When AI Settles Down: Late-Stage Stability as a Signature of AI-Generated Text Detection

2026-01-08

arXiv.org

AI Summary

This work addresses a critical limitation in existing zero-shot AI text detection methods, which overlook the temporal dynamics inherent in autoregressive generation and consequently struggle to distinguish human- from AI-authored text. Through analysis of over 120,000 samples, the study reveals—for the first time—that log-probability fluctuations in AI-generated text exhibit significant attenuation in later generation stages. Leveraging this insight, the authors propose two lightweight statistical features derived exclusively from the latter half of a text: derivative dispersion and local volatility. Requiring neither perturbation-based sampling nor access to additional models, the method achieves state-of-the-art performance on the EvoBench and MAGE benchmarks and demonstrates strong complementarity with existing global approaches. Empirical results show that the volatility in the second half of AI-generated text is 24%–32% lower than that of human-written text.

Technology Category

• Natural Language Processing: Generation
• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Zero-shot detection methods for AI-generated text typically aggregate token-level statistics across entire sequences, overlooking the temporal dynamics inherent to autoregressive generation. We analyze over 120k text samples and reveal Late-Stage Volatility Decay: AI-generated text exhibits rapidly stabilizing log probability fluctuations as generation progresses, while human writing maintains higher variability throughout. This divergence peaks in the second half of sequences, where AI-generated text shows 24--32\% lower volatility. Based on this finding, we propose two simple features: Derivative Dispersion and Local Volatility, which computed exclusively from late-stage statistics. Without perturbation sampling or additional model access, our method achieves state-of-the-art performance on EvoBench and MAGE benchmarks and demonstrates strong complementarity with existing global methods.

Problem

• AI-generated text detection
• temporal dynamics
• late-stage stability
• zero-shot detection
• autoregressive generation

Innovation

• Late-Stage Volatility Decay
• Zero-shot Detection
• Derivative Dispersion
• Local Volatility
• AI-Generated Text Detection