Why does AI use em dashes? The tells of AI writing

Short answer: Chatbots overuse em dashes (—), bold phrases, tidy bullet lists and certain words because of what they learned from and how they were tuned. They were trained on lots of edited, published writing, where em dashes are common, and then rewarded by human raters for answers that looked clear and well organized. The result is a recognizable "house style". But none of these tells proves a text was written by AI, and AI detectors are unreliable enough that you shouldn't trust them on their own.

The em dash, explained

An em dash is the long dash used to set off a thought — like this. Professional writers, journalists and book editors use it a lot. Chatbots learned from huge amounts of that kind of polished text, so the em dash became one of their favorite tools for linking ideas.

There's a second reason. When companies fine-tune a model, people rate its answers, and the model learns what gets higher ratings. Sentences that flow, with an aside or a punchline after a dash, tend to read as fluent. Do that across millions of examples and the habit becomes strong.

Some chatbots now let you ask for no em dashes, and newer versions use them less, but the reputation has stuck.

Other common tells

Writers and editors who see a lot of AI text often point to the same habits:

• Bold words in the middle of sentences, even in casual answers.
• Headings and bullet points for everything, including a simple question.
• Groups of three: "fast, reliable and affordable".
• Signature words, for a while famously "delve", plus "tapestry", "testament", "navigate", "landscape", "robust".
• "It's not just X, it's Y" constructions.
• A neat closing summary that restates what was just said.
• Relentless positivity: "Great question!"
• Emojis as section markers in some apps.
• Hedging everything: "It's important to note that…"

Wikipedia's editors keep a long public guide to these patterns, because they spend a lot of time cleaning up AI-written articles.

Why the tells change over time

Each tell is a side effect of training. Once people notice one and complain, companies adjust the next model, and the tell fades while a new one appears. "Delve" became a joke in 2024 and then dropped off. That's why lists of "AI words" go out of date quickly.

Tells are clues, not proof

Every one of these habits came from humans first. Plenty of people love em dashes, bullet points and the word "robust". Copywriters, academics and people who learned English from textbooks often write in a tidy, formal way that looks "AI-like".

So a text with three em dashes is not evidence of anything. A text that is generic, confident, oddly balanced, full of the habits above and doesn't contain any specific detail only the writer would know is more suspicious. Even then, it's a guess.

Why AI detectors get it wrong

Tools that promise to detect AI writing are much less accurate than they sound.

• In 2023, OpenAI shut down its own AI text classifier because of its low accuracy. At launch it correctly flagged only about a quarter of AI-written text and sometimes labeled human writing as AI.
• A Stanford study the same year found that popular detectors flagged essays by non-native English speakers as AI-written far more often than essays by native speakers, because simpler, more predictable wording looks "machine-like" to them.
• Light editing or asking the model to change its style is often enough to fool a detector.

If you're a student or employee accused because of a detector score, ask what else the accusation is based on, and offer your drafts, notes or version history. Google Docs and Word keep an edit history that shows how a text was written.

How to make AI writing sound like you

If you use AI to help you write, a few habits make the result less generic:

• Give it your own notes and examples, not just a topic.
• Ask for plain style: "no em dashes, no bold, no bullet points, short sentences".
• Paste a sample of your own writing and ask it to match the tone.
• Edit the result. Cut the opening and closing summaries, add one real detail, change words you'd never use.
• Be open about it where it matters, such as school or work, and follow the local rules.

The best defense against sounding like a chatbot is to say something only you could say.

Based on published research and product documentation as of October 2026.