How does GPTZero work, and how accurate is it?
What GPTZero actually measures today, why perplexity and burstiness are only part of the story, and how to read its result fairly.
GPTZero is one of the best-known AI detectors, especially in schools. It started as a student project in January 2023 and is now part of Superhuman. Here is what it looks at, what its accuracy numbers do and don't tell you, and what to do if it flags your work.
Perplexity and burstiness: where it started
The first version of GPTZero used two simple measures. Perplexity asks how predictable each word is to a language model: very predictable text looks machine-made. Burstiness asks how much that predictability changes from sentence to sentence: people mix plain and surprising sentences, while models tend to keep an even pace.
GPTZero says it moved to a deep-learning model in autumn 2023. Perplexity and burstiness are now only one of seven signals it combines, alongside its own trained classifier and text search.
What the accuracy numbers mean
- GPTZero advertises about 99% accuracy across its test sets.
- Detector makers each publish comparisons that put themselves first, and the numbers disagree. For example, GPTZero's own comparison reports a much higher false-positive rate for Copyleaks than Copyleaks reports for itself.
- Accuracy on a benchmark is an average. It does not tell you how the detector treats your genre, your length or your writing style.
Treat any single vendor's accuracy claim as marketing until an independent test on text like yours confirms it.
Who gets flagged unfairly
Short texts, formal templates (cover letters, lab reports, legal writing) and plain, simple English are the usual false positives on any detector. Researchers have shown that non-native English writers are flagged more often, because they tend to use common, predictable words (Liang et al., 2023).
If GPTZero flags your own writing
- Look at the highlighted sentences, not just the headline percentage.
- Keep your drafts, notes and document version history: they show how the text was made.
- Ask for a second opinion from a different detector, and remember that a score is an estimate, not proof.
- When you revise, add what only you know: a real example, a number, your own view. That helps readers as much as detectors.
What we saw in our own tests
While building TrulyWrite we checked a few rewritten passages with GPTZero in November 2026. Long rewrites that kept real details were rated as human; we have not run a large GPTZero test, so we don't publish an accuracy figure for it.