DeepSeek, ten essay topics each, roughly 240,000 words of generated prose. The human baseline came from eight published essays spanning literary criticism, journalism and technical writing: 57,232 words, 3.23 em dashes per 1,000.
GPT-4.1 unconstrained produced 10.62 per 1,000 words. Divide, and you get the 3.28x that McGill’s Office for Science and Society repeated in May.
Two caveats a technical reader deserves up front. This is one preprint, not peer reviewed, and the human baseline is eight essays. And that baseline has a range of 0.33 to 17.12 per 1,000 words, meaning some perfectly human essayists out-dash GPT-4.1 by a comfortable margin. Freeburg says so explicitly: the finding is not that LLMs uniformly overuse the mark relative to all human writing.
The result that matters is what happens when you say stop
Condition B of the experiment instructed each model: no markdown formatting, headers, bullet points, bold text, or lists. It worked. Overt markdown dropped to zero across all twelve models, from every provider, at every capability tier. Only Claude Haiku 3.5 retained a residual, 0.9 features per 1,000 words.
The em dashes stayed. GPT-4.1 went from 10.62 to 9.10, a 14% reduction. DeepSeek V3 went 6.95 to 5.41.
So Freeburg added Condition C: same instruction, plus “do not use em dashes.” GPT-4.1 still produced 3.86 per 1,000 words. Pushed to 5,000-word outputs, it retained 6.97 per 1,000 under direct prohibition — twice the human mean, while being told not to.
Freeburg’s mechanism is the good part of the paper. In markdown, dashes of every kind signal structural boundaries: thematic breaks, list markers, YAML delimiters. The em dash is the one member of that family that is also legal prose punctuation. So an instruction to write prose rather than markdown sails straight past it, because it does not look like formatting.
The em dash is the smallest possible unit of markdown thinking that survives compression into plain prose.
You need not buy the full genealogy. Sean Goedecke argued last October for a different cause: digitised late-1800s print books, where em dash rates ran ~30% higher than contemporary English. Either way, something in these models treats the mark as structure, and formatting instructions do not reach it.
It’s a fingerprint of the lab, not of the machine
Here is where the “AI uses em dashes” folk wisdom falls apart.
Meta’s Llama 3.1 8B Instruct and Llama 3.3 70B Instruct produced zero em dashes across roughly 40,000 words, in both conditions. Freeburg checked that they weren’t substituting double hyphens or en dashes; they simply don’t make dash-mediated clause transitions. Claude Opus 4.6 started at 9.09 unconstrained and dropped to 0.19 under markdown suppression, a 98% reduction, then to zero under explicit prohibition. Gemini 2.5 Pro went 3.53 to 0.00. GPT-5.4, OpenAI’s newest, produced 1.43 unconstrained.
Same markdown-soaked training corpora, wildly different outcomes. The base-vs-instruct comparison isolates why: Llama 3.1 8B in base form produced 0.49 per 1,000 words, and Meta’s RLHF drove that to zero. The latent tendency exists before post-training; the fine-tune decides its amplitude. Sam Altman has publicly acknowledged that ChatGPT’s em dash frequency was adjusted in response to user preference, which confirms the lever exists and that someone at OpenAI pulled it in the other direction.
Models raised on markdown default to structured documents. That is the same instinct behind the whole ecosystem of all-caps .md routing files agents now read, which I wrote about a few days ago. The em dash is that instinct leaking one character at a time into prose.
The cost lands on humans
Meanwhile the tell is being used, badly, by everyone.
In February, Viktor Löfgren of Marginalia scraped HN’s /newcomments and /noobcomments feeds and found that comments from newly registered accounts contained em dashes, arrows and similar symbols 17.47% of the time against 1.83% for established accounts, across about 700 comments each (p = 7e-20). It hit 717 points and 598 comments on Hacker News. Freeburg’s actual paper, posted a month later, got one point. The vibe check outdrew the measurement by a factor of 700.
As a detector feature, the em dash sits in a vector alongside perplexity and burstiness — and that vector already has a known failure mode. Liang et al. ran 91 TOEFL essays by non-native English speakers through seven GPT detectors: 19% were unanimously flagged as AI-generated, and 97% were flagged by at least one. Add a punctuation mark that correlates with formal register and you make that worse, not better.
The adaptation is already happening, and it’s pure deadweight loss. A McGill student contributor described cutting long sentences, semicolons and dashes from her own writing to avoid being flagged. One HN commenter turned off macOS’s automatic double-hyphen conversion and went back to typing --, calling it “even less correct than before, but at least people” won’t assume a bot. Another pointed out mid-thread that it’s Option+Shift+Hyphen for the em dash and plain Option+Hyphen for the en dash, which is exactly the kind of thing you know only if you cared enough to learn it.
Use the mark as a prior if you like. Never use it as evidence. And note the trajectory: GPT-4.1 at 10.62, GPT-5.4 at 1.43. Once a tell gets named it gets tuned, and the em dash panic will comfortably outlive the evidence for it.
Sources
- E. M. Freeburg, “The Last Fingerprint: How Markdown Training Shapes LLM Prose”, arXiv:2603.27006, 27 March 2026 (code and data)
- Sean Goedecke, “Why do AI models use so many em-dashes?”, 30 October 2025
- Viktor Löfgren, “New accounts on HN 10x more likely to use EM-dashes”, Marginalia, 25 February 2026
- Lia Erisson, “Why Did LLMs Steal Our Em-Dashes?”, McGill Office for Science and Society, 8 May 2026
- Liang et al., “GPT detectors are biased against non-native English writers”, Patterns, 2023 (Stanford HAI summary)
Coverage
- New accounts on HN more likely to use em-dashes — Hacker News, 717 points
- The Last Fingerprint — Hacker News, 1 point
Related on this blog