Researchers discover that LLMs' overuse of em dashes stems from markdown formatting in training data, not inherent model design.

Em dashes in AI-generated text are identified as 'markdown leaking into prose' — the smallest surviving unit of structural patterns learned from markdown-heavy training corpora

Study tested suppression across 12 models from 5 major providers (Anthropic, OpenAI, Meta, Google, DeepSeek) to understand the em dash phenomenon

Two-condition experiment showed that when models are instructed to avoid markdown formatting, overt features like headers, bullets, and bold text are eliminated

The research provides the first mechanistic explanation for why some LLMs overuse em dashes, a widely recognized marker of AI-generated content