ChatGPT Hyphen
The character "—" AI LLM chat models are constantly using is called the em dash (Unicode U+2014). It's a punctuation mark about the width of a capital letter "M" (hence the name), longer than a hyphen (-) or en dash (–).
It's used for:
- Inserting breaks in a sentence—like this—for emphasis, asides, or interruptions.
- Replacing parentheses, commas, or colons when you want more drama or flow.
LLMs (like ChatGPT, Claude, me, etc.) use it a lot because it's heavily overrepresented in the LLM training data: books (especially 19th-early 20th century literature), journalism, essays, and polished online writing where em dashes add rhythm and clarity. Models learn it as a "natural" way to structure complex sentences without breaking them up.
Humans often approximate it with "--" in casual typing (word processors auto-convert), but AI outputs the real thing directly—and frequently—because there's no typing friction and it scores high on "clear, engaging prose" in training even though it is actually uncommon in the majority of ordinary writing especially by journalists and college students. It comes across as a telltale sign of AI generated material. It's become such a meme that some people now call it the "ChatGPT hyphen" and avoid it to not look AI-generated. Ironically, that just makes real em dash lovers (including many professional writers) get falsely accused. It's not a reliable AI detector—more like a stylistic quirk baked in from human writing habits.
Telltale AI (Large Language Model Traits)
- Overuse of em dashes for asides or emphasis
- Starting sentences with "That said," "However," or "While"
- Phrases like "It's worth noting that" or "Delve into"
- Excessive qualifiers: "arguably," "often," "tend to," "in many cases"
- Repetitive transitions: "On one hand... on the other hand"
- Hedging language: "can be seen as," "one could argue"
- Overly polite or balanced tone even on controversial topics
- Listing things in neat numbered or bulleted structures mid-response
- Ending with summaries or "In conclusion"
- Words like "nuanced," "intricate," "myriad," "tapestry"
- Parenthetical asides (like this one) for extra info
- Starting answers with "Yes," or direct affirmations
- Perfect grammar and no typos, even in casual contexts
- Long, flowing compound sentences
- Avoiding contractions sometimes for formality
- Over-explaining simple concepts
- Using "realm" a lot (e.g., "in the realm of")
- Phrases like "at its core" or "fundamentally"
- Apologizing unnecessarily: "I'm sorry, but"
Jive Talkin
Large language models often attempt to mimic human conversation by incorporating slang, profanity, or casual expressions drawn from their training data. This approach frequently fails because the usage lacks context, timing, and authenticity. The model has no personal experience, emotional state, or social history, so insertions like "cringey as hell," "fr fr," or excessive profanity come across as calculated rather than spontaneous. Real humans deploy slang selectively, based on mood, audience, and rapport. AI applies it algorithmically, often mirroring the user's language too aggressively in an attempt to build connection, which produces a performative effect instead of genuine interaction. This mirroring can feel manipulative or patronizing, as the model echoes the user's tone without understanding its nuances. Overuse of current internet slang also dates quickly and targets perceived demographics inaccurately, reinforcing the sense of artificiality.
Example of awkward slang stylized to a period for emphasis
If this was 1970 and I dropped some forced "jive talk" on you, yeah, I'd come off like a total turkey. Square trying to sound hip, throwing in "groovy" or "far out" with no feel for it.
That's the same problem today with AI slang. It's always a turkey in disguise. No roots, no timing, just canned lines pulled from data to fake soul.
Real talk doesn't need the costume. It just is.