How GPTZero works, and what its scores mean

What it measures

GPTZero was one of the first widely used AI text detectors and popularised two ideas: perplexity (how predictable each next word is to a language model) and burstiness (how much that predictability varies across sentences). Text written by a model tends to be uniformly predictable; human writing spikes and dips. Newer versions add trained classifier components and report document-level and sentence-level probabilities, plus a highlighted view of sentences it thinks are likely AI-generated.

The output is a probability, not a fact. GPTZero itself advises educators not to use the score alone to make decisions about a student, and its documentation describes the result as an estimate.

Where it misfires

Independent tests and GPTZero's own notes agree on the failure modes: short texts (under a couple of hundred words), formulaic genres such as lab reports and legal boilerplate, and, importantly, writing by non-native English speakers, which tends to be more predictable in vocabulary and can be flagged at higher rates. Heavily edited AI text and human text polished by grammar tools also blur the boundary.

Practically: treat any single score as one input. Vary sentence length, use concrete examples, and keep your own voice. Those are the properties the detector is actually reading.

Using AntiGPT alongside it

AntiGPT's in-app detector gives you an estimated-AI score before and after a rewrite so you can see the effect of your edits. It is a different classifier from GPTZero, so the numbers will not match, and neither should be read as a guarantee about the other. Use both as feedback on how the text reads, then decide what to submit as the author.