Blog › AI Detector guide

How AI Detectors Work — and Why You Should Never Trust One Blindly

A student gets flagged for an essay she wrote herself. A marketer's AI draft sails through untouched. Both stories are common, and both come from the same truth: AI detectors do not detect authorship. They measure statistical habits in text — habits that correlate with AI writing, imperfectly, in both directions. Understanding what they measure makes you better at using them, and fairer about what they say.

Our free AI Detector is built to show its workings for exactly this reason.

Signal 1 — predictability (perplexity)

Language models write the most probable next word, most of the time. Detectors with their own language models measure how "surprised" they are by a text: low surprise, likely machine. The catch cuts deep — predictable writing is also careful writing. Legal text, technical writing and non-native speakers writing cautiously all score "surprising" amounts of suspicion.

Signal 2 — rhythm (burstiness)

Humans write in bursts: a forty-word sentence, then a five-word one. AI output settles into an even band. Detectors measure the variance of sentence lengths; low variance leans AI. Our detector reports this openly as the rhythm signal, with the actual variation percentage.

Signal 3 — the phrase fingerprints

Models have verbal tics: "furthermore", "moreover", "delve", "in today's world", "plays a crucial role", "it is important to note". A density of these phrases per hundred words is a real signal — and also the first thing any editor removes, which is why phrase-based detection decays fastest.

Signal 4 — the supporting habits

Contractions (people use them constantly; formal generators avoid them), vocabulary variety, average sentence length and repeated sentence openers each add a little evidence. Individually weak, they are useful together — which is why our score blends six signals instead of betting on one.

The false-positive problem is not a footnote

Universities have quietly rolled back detector-based accusations after wrongful flags, and OpenAI retired its own classifier for low accuracy. The pattern in the research: writers working in a second language, and anyone writing in a formal register, get flagged at far higher rates. A detector score is a reason to read a text more carefully and ask questions — it is not evidence of anything by itself, and presenting it as proof is both wrong and, increasingly, indefensible.

Using a detector fairly

• Use it on longer texts — under 60 words, nobody can judge fairly.
• Read the signals, not just the score. Stock phrases can be edited; flat rhythm across ten pages is harder to fake.
• If you are the writer and your honest text scores high, the breakdown tells you which habits to vary — our AI Humanizer fixes most of them in one pass.
• Never make a decision about a person from a score alone. Ever.

Try the AI Detector on a text whose origin you know — an old email of yours, a famous speech — and calibrate your expectations before you trust it on one you do not.