How AI Text Detectors Actually Work (No Hype, Just Mechanics)

AI detectors get talked about like they're a magic lie detector for text. They're not. Under the hood, most of them lean on a small handful of statistical signals, and understanding those signals makes the whole 'AI detection' conversation a lot less mysterious.

1. Perplexity

Perplexity measures how 'surprised' a language model is by each word in a sequence, given the words before it. AI models tend to generate the statistically most likely next word most of the time, which produces low perplexity — smooth, predictable text. Human writing is messier: we make less-likely word choices, take tangents, and vary phrasing in ways that push perplexity higher. Detectors use unusually low perplexity as a signal that a model, not a person, probably generated the text.

2. Burstiness

Burstiness looks at variation across a whole document rather than word-by-word. Human writing naturally 'bursts' — a short punchy sentence, then a long winding one, then a fragment. AI-generated text tends to even that out, producing sentences that cluster around a similar length and rhythm throughout. Low burstiness (i.e., very consistent sentence structure) is a second common tell.

3. Watermarking (less common, more controversial)

Some AI providers have experimented with watermarking — subtly biasing which words a model picks so the output carries a statistical signature that a matching detector can identify later. This only works if the detector knows which model's watermark to look for, and it's easily broken by any rewriting, paraphrasing, or translation of the text. It's far less commonly deployed at scale than perplexity/burstiness-based detection.

Why detectors disagree with each other

• Each tool sets its own thresholds for what counts as 'AI-likely'
• Each tool trains on a different mix of human and AI writing samples
• Some weigh perplexity more heavily; others weigh burstiness or sentence-level structure more
• None of them are checking a database of 'known AI text' — they're all estimating from patterns

The practical upshot: since detectors are pattern-matching for predictability and uniformity, text that restores natural variation — in sentence length, word choice, and structure — reads as more human to virtually all of them, even though each detector weighs the signals slightly differently.