Em-dashes, the word "delve", suspiciously perfect grammar — the internet is full of checklists for spotting AI text. Here's the uncomfortable truth: most of these signs stopped working, and the ones that remain flag real human writers just as often.
Rather not guess? Paste any text above — our tropa-3 model checks it in seconds.
Each of these shows up in every "how to spot ChatGPT" listicle. Here's what each one actually tells you.
The claim: “AI "loves" em-dashes, so lots of them means ChatGPT wrote it.”
Plenty of careful human writers use em-dashes constantly (journalists, essayists, anyone who read a lot of good prose). Newer models were tuned to use them less. This "tell" flags editors and stylists far more reliably than it flags AI.
The claim: “Certain vocabulary is a dead giveaway for AI.”
This was briefly true for 2023-era GPT output. Current models rarely lean on those words, and prompting them away takes one sentence. Meanwhile "delve" is completely normal in academic and non-native English — so this mostly catches real people, not machines.
The claim: “No typos, no slips — must be a machine.”
Grammarly, spellcheck and a careful proofread produce flawless human text every day. And AI is happy to write with deliberate errors if asked. Flawlessness tells you someone edited — not who (or what) wrote it.
The claim: “Humans vary sentence length ("burstiness"); AI is monotone.”
The most reliable of the classic signals — but still soft. Technical and academic writing is naturally uniform, and modern models deliberately vary rhythm. Useful as a hint, useless as proof.
The claim: “That bland "it is important to note" voice screams AI.”
A real pattern in un-edited AI output — but corporate and student writing sounds exactly the same, and one editing pass removes it. It points at effort level, not authorship.
The claim: “Formulaic scaffolding is an AI fingerprint.”
This is how millions of people were taught to write essays. It flags a five-paragraph format, not a language model — and current AI drops the scaffolding on request.
Notice what the classic signs have in common: em-dashes, formal vocabulary, clean grammar, structured paragraphs. Those aren't signs of a machine — they're signs of careful, educated, or non-native writing. Which means manual "AI spotting" systematically accuses the wrong people: the diligent student, the ESL writer, the professional editor.
This is exactly why we built a trained detection model instead of a checklist. On 1,314 verified human texts, heuristic-style detection flagged 9.7% as AI; our first model, tropa-1, flagged 1.2%. Our current model, tropa-3, flags 0.10% of 239,490 held-out human texts. A real model learns the difference the eye can't.
No single sign is proof. But some signals are more meaningful than the surface-level tells — especially when several appear together.
Confident citations, statistics or quotes that don't check out. AI invents plausible-sounding detail; humans usually leave gaps instead of inventing. This is one of the strongest tells you can verify yourself.
Real writing has concrete, idiosyncratic detail — a name, a date, a mistake, an opinion with a cost. Generic writing that could describe anyone is a softer but real signal.
The one thing you can't see: the deep token-level patterns a model leaves behind. This is what a trained detector measures — and why it beats eyeballing every time.
Even these aren't definitive. If the stakes are real — grading, hiring, publishing — treat any signal as a reason to look closer, not a verdict. See our guide for educators on using detection responsibly.
The signs are unreliable — a model trained on hundreds of thousands of texts isn't. Paste any writing and get a sentence-level read from tropa-3 in seconds.
No. Em-dashes are common in polished human writing, and newer models were tuned to use them less. Counting em-dashes flags careful writers far more often than it flags AI — it is not a reliable sign on its own.
Sometimes, with obvious un-edited output — but not reliably. Every surface-level "tell" (vocabulary, grammar, structure) can be edited away in seconds, and each one falsely flags real human writing, especially formal or non-native English. Reading is a starting point, not a verdict.
For a human reader: fabricated but confident specifics — citations, stats or quotes that do not check out. Beyond that, the strongest signal is statistical and invisible to the eye, which is what a trained detector measures.
Because heuristic-style detection keys on formality, clean grammar and even sentence rhythm — traits shared by good human writers and non-native speakers. It is the core problem in AI detection, and the reason we train our models to cut false positives. tropa-1 cut them from 9.7% to 1.2% on verified human text; tropa-3 flags 0.10% of 239,490 held-out human texts.
The text tells here are specific to writing. AI images have their own (also fading) tells — extra fingers, warped text, impossible reflections. Our detector checks images, audio and video as well as text.