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

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.

Key Contributions

Discovery of Late-Stage Volatility Decay: AI-generated text exhibits 24–32% lower log-probability volatility in the second half of sequences compared to human text
Two zero-shot detection features — Derivative Dispersion and Local Volatility — computed exclusively from late-stage token statistics without perturbation sampling or additional model access
State-of-the-art detection performance on EvoBench and MAGE benchmarks with strong complementarity to existing global methods

🛡️ Threat Analysis

Output Integrity Attack

Proposes novel AI-generated text detection features (Derivative Dispersion, Local Volatility) that identify whether text was produced by an LLM — directly addressing output integrity and content authenticity.