Twenty short snippets. Half written by humans, half by models. Classify each. Reveal at the end with a one-line tell per snippet. Your best score is saved; the leaderboard is opt-in.

Tells that usually work

AI tells

• Hedged symmetry. "While this approach has its merits, it also presents certain challenges…" Whenever a sentence balances two sides with no specificity, suspect a model.
• Empty superlatives. "Truly transformative", "an exciting new chapter", "the possibilities are endless." Humans reach for superlatives rarely and specifically. Models reach for them as punctuation.
• Corporate vocabulary. "Leverage synergies", "empower users", "navigate complexities", "sustainable growth" — content-farm vocabulary clusters hard in LLM output.
• Self-reference. "As an AI, I'm happy to…" — the giveaway, though usually stripped out before snippets go out.

Human tells

• Specific nouns you wouldn't invent. A time (7:12am), a name (Dave), a cat, a printer. Humans ground their jokes in stuff.
• Self-deprecation that lands. Good humans are mean to themselves in a way that's not performative. Models trained on safe text usually can't.
• Grammatical irregularity on purpose. A lowercase "ok", a fragment, a dropped period. Humans break grammar to convey tone. LLM outputs almost always end with a period.
• Admissions of not-knowing. "nobody knows why", "this is cursed", "refactor later" — the voice of someone who has lived with consequences.

Why any of this matters

The stock question "how do we detect AI-generated text" has no stable answer — the generators keep improving and every tell rotates out eventually. What doesn't rotate is the underlying shape: models are trained to sound like text, but humans write to sound like themselves. The gap is usually specificity.

Also useful: the same pattern-matching you use to spot AI in other people's writing is the pattern-matching you need to avoid it in your own. If your first draft reads like item #8 above ("Whether you're a seasoned professional…"), something is wrong — either a model wrote it or you're imitating one without meaning to.