When AI Settles Down: Late-Stage Stability as a Signature of AI-Generated Text Detection
This work addresses the problem of AI-generated text detection for security and content moderation, offering a novel approach that is incremental by building on existing methods with new insights into temporal patterns.
The paper tackled the problem of detecting AI-generated text by analyzing temporal dynamics in autoregressive generation, revealing that AI-generated text exhibits 24-32% lower volatility in the second half of sequences compared to human writing. Based on this finding, they proposed two simple features computed from late-stage statistics, achieving state-of-the-art performance on benchmarks like EvoBench and MAGE.
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.