How do AI detectors work?

Strip away the marketing and almost every AI detector is one of two designs. The first is statistical: it measures properties of the text and compares them to what models typically produce. The second uses a language model itself as a judge. Both work. Both can be wrong. Understanding the difference tells you a lot about when to trust a score.

Statistical detectors: perplexity and burstiness

The classic approach — used by early-2023 detectors like GPTZero — rests on two measurements. Perplexity asks: how surprised would a language model be by this next word? Model-written text tends to choose high-probability, unsurprising words — so it has low perplexity. Human writing wanders more, reaches for the odd unexpected word, and so reads as less predictable.

Burstiness measures variation in sentence length and complexity. People write in bursts — a long, winding sentence, then a short one. Models, left alone, tend to produce sentences of strikingly uniform length. Low burstiness is a model tell.

The weakness: both signals are easy to disturb. Vary your sentence lengths, drop in a few less-predictable words, and a purely statistical detector loses confidence fast. And formal human writing — academic, legal, diplomatic — is naturally low-burstiness, which is exactly why these detectors over-flag it.

LLM-judge detectors

The newer approach asks a capable language model to do the judging directly: here is some text, assess how it was likely produced. A strong model can weigh things a perplexity score can't — whether an argument has a real point of view, whether the rhetorical craft looks deliberate, whether hedging and counter-argument are present the way a thoughtful human writer would use them.

This is the family snizzly's detector belongs to. It's better at the hard cases — the skilled essayist, the non-native writer — because it can recognise human craft as a human signal instead of mistaking its polish for a machine. But it has its own failure mode: a judge is only as fair as its calibration. An LLM judge that's been told "polished equals AI" will punish good writers.

Every detector has a failure mode. The honest question isn't "is it accurate?" but "accurate for whom, and wrong in which direction?"

Why no detector can be perfect

There's a deeper reason detection will always be probabilistic. Modern models are trained to write like people. As they improve, the gap between the best model output and good human writing narrows — by design. A detector is measuring that shrinking gap. So a detector's job isn't to deliver certainty; it's to give you a well-calibrated estimate and be honest about its confidence.

That's the standard snizzly holds itself to: a calibrated judge, a confidence level on every result, and a benchmark the detector is re-tested against after every change. You can read the exact calibration approach in the project's detector notes — we'd rather show the working than claim a number we can't stand behind.

❦