The Peculiar Persistence of the Em Dash in AI Writing

In the intricate world of large language models (LLMs), the humble em dash has emerged as an unexpected subject of scrutiny. Seen as a hallmark of AI-generated prose, its presence, or absence, can tell us more than we might initially presume about the underlying processes that guide these models.

The Role of Training Data

Let's start with the foundation: training data. Models from OpenAI, Anthropic, Meta, Google, and DeepSeek ingest an abundance of markdown-formatted text. This saturation leaves an indelible mark, or rather a dash, on the prose they generate. Em dashes represent a vestige of markdown's structural orientation within these models, it's the residue that surfaces even when other markdown features, like headers or bullet points, are suppressed.

To enjoy AI, you'll have to enjoy failure too, and in this context, failure to completely suppress the em dash speaks volumes. Consider Meta's Llama models, which interestingly produce none at all, starkly contrasting with the GPT-4.1 model's 9.1 em dashes per 1,000 words, even under suppression. What differentiates these models? It's not just a stylistic choice but a reflection of the fine-tuning methodologies applied to them.

Fine-Tuning and Its Discontents

Fine-tuning is where the plot thickens. The better analogy is that LLMs are like musicians who, despite being trained in classical music, can't help but let a little jazz slip into their performances. The em dash is the jazz riff in this scenario. It's the feedback loop that resists suppression, revealing the structural quirks imbued during training.

Meta's models, for instance, manage to avoid this entirely, suggesting a unique fine-tuning process. In contrast, other LLMs, even when given explicit instructions to avoid markdown, can't quite kick their em dash habit. This discrepancy reframes the em dash not as a stylistic defect but as an indicator of the unique paths these different models have traveled.

Why Does This Matter?

Does the frequency of em dashes really matter? Pull the lens back far enough, and the pattern emerges: this quirk offers valuable insights into the structural methodologies behind AI training. It bridges previously isolated discussions about markdown formatting and AI text generation, connecting the dots between what we train these models with and how they express themselves.

In a rapidly evolving field, understanding these nuances is important. It's a reminder that the mechanics of AI are as much about what they unconsciously retain as what they consciously produce. As we advance, the question isn't merely about eliminating the em dash but about what other latent tendencies we might uncover in these sprawling neural networks. The proof of concept is the survival of these quirks, which continue to challenge our expectations and push the boundaries of AI understanding.