Why AI Text Is Full of Em Dashes, and Why That Alone Proves Nothing

Quick answer

The internet has spent two years arguing about em dashes. One camp treats them as a dead giveaway of AI-generated text. Another camp, fairly, points out that plenty of human writers have loved the em dash for decades. A March 2026 paper gives the first real mechanism behind the pattern instead of another anecdote: large language models learn a "structural orientation" from markdown-heavy training data (GitHub, Stack Overflow, technical docs), and when that structural impulse gets compressed into plain prose, the em dash is the one punctuation mark that survives, because it's the only dash-family character that's both a markdown structural marker and legitimate prose punctuation at once. Testing twelve models from five providers, the researcher found em-dash rates ranging from 0.0 to 9.10 per 1,000 words even after explicitly instructing every model to avoid markdown formatting, against a human baseline of 3.23. That's a real signal. It's also nowhere near proof on its own, and the same paper's own numbers show exactly why.

The mechanism: markdown training, then RLHF turns up the volume

The paper, "The Last Fingerprint: How Markdown Training Shapes LLM Prose" by independent researcher E. M. Freeburg (posted to arXiv in March 2026), starts from an observation that's easy to miss because it usually gets discussed in two separate places. Writers complain online that AI text overuses em dashes. Developers separately complain that AI models "think in markdown," defaulting to headers, bullet points, and bold text even when nobody asked for formatting. Freeburg's argument is that these are the same phenomenon at different resolutions. Training corpora like GitHub, Stack Overflow, and technical documentation are saturated with markdown, where dashes of every kind (horizontal rules, list markers, YAML delimiters) consistently mark a structural boundary. A model trained on that data doesn't just learn markdown syntax, it internalizes dashes as "where one unit of meaning ends and another begins."

Tell a model to stop using headers, bullets, and bold, and it mostly complies, because those are obviously markdown. But the em dash is prose-legal punctuation too, so an instruction to "write flowing prose, not markdown" passes right over it. That's the paper's central claim: the em dash is the smallest unit of markdown-trained structural thinking that survives compression into plain text.

Training data alone doesn't explain why the habit is so much stronger in some models than others, and that's where reinforcement learning from human feedback (RLHF) comes in. Freeburg argues RLHF amplifies whatever the base model already leans toward, because the human evaluators who rate model outputs during training tend to reward prose that reads as precise and well-organized, and em-dash-heavy writing reads that way. Sam Altman gave this framing some unintentional real-world support in November 2025, posting on X that telling ChatGPT not to use em dashes in custom instructions had "finally" started working, a small admission that the behavior needed deliberate fine-tuning to fix, not something an ordinary prompt could just talk it out of.

What the numbers actually show

Freeburg ran essay-writing tasks on ten everyday topics, twelve models, five providers, two conditions each time: write freely, or write in flowing prose with markdown explicitly banned. Headers, bullets, and bold text collapsed to near zero across the board once that instruction landed. Em dashes didn't cooperate the same way, and the model-by-model spread is where the paper gets interesting.

GPT-4.1 was the stubborn one. It opened at 10.62 em dashes per 1,000 words with no constraints, and suppression only knocked it down to 9.10, a 14 percent drop that barely counts as compliance. Claude Opus 4.6 went the opposite direction, falling from 9.09 to 0.19, a 98 percent collapse once told to write plain prose. DeepSeek V3 sat in the middle of that spread, moving from 6.95 to 5.41. Meta's Llama models, both the 8B and 70B instruction-tuned builds, never produced a single em dash across roughly 40,000 words in either condition, constrained or not. And the number all of this gets compared against, the human baseline, came out to 3.23 em dashes per 1,000 words across eight published essays (57,232 words total, spanning journalism, literary criticism, and technical writing), with individual writers ranging anywhere from 0.33 to 17.12 depending on genre and personal habit.

