Text Detection Concepts

AI text detection relies on identifying statistical patterns that distinguish model-generated text from human-written text.

Perplexity and Burstiness

Perplexity measures how "surprised" a model is by the text. LLMs tend to generate text with lower perplexity (more predictable) than humans. Burstiness measures the variation in perplexity. Humans tend to write with more variable sentence structures, leading to higher burstiness.

Zero-Shot Detection

Methods like DetectGPT and Binoculars use the model itself (or a proxy) to detect its own output without needing training data. They rely on the observation that model-generated text occupies "negative curvature" regions of the model's log-probability function.

Outlier Detection

HumanOOD treats human text as the "in-distribution" data and AI text as outliers. By clustering embeddings of human text, we can detect AI text as being far from the cluster center.