If you have generated essays, blog posts, or marketing copy with ChatGPT recently, you have almost certainly noticed an unmistakable pattern: an epidemic of em-dashes (—).
Whether it is GPT-4o, o1, or Claude 3.5 Sonnet, modern large language models have developed an almost compulsive addiction to interjecting thoughts—often two or three times in a single paragraph—using the long em-dash. To human readers, editors, and hiring managers, this has become one of the most obvious tells that a piece of writing was generated by artificial intelligence.
Why Does ChatGPT Overuse Em-Dashes?
The reason ChatGPT defaults to em-dashes comes down to three technical factors in how transformer architectures are trained:
- Reinforcement Learning from Human Feedback (RLHF): During human evaluation, annotators favored responses that felt authoritative and grammatically sophisticated. High-end journalism (like The Atlantic or The New York Times) makes frequent stylistic use of em-dashes, causing the reward model to associate em-dashes with high-quality prose.
- Token Transition Probabilities: Once an LLM starts an explanatory clause, the em-dash token has a statistically high probability of bridging the parenthetical thought without requiring complex sentence restructuring.
- Lack of True Fatigue: Human writers consciously recognize when they have used a stylistic device too many times. An autoregressive language model only predicts the next token based on statistical context, meaning it will happily deploy four em-dashes in consecutive sentences.
Method 1: Stop Em-Dashes via Custom Instructions (The Set-and-Forget Fix)
The most reliable way to curb em-dashes across all your ChatGPT chats is by adding a negative constraint to your Custom Instructions.
Avoid using em-dashes (—) or en-dashes (–) to connect clauses. Use standard periods, commas, or semicolons instead. Keep sentence structures varied and direct. Never use more than one dash per 1,000 words.
How to apply: In ChatGPT, click your profile icon in the bottom-left corner → select Customize ChatGPT → paste the snippet above into the box labeled "How would you like ChatGPT to respond?"
Method 2: In-Chat Prompt Modifiers
If you are using ChatGPT without Custom Instructions or using the API directly, append this stylistic modifier to the end of your generation prompt:
Notice that positive instructions ("Break long thoughts into two separate sentences") consistently perform better than purely negative constraints ("Do not use dashes") because language models struggle with absolute negation.
Why Prompts Still Fail (Attention Drift & Long Outputs)
Even with Custom Instructions enabled, you have probably noticed that during long generations (over 800 words), em-dashes start creeping back in.
This occurs due to attention degradation: as the output grows longer, the model’s attention mechanism allocates more compute to maintaining coherence with recent paragraphs and less compute to negative system instructions established thousands of tokens earlier.
The Permanent Solution: Normalize Dashes in 1 Click
Instead of constantly arguing with ChatGPT prompts, use our free client-side cleaner. With the Normalize dashes rule enabled, all em-dashes (—) and en-dashes (–) are instantly converted into clean standard hyphens (-), spaced hyphens, or commas in under 15ms.
Pro Tip: Strip Hidden Formatting When Pasting
When pasting AI text into Google Docs, Microsoft Word, or CMS platforms like WordPress, standard Ctrl+V pastes rich HTML styles, non-breaking spaces (\u00A0), and font sizing.
Always use Paste as Plain Text:
- Windows: Ctrl + Shift + V
- Mac: Cmd + Shift + V (or Cmd + Option + Shift + V in Word)
Frequently Asked Questions
Do AI detectors flag em-dashes?
Yes. While an em-dash alone will not trigger an AI detector, tools like GPTZero and Turnitin measure perplexity (predictability) and burstiness (sentence variation). Repetitive em-dash clause insertions lower perplexity and create robotic rhythm that detectors frequently associate with LLM outputs.
Is an em-dash the same as a hyphen?
No. A hyphen (-, U+002D) is used for compound words (e.g., "well-known"). An em-dash (—, U+2014) is as wide as the letter "M" and represents a sudden break in thought. ChatGPT frequently uses em-dashes where a comma or period would be more natural.