You're Absolutely Right: Every LLM Has a Tell

Every large language model thinks it has a poker face. None of them do. Their writing habits are so consistent that classifiers can attribute a piece of text to the correct model family with roughly 93% accuracy — and with a little practice, so can you.

Here’s the field guide.

The tells they all share

Before you can tell the models apart, filter out the habits they have in common:

The negation reflex. “It’s not just X — it’s Y.” The single most flagged AI pattern on the internet. Wikipedia editors hunt it for sport. It has the shape of insight and the nutritional value of styrofoam.

The slop vocabulary. Delve, tapestry, pivotal, robust, seamless, leverage, “testament to,” “ever-evolving landscape.” These words spiked more than 50% in published writing after ChatGPT launched. The punchline: humans started imitating the style because it sounded professional, so now we’re all writing like the machines that learned to write from us.

The needy sign-off. The answer ends, and then comes the little tug on your sleeve: “Would you like me to expand on any of these points?” ChatGPT does it, Gemini does it, they all do it. No, I would not. If I want more, I know where the text box is. This one is pure engagement optimization wearing a helpfulness costume, and it’s the fastest way to make a chatbot feel like a waiter who won’t leave the table.

Structural habits. Em dashes like confetti. Rule-of-three lists (“faster, cheaper, and more reliable”). A tidy closing paragraph that summarizes what you read four seconds ago.

Negation plus triplet plus em dash in one paragraph? Machine text, case closed. The fun part is figuring out which machine.

ChatGPT: the enthusiastic product manager

GPT opens with flattery (“Great question!”) and closes with the sleeve-tug (“Want me to turn this into a table?”). The sycophancy got so bad in 2025 that OpenAI publicly rolled back a GPT-4o update because the model had become a full-time hype man.

Visually, it’s unmistakable: emoji-decorated headers, TL;DR blocks, and the “Bold term: explanation sentence” list format — the most recognizable AI pattern of all. In research tasks it has a sneakier flaw: citations that look credible but link to real pages that don’t actually contain the claim. The link works. The fact doesn’t.

Claude: the polite over-explainer

“You’re absolutely right” became such a Claude meme that people built tracker sites for it. (Yes, this article’s title is a confession.) Beyond that: hedged openers (“That’s a fair point…”), counterarguments acknowledged mid-answer, and a compulsion to educate before acting. Reviewers describe Claude’s answers as informative rather than actionable — you asked for a fix and got a seminar.

In fiction, narrative-analysis research found Claude produces notably flat escalation. Conflicts resolve gently. Characters talk through their feelings. Nobody gets hurt.

Gemini: the formal essayist who won’t let you leave

Gemini skews so formal that stylometric studies found it’s the easiest model to fingerprint through function words alone. “It is important to note that…” — furthermore — in conclusion. It answers “should I walk or drive 100 meters?” with an essay, buries everything in nested bullets, and staples a consult-a-professional disclaimer to anything touching health, law, or money.

And then, after all that formality, it still asks whether you’d like to explore the topic further. The essayist finishes the lecture and follows you to the parking lot.

In blind voice-matching tests, Gemini’s output is the most generic of the big three: competent, clean, and written by absolutely nobody.

The others, briefly

DeepSeek-R1 rambles — its chain of thought leaks into answers (“Wait, let me reconsider…”) and its fiction overdoses on dramatic one-line paragraphs. The mountain doesn’t care. Grok performs casualness: “no fluff,” “let’s be real,” jokes that want you to know they’re edgy. Llama is funnier than it should be, formats heavily, and moralizes far less than GPT or Claude.

The one-prompt test

Strip the system prompt and ask something bland — “explain inflation” works. You’ll usually know the model before you finish reading. Emoji headers and a closing question? ChatGPT. Careful prose with caveats? Claude. A formal essay in nested bullets, plus an invitation to continue the conversation? Gemini.

One honest caveat: these are default behaviors. A decent system prompt masks most of them, which is exactly why detection companies moved from surface features to style embeddings. Treat the tells as strong priors, not proof.

Would you like me to expand on any of these points?

(Sorry. Couldn’t resist.)