What is AI detection?

An AI detector is a tool that reads a piece of text and estimates how likely it is that a language model — ChatGPT, Claude, Gemini and the like — wrote it. That's the whole job. It doesn't read minds, it doesn't see your edit history, and it never sees you type. It only ever sees the finished words, and it makes a statistical guess from them.

That last point matters more than anything else in this article, so it's worth saying twice: a detector judges the surface of the text, not the process behind it. Two paragraphs that read identically will score identically, no matter how they were actually written.

The three-way split

Older detectors returned a single number — "82% AI". Modern ones, snizzly's included, break that into three parts that add up to 100:

• ❧AI-generated — the text reads as though a model wrote it from a prompt, with little human shaping.
• ❧AI-refined — a human wrote the substance and a model polished the surface: smoothed the rhythm, swapped vocabulary, tidied transitions.
• ❧Human-written — the text reads as something a person drafted and revised themselves.

The split is more honest than a single percentage because most real writing in 2026 isn't purely one or the other. A student drafts an argument, then asks a model to fix the grammar. A professional dictates a rough memo and has a tool tighten it. "AI-refined" is the bucket for that middle ground — and a fair detector treats it as a partial signal, not a verdict.

What detectors are good at

On long, untouched, model-generated text, a good detector is genuinely accurate. Unedited model output has consistent fingerprints — even sentence lengths, a particular fondness for certain connectives, confident claims with no hedging — and 300+ words of it is hard to mistake for anything else.

What they are not good at

Short text is the big weakness. Under roughly 80 words there simply isn't enough signal, and any detector that returns a confident score on a single sentence is bluffing. The other weakness is the false positive: fluent, polished human writing — by a strong essayist, or by a non-native English speaker who learned the language formally — can carry some of the same evenness that models produce. A detector that hasn't been calibrated for this will flag real human work.

A detector score is evidence, not a verdict. Treat a high score as a reason to look closer, never as proof on its own.

This is why snizzly shows you the full three-way breakdown and a confidence level instead of a lone number — and why the detector is free to use with no sign-up. You should be able to see exactly where your writing stands before anyone else does.

❦