Probability-based detection¶
Probability-based detection methods leverage statistical properties of probabilistic models—such as likelihood scores, entropy, or distributional curvature—to identify synthetic or anomalous content. These approaches exploit the assumption that authentic and generated content have distinct probabilistic signatures under appropriate models.
Key approaches¶
- Likelihood thresholding: Classifying based on the raw log-probability of content under a reference model
- Curvature analysis: Using derivatives or Hessian properties of the probability landscape to detect anomalies
- Entropy-based methods: Leveraging information-theoretic properties to identify content distribution shifts
- Rank-based statistics: Analyzing the rank of tokens or sequences under the model's conditional distributions
Key papers¶
- Mitchell et al. (2023) — DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature: Proposes analyzing the negative trace of the Hessian of log-probability to detect LLM-generated text; outperforms simple likelihood and entropy-based baselines.
Related topics¶
- Machine-generated text detection — primary application
- Fake news detection methods — broader category