Explainers
How AI Detectors Work: Perplexity, Burstiness & Classifiers Explained
November 12, 2026 · 8 min read
GPTZero, Turnitin, Originality.ai, Copyleaks and others all try to answer one question: does this text look machine-written? Most rely on three kinds of signal.
1. Perplexity
Perplexity measures how surprised a language model is by each next word. AI text tends to pick likely words, so it has low perplexity. Human writing has more unexpected choices.
2. Burstiness
Burstiness is variation: long sentences next to short ones, simple next to complex. People write in bursts. AI writes evenly. That's why HumanizePaper shows a sentence variety score.
3. Trained classifiers
Many detectors train a model on labeled human and AI text. These pick up subtler cues such as favorite words, paragraph shapes and hedging patterns.
Why detectors make mistakes
- Short texts give too little signal.
- Formulaic genres look predictable even when human-written.
- New AI models write differently from the ones detectors trained on.
The takeaway
Detector scores are probabilities, not verdicts. The most reliable way to sound human is to write with specifics, opinions and varied rhythm. Here are 10 techniques.