Has AI ruined the em dash? Why LLMs overuse professional writers' favorite punctuation
Nov 21, 2025
Key Points
- Large language models overuse em dashes because they're trained on professional writing, where the punctuation appears 4.4 times more frequently than in average text, not because of AI hallucination.
- A sharper tell for AI writing is antithetical parallelism, the "it's not this, it's that" construction that appears identically across OpenAI, Google, and Anthropic outputs.
- Social media has weaponized the em dash as proof of AI authorship despite zero correlation with actual machine generation, forcing even OpenAI's CEO to disable it via custom instructions.
Summary
The em dash has become a tell for AI-written text. People now shame it as a GPT cliché, but Joel Stein argues in The Wall Street Journal that the real culprit isn't the punctuation mark itself but where language models learned to write.
The panic started when a high school journalist's source claimed a school administrator's email was AI-generated based purely on em dash usage. AI detection tools returned a 0% confidence score that a computer wrote it, yet the source remained convinced because of the punctuation. On Instagram, beauty influencer Lux Jen warned followers to remove em dashes if they didn't want to be accused of using AI. OpenAI's CEO posted a small victory: telling ChatGPT not to use em dashes in custom instructions finally makes it comply.
Stein digs deeper. Large language models train on professional writing. OpenAI has a deal with The Wall Street Journal. Professional writers use em dashes 4.4 times more frequently than average people. The em dash isn't a GPT hallucination but a learned pattern from good writing. Emily Dickinson, Virginia Woolf, and David Foster Wallace all relied on the em dash as a mark of nuance and digression, the kind of human breath marks that make prose feel alive rather than mechanical.
Stein calls the em dash a declaration of humanity against AI's onslaught but adds a practical caveat. Most people can't find one on their keyboard and wouldn't know when to use it if they did. The punctuation mark requires more nuance than a thumbs up or heart emoji. It's less jarring than parentheses but a bigger interruption than a comma.
A sharper tell for AI writing emerged during the discussion: antithetical parallelism, the "it's not this, it's that" construction that appears across models and platforms. One host tested Google's Gemini 3 Pro with a Porsche GT3 RS ad prompt and received almost identical phrasing: "The noise isn't just sound, it's frequency. The vibration isn't just an engine, it's a heartbeat." This construction shows up in OpenAI, Google, and Anthropic outputs alike, suggesting models have converged on what reads as persuasive rather than learning from each other directly.
The practical puzzle remains: why has this specific rhetorical pattern, distinct from the em dash issue, become baked into multiple model families as a default cadence for "good" writing?