GPT-4o uses roughly 10x more em-dashes than GPT-3.5. Usage in scientific papers more than doubled between 2021-2025. Human writers use em-dashes like salt; AI uses them like rice. Model-dependent: ChatGPT overuses them, Claude uses few, Gemini uses none.
The Pattern
Good writers use em-dashes. But GPT-4o uses them the way a nervous cook uses garlic -- on everything, in everything, at rates roughly 10x higher than its predecessor GPT-3.5.
It's model-specific, which makes it unusual among AI tells. ChatGPT (GPT-4o above all) massively overuses them. Claude uses few. Gemini barely uses any. So heavy em-dash density doesn't just flag "AI." It flags ChatGPT.
The Washington Post profiled the phenomenon in April 2025, and the em-dash became the punctuation mark most publicly associated with AI writing. It also became a meme.
Density is the tell, not presence. Humans reach for an em-dash now and then. Three or four per paragraph, every paragraph, is a machine pattern.
Examples
The Research
The Washington Post ran "The em-dash is AI's favorite punctuation mark" in April 2025. Readers and editors had already been flagging em-dash density as a first-pass AI check; the article gave the practice a mainstream stamp. It was widely shared, partly because everyone had noticed the pattern and nobody had written the definitive piece on it yet.
Sean Goedecke dug into the technical cause. Tokenization treats em-dashes differently from other punctuation, and RLHF training rewarded their use as a sophistication signal. The model learned that em-dashes make text look polished. So it reaches for them constantly, the way a student who just learned the word "moreover" puts it in every paragraph.
Model comparisons confirm this is a ChatGPT problem, not a generic AI problem. Claude and Gemini don't exhibit the pattern. That makes em-dash density rare: it's not just an AI detection signal, it's a model attribution signal. Heavy em-dashes point at ChatGPT with unusual specificity.
Know Your Meme gave it an entry. Users started posting screenshots of em-dash-heavy text as proof of AI authorship, and the em-dash itself became shorthand for "ChatGPT wrote this." It's one of the few AI writing patterns that crossed over into mainstream internet humor -- people now insert gratuitous em-dashes into sentences as a joke.
Caught in the Wild
The April 2025 profile documented what readers and editors had been doing informally: flagging em-dash density as a first-pass AI check. The piece went wide. It gave a name to something everyone had already noticed but nobody had written up properly.
Washington Post →Originality.ai found 54% of long-form LinkedIn posts were AI-generated, with em-dash density as one detection signal. Those AI posts got 45% less engagement. Thought-leadership and personal branding posts were the worst offenders -- em-dash-heavy prose became a reliable marker for low-effort AI content on the platform.
Originality.ai →The tell got so well-known it earned a Know Your Meme page. Users post screenshots of em-dash-saturated text as evidence. The meme format itself is the joke: people insert absurd numbers of em-dashes into normal sentences to mock AI output. The punctuation mark got its own punchline.
Sources