AI detectors feel like magic — paste in some text and get back "85% AI." But underneath, they're doing something much more mundane and much less certain than people assume. Understanding how they actually work explains both what they catch and why they're so often wrong.
They measure predictability, not authorship
A detector cannot see who typed something. What it can do is run the text through a language model and ask: how predictable is each word, given the words before it? Language models are trained to produce the most likely next word, so AI-generated text tends to be very predictable. Human writing is messier and more surprising. Detectors turn that difference into a score.
The two main signals: perplexity and burstiness
Perplexity
Perplexity is a measure of how "surprised" a language model is by the text. Low perplexity means the words are highly predictable — exactly what a model produces. High perplexity means the writing takes turns the model wouldn't expect. Detectors broadly treat low perplexity as a sign of AI.
Burstiness
Burstiness measures variation in sentence structure and length. Humans write in bursts — a long sentence, then a short one, then a fragment. AI tends to be more uniform. Low burstiness (everything roughly the same) leans "AI"; high burstiness (lots of variation) leans "human."
Why detectors produce false positives
Here's the core problem: "predictable and even" is not the same as "written by AI." Plenty of human writing is predictable and even, including:
- Writing by non-native English speakers, who often use simpler, more regular phrasing
- Technical, legal or academic writing, which is formulaic by design
- Anyone taught to write in a clear, structured, "five-paragraph" style
These get flagged as AI even when a person wrote every word. There have been widely reported cases of detectors flagging classic human texts and students' genuine essays. That's why responsible institutions warn against treating a detector score as proof of misconduct.
Beyond perplexity and burstiness: classifiers and watermarking
Perplexity and burstiness are the core idea, but most production detectors don't stop there. Many layer a trained classifier on top — a model fine-tuned on large sets of known human-written and known AI-written text, learning to recognize combinations of patterns beyond the two headline signals. It's still statistical pattern-matching, just with more dimensions than a human could manually inspect. This is one reason different detectors score the same text differently — each one was trained on a different mix of examples.
A separate approach, used by a handful of AI providers rather than by third-party detectors, is watermarking: embedding a statistical pattern into generated text at the point of creation (Google's SynthID for Gemini output is one public example) so the same provider's tools can later check for it. This only works on text produced by that specific model and scheme — it can't tell you anything about text from a different model or a human. It's also why a detector generally can't say which AI wrote something, only that the writing looks machine-like: without a matching watermark, the perplexity and burstiness patterns of one model's output look broadly similar to another's.
Why detection is an arms race
As models improve and as people learn to write with more variation, the gap detectors rely on shrinks. Each new model is harder to distinguish from human writing, and detectors are constantly retrained to keep up. The result is a moving target where scores swing as both sides evolve — another reason a single percentage shouldn't be treated as fact.
What this means for you
Three practical takeaways:
- A score is an estimate, not evidence. Detectors are probabilistic and demonstrably fallible.
- Even, predictable writing scores higher — so genuinely human writing that happens to be flat can be flagged.
- Writing with natural variation — varied sentence lengths, plain word choice, real voice — reads as human because that's literally what detectors are measuring.
The takeaway
AI detectors are pattern-matchers measuring predictability, not authorship verifiers. They're useful as a rough signal and unreliable as a verdict. If your goal is writing that reads naturally and human, focus on the qualities detectors reward — variation and natural phrasing. A humanizer can help by rewriting flat text into something with more natural rhythm, but the principle matters more than any tool: human writing is varied writing.