AI detectors work by measuring two statistical properties of text and then making a guess: perplexity, which is how predictable each word is, and burstiness, which is how much sentence length and complexity vary. AI-generated writing tends to score low on both, because language models pick high-probability words and produce steady, even sentences. Human writing is usually messier. The key word is "usually," because that is where detectors go wrong. A detector never knows that a machine wrote something. It estimates, and it reports the estimate as a confidence score.
Key facts
- Perplexity measures how predictable a text's word choices are. Lower perplexity reads as more machine-like.
- Burstiness measures variation in sentence length and structure. AI text is more uniform, so lower burstiness also reads as machine-like.
- Detectors output a probability, not a fact, and accuracy drops sharply on text under 200 words and on edited text.
- The same two signals that flag AI also flag formal and non-native human writing, which is the root cause of false positives.
What is perplexity?
Perplexity is a measure of how surprised a language model is by the next word. If a model can easily predict each word, perplexity is low. Language models are built to choose the most probable next word, so their own output is, by design, low-perplexity. Human writers make less predictable choices, so human text usually has higher perplexity. Detectors treat low perplexity as a sign of AI. GPTZero's own documentation describes perplexity as a measure of how likely it is that an AI model would have chosen the exact same words. The flaw: simple, common, or formal human vocabulary also produces low perplexity, so it gets flagged too.
What is burstiness?
Burstiness describes the variation in a text's sentence structure. Human writing tends to mix short, punchy sentences with longer, complex ones, which creates a bursty, uneven rhythm. AI writing tends to produce sentences of similar length and structure, giving a flatter, steadier pattern. Detectors read low burstiness as machine-like. The same flaw applies: tightly edited or highly conventional human writing can be very even too.
Why detectors produce a probability, not a verdict
Because perplexity and burstiness exist on a spectrum and human and AI text overlap on both, there is no clean dividing line. Detectors pick a threshold and report how far a text sits from it, expressed as a percentage. "95% AI" means the text matches the model's AI patterns at that confidence, not that there is a 95% chance a machine wrote it. Those are different statements, and confusing them is how a statistical guess gets treated as proof. This is not just theory: OpenAI retired its own AI text classifier in 2023 because of its low rate of accuracy. See can AI detectors be wrong for what that confusion costs people.
Why the same method causes false positives
The honest problem with this approach is that the signals it relies on are not unique to AI. Non-native English writers, academics, and anyone writing in a formal register naturally produce low-perplexity, even-rhythm text, and detectors flag it. A 2023 Stanford study published in Patterns found seven detectors flagged non-native essays as AI 61% of the time. For the full data on accuracy and false-positive rates, see our AI detector accuracy reference.
How to check your text
If you want to see what your writing looks like through this lens, paste it into Metric37's free AI detector. It returns a 0 to 100 human score using the same kind of statistical reading detectors use, so you can see whether your text trends machine-like before anyone else runs it through a tool. As with every detector, treat the number as information, not a verdict.