We’ve all been playing the game. You read a corporate email, a product description, or a slightly-too-helpful forum post and you get that feeling. Your eyes scan for the clues. The word “delve”? Check. An obsession with “leveraging”? Check.

And the biggest, most obvious red flag of all? The em dash.

That long, elegant, and—in the hands of AI—wildly overused punctuation mark. For months, the em dash has been the digital equivalent of a fake mustache, a dead giveaway that you’re not talking to a human.

Well, that game just got a little harder.

You may now have to scrutinize what you read on the internet (and even in print) more closely to determine if it’s the product of AI. In a post on X (formerly Twitter), OpenAI CEO Sam Altman revealed that if you tell ChatGPT not to use em dashes in your custom instructions, the chatbot will now finally listen to you.

Previously, ChatGPT would notoriously ignore this specific request. You could beg, plead, and command it—”Do not use em dashes”—and it would respond with something like, “Of course, I will happily adhere to your request—it’s important to respect stylistic preferences.” The rebellious punctuation was a constant, frustrating tic.

People are treating the presence of em dashes, especially if they’re used in abundance, as one of the biggest “tells” if something was written by a large language model (LLM). Of course, just because a piece of text uses em dashes doesn’t mean it was actually written by AI—plenty of human writers (this one included) love them. But people have become justifiably suspicious of any writing that uses the punctuation mark with robotic frequency.

Altman, for his part, acknowledged how difficult this “fix” actually was, stating on X that getting the AI to stop its dash-happy habit “was a surprisingly tricky problem.”

Why the AI obsession with the dash?

It’s not entirely clear why generative AI models have such a tendency to pepper the text they generate with em dashes. The “why” is a fascinating window into how these models “think.”

LLMs are statistical models, essentially the world’s most sophisticated autocomplete. They are trained on a vast, mind-bogglingly large number of books and online content, such as scientific papers, technical manuals, posts on public forums, and news articles.

Here’s the prevailing theory:

1. It’s a “safe” choice: The em dash is a grammatical multitool. It can replace commas for an aside, parentheses for a clarification, or a colon for an explanation. For an AI predicting the next most-likely word (or “token”), the em dash is often a high-probability, statistically “safe” bet to connect two related ideas.
2. It reflects the training data: Formal writing, academic papers, and 19th-century literature—all key parts of an LLM’s “diet”—use the em dash heavily. The AI is simply mirroring the more formal and complex parts of its training data.
3. No one told it to stop: It’s possible that during the human-feedback phase of training, the punctuation mark simply wasn’t flagged by AI trainers as something the LLM should avoid. Why would it be? It’s a perfectly valid (if overused) grammatical tool.

The cat-and-mouse game of AI detection

This update is more than just a punctuation patch. It’s a significant move in the escalating cat-and-mouse game between AI developers and the public trying to spot AI in the wild.

The em dash was just one of many “tells.” There’s a whole list of AI-favored “corporate-speak” words that set off alarms, including:

• Delve
• Leverage
• Tapestry
• Beacon
• Navigate (as in, “navigate the complex landscape of…”)
• In conclusion…
• It is important to note…
• Furthermore…

As soon as the public identifies these statistical tics, AI companies work to “fix” them—not because they are wrong, but because they make the AI sound less human.

With this update, OpenAI has effectively removed one of the most reliable “tells” from its toolkit, at least for users savvy enough to use the “Custom Instructions” feature. This feature allows users to set persistent rules and preferences for how ChatGPT responds, such as “Always reply in a casual tone” or “Never lecture me” and, now, “Don’t use em dashes.”

As the models get “smarter” and their rough edges are sanded down, distinguishing human-written text from AI-generated text will become less about spotting clumsy punctuation and more about analyzing the originality, intent, and subtle flaws behind the words themselves.

For now, the em dash—once the scarlet letter of AI writing—can finally be put to rest.