<ul><li>Existing methods for the zero-shot detection of machine-generated text are dominated by three statistical quantities: log-likelihood, log-rank, and entropy.</li><li>A new method is introduced, targeting a defect in the way that decoding strategies normalize conditional probability measures, for detecting machine-generated text.</li><li>The new method is rigorously theoretically justified, easily explainable, and conceptually distinct from existing methods.</li><li>The performance of the new detector is at least comparable to other state-of-the-art text detectors and in some cases, it strongly outperforms them.</li></ul>

TempTest: Local Normalization Distortion and the Detection of Machine-generated Text

Discover more