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

TL;DR

This paper identifies late-stage stability patterns in AI-generated text, proposing simple features based on temporal dynamics that improve detection accuracy without needing model access.

Contribution

It introduces the concept of Late-Stage Volatility Decay and two novel features for AI text detection, achieving state-of-the-art results without perturbation sampling.

Findings

AI-generated text shows 24-32% lower volatility in late stages.

Proposed features outperform existing methods on benchmarks.

Method is effective without additional model access.

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.