That spread matters more than any single figure in the table. Several of the tested models land inside or right next to the normal human range once suppressed. GPT-4.1 doesn't; it sits well above the human ceiling even after being told point-blank not to use markdown. Freeburg is careful not to overreach here: the claim isn't that AI writing uniformly outpaces human em-dash use, it's that suppression resistance varies by provider in a way that tracks how hard each company's fine-tuning actually leans on the habit.

One more data point makes the RLHF story concrete. Freeburg ran Llama 3.1 8B locally in both its raw base-model form (before any human-feedback tuning) and its instruction-tuned form. The base model produced a small but real em-dash tendency, 0.49 per 1,000 words. Meta's RLHF process drove that number to exactly zero. The underlying orientation from training data was there before any fine-tuning touched it; what a specific company's RLHF pipeline does with that orientation, amplify it or erase it, is what actually shows up in the deployed product.

Why the "it's just human writers" counterargument is also right

None of this makes the online "em dashes mean AI" shorthand correct, and it's worth taking the counterargument seriously instead of waving it off. Writer Brian Phillips made the case bluntly in The Ringer in August 2025: the em dash has been standard punctuation in professional writing for centuries, and if AI models use it a lot, that's downstream of the fact that the human writing they trained on used it a lot too, including the enormous troves of books and journalism scraped, licensed, or in some documented cases pirated to build those training sets.

That's not a rebuttal of Freeburg's mechanism, it's a different layer of the same explanation. The training-data saturation argument and the "human writers already loved this punctuation mark" argument aren't in conflict. They're both true at once: em dashes are common in the high-quality human prose that trained these models, and the models' post-training process then decides how much of that tendency survives into deployed output. What both sides agree on, and what the research confirms with real numbers instead of vibes, is that punctuation alone can't distinguish a careful human writer who loves a good dash from an AI model whose fine-tuning happened to preserve the habit. We've written about the broader family of AI writing tells readers rely on, and the same caveat applies to every one of them: real pattern, unreliable proof, especially in isolation.

What this means if you're actually trying to tell

If you've ever eyeballed a piece of writing and flagged it as AI because of the dashes, you were pattern-matching on something real. The mechanism in this paper explains why that pattern exists at all, and why it varies so much model to model. But treating any single stylistic tell as a verdict has the same failure mode we've covered before: a signal that correlates with AI generation on average will always misfire on the specific human writer who happens to sit at the high end of the normal range, and it will always let through the AI-generated text from whichever provider's fine-tuning suppressed the tell.

Statistical detection built on dozens of signals working together, not one visible punctuation quirk a reader can spot by eye, is the only version of this that scales past "well, my professor loves dashes too." If you've got a piece of writing you're actually trying to check, run it through TheChecker.AI's free demo and look at the sentence-level breakdown instead of counting dashes.

FAQ

Does using em dashes mean my writing will get flagged as AI? Not by itself, and it shouldn't. The human baseline in this study ranged up to 17.12 em dashes per 1,000 words for some writers and genres. A real detector weighs dozens of statistical signals together, not one visible punctuation mark, precisely because any single tell like this one produces both false positives on heavy-dash human writers and false negatives on models whose fine-tuning suppressed the habit.

Which AI models use em dashes the most? In Freeburg's test, GPT-4.1 was the heaviest user even after being told to avoid markdown formatting (9.10 per 1,000 words), followed by DeepSeek V3 (5.41). Anthropic's Claude models suppressed the habit almost completely when instructed. Meta's Llama models produced none at all, in either condition.

Can I get ChatGPT to stop using em dashes? Sam Altman said in November 2025 that custom instructions telling ChatGPT to avoid em dashes had started working reliably. Before that, users widely reported that even explicit prompts often failed to suppress the habit, consistent with the paper's finding that the tendency survives ordinary formatting instructions in several models.

Is the em dash actually connected to markdown formatting? Not syntactically. The em dash itself isn't a markdown element (CommonMark's dash-family syntax is horizontal rules, list markers, and YAML delimiters). The paper's claim is more specific: models trained on markdown-heavy data absorb dashes generally as structural boundary-markers, and because the em dash is the one member of that family that's also valid in ordinary prose, it's the one that slips through when a model is told to stop formatting and just write.