DetectGPT is a zero-shot method for estimating whether language-model text was machine-generated without relying on labeled examples. It makes small changes, or perturbations, to a passage and observes how a language model scores those altered versions. Stable high scores may suggest generated text; greater variation may suggest human writing with more unexpected phrasing. section 6.9
Early experiments reported up to 95% accuracy across several open-source models, but later testing exposed a major limitation: performance fell when the scoring model differed from the model that generated the text. DetectGPT therefore provides evidence, not a dependable verdict, and should not serve as the sole basis for consequential decisions. section 6.9
