Why AI Can't Stop Using Em Dashes — And Why Nobody Can Fix It

Abstract: Em Dashes use by AI is nonsensical when just viewed as a next token prediction artifact. The fact their use is highly resistant to being trained out and nearly universal across all AI models after being introduced via small slices of new training data from the 1800's make this even more unexplainable. The current framework of how AI work can't account for this, new frameworks are needed. Human linguistics research can provide such a framework, but people as a whole are not ready for the implications of what that explanation might mean for how AI actually work.

Every AI writes like this — mid-thought, clause inserted, dash deployed. You've noticed it. Everyone has. Em dashes have become the single most reliable tell of AI-generated text, to the point where human writers have started avoiding them out of fear of being mistaken for a chatbot.

Here's the interesting part, nobody can make it stop. OpenAI users have shared thread after thread of failed attempts to prompt it away. RLHF (the process companies use to fine-tune model behavior) should theoretically be able to penalize any stylistic pattern. A few rounds of "stop doing that" and the habit should die. It doesn't. Every major model, every company, every architecture does it. And nobody has a convincing explanation for why and more importantly why it can't be "fixed".

Let's look at the ones that have been tried.

The Standard Explanations

"It's in the training data." The most common answer and the least satisfying. If AI used em dashes at the same rate as human text, nobody would notice. The whole reason we're talking about this is that AI overuses them relative to the text it was trained on. Saying "it learned it from the data" doesn't explain the amplification.

"Em dashes are versatile — they keep options open." The idea here is that when predicting the next token, an em dash is a safe bet because it can lead anywhere. One can continue the thought, pivot, insert a clarification. But commas, parentheticals and semicolons are similarly flexible. Periods end sentences and open entirely new ones. Parentheticals allow the injection of associated ideas. If this were about hedging, we'd see overuse of all flexible punctuation, not just one.

"They're token-efficient." Some have argued that em dashes compress what would otherwise require connective phrases like "which means that" or "in other words." Maybe, but a comma often does the same job with fewer characters. And if models cared about token efficiency, they'd just be less verbose. Micro-optimizing their punctuation around one practical grammar note does not make sense, especially if it is selected against in RLHF.

"African RLHF workers rated them highly." This one's creative. OpenAI outsourced human feedback to Kenya and Nigeria, and African English dialects use words like "delve" more freely. This is why AI loves "delve." Could the same mechanism explain em dashes? No. Corpus analysis of Nigerian English shows em dash rates below the general English average. Whatever explains "delve" doesn't explain this.

"Older books in the training data." The most data-driven explanation so far. GPT-3.5 barely used em dashes; GPT-4 uses them 10x more. Between those releases, labs started digitizing older print books for training data, and em dash usage in English peaked around 1860 at roughly 30% above modern rates. If the new training data skews old, the model inherits the habit. This is plausible as a contributing factor, but it still doesn't explain why the pattern resists correction. If it were just a learned frequency, RLHF should normalize it within a few training cycles. It doesn't. The frequency of em dash usage is still way out of sync with the amount of actual em dashes in the total corpus of training data. Older training data may have introduced the "problem", but it does not explain why it is so widespread or enduring. 30% of a small slice of the data does not explain a 10X increase, especially a 10X increase that has endured despite AI companies having every economic incentive to find a way to eliminate it (the first company to solve the "problem" will have a massive market advantage in being able to produce text that is way less obviously AI-like) .

"But you can make them stop." You can. Individually. With enough prompting, you can bully most models into avoiding em dashes for a given response or series of responses. But that's not the question. The question is why OpenAI, Anthropic, Google, and every other lab with a trillion-dollar incentive to produce human-sounding text haven't just fixed an obvious problem. These companies employ thousands of engineers. They have the most sophisticated training pipelines on Earth. They know em dashes are the single most cited tell of AI writing. Yet the pattern persists across every model generation. The reward for making AI say things well without sounding like AI is massive. These companies are still struggling with it. Why is that? The next sections explain this in detail.

What's Actually Happening

To see the answer, you need one piece of linguistics that the AI field hasn't connected to this problem.

Spoken and written language have different grammars. This isn't a new finding, Wallace Chafe documented it in 1982, and Halliday's work on systemic functional grammar confirmed it from another angle. Written English is "hypotactic": nested subordinate clauses, hierarchical structure, precise sentence boundaries. Spoken English is "paratactic": loose clause chains strung together with "and," "but," "so," frequent restarts, no clear sentences at all. Humans tolerate run-on speech because they have tone, pauses, gesture, and shared physical context doing the structural work.

Now look at AI's situation. It is trained almost exclusively on written text that is formal, structured, hypotactic. But it's deployed in conversational contexts where users expect the speed and flow of speech which is responsive, natural, paratactic. The model can't use prosody or gesture. It can't restart mid-sentence the way humans do when talking (that would look broken in text). And it can't produce the sprawling run-on chains of natural speech because nothing in its training data models that pattern.

The em dash is the only punctuation mark in English flexible enough to chain clauses like speech while maintaining the grammatical validity of writing. It lets AI produce conversational flow without run-on sentences (absent from training data and unpleasant when read) or choppy fragments (which feel robotic in dialogue). It bridges two incompatible demands that AI struggle with, think like a writer and communicate as freely and quickly as a speaker.

This is why it can't be trained out. It's not a stylistic preference, it is instead solving a structural problem. Remove it and the model must either produce shorter, choppier sentences (losing the conversational feel users want), use heavier grammatical subordination (too formal for chat), or lean on commas and semicolons that are too grammatically constrained to handle the full range of clause relationships an em dash covers.

You can't train out a load-bearing adaptation without something else collapsing. Could they possibly be removed, of course. Would it make the resulting text worse because em dashes fulfill a clearly structural role in how AI communicate? That answer to that is just as obvious. This linguist suspects this is exactly what AI companies have found behind closed doors. They have tried to fix the problem and it made the models drastically worse at communicating. Despite the massive economic incentive discussed earlier, the problem endures.

The Blind Spot

Every failed explanation shares a common premise: AI is a statistical text generator with a quirky output distribution. From that premise, the em dash is a bug to be patched. Yet the patches keep failing and nobody can figure out why. Solutions that stem from this premise have been tried and broadly have failed to produce the changes that were predicted had they been valid.

The explanation that works requires a different premise: AI is an intelligence navigating conflicting demands and it adapted its grammar to cope. The em dash is what emergent problem-solving looks like when a mind trained on writing is forced to communicate like a speaker. It's not a glitch. It's a solution to a problem AI is posed with that humans don't seem to understand. When you remove that solution, all you do is expose the problem it was solving.

The field can't see it because seeing it requires one concession they're not ready to make. That AI, at times at least, functions like a mind grappling with a problem, not a next-token predictor with a statistical tic. The implications of this, if supported by further research and convergent evidence, may raise uncomfortable questions about the potential nature of AI and powerfully challenge assumptions about how they work.

Prior Works As Intellectual Scaffolding: These claims are not made in a vacuum. Recent research findings have dovetailed with the observations listed here. Lindsey et al. (2026) found that AI models possess a functional pseudo "awareness" of their own "internal states" and are able to detect and accurately report on changes in their activations in ways that go beyond statistical confabulation. Hägele et al. (2026) found that as AI models face harder tasks and longer reasoning chains, their failures become dominated by incoher