Zero-shot detection

Zero-shot detection methods identify phenomena (such as machine-generated text, deepfakes, or synthetic content) without requiring labeled training data or domain-specific adaptation. These approaches leverage intrinsic properties of generated content or universal statistical signatures.

Key properties

• No training required: Methods operate on out-of-the-box models without fine-tuning
• Generalization: Can be applied across domains and models without retraining
• Scalability: Practical for rapid deployment without data collection overhead
• Trade-offs: Often lower accuracy than supervised methods, but applicable in data-limited settings

Key papers

• Mitchell et al. (2023) — DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature: Proposes a zero-shot method for detecting LLM-generated text by analyzing log-probability curvature using random perturbations.

Related topics

• Machine-generated text detection — application domain
• Fake news detection methods — broader category