On the Salience of Low-Probability Tokens for AI-Generated Text Detection: A Multiscale Uncertainty Perspective

AI-generated text increasingly blends with human writing, raising practical risks such as misinformation, academic misuse, and corpora contamination. While statistical detectors are appealing for efficiency and generalization, they suffer from two key limitations. (i) Boilerplate dominance, boilerplate tokens shared across human and LLM writing can overwhelm discriminative signals. (ii) Brittle point estimates, relying on a single probability score yields unstable decisions under adversarial manipulations. To address these issues, we propose $\textbf{Uncertainty}$, a multiscale uncertainty estimator that focuses on informative low-probability tokens, which more clearly expose distributional discrepancies. Locally, it alleviates boilerplate dominance by averaging the log-probabilities of low-probability tokens; globally, it reduces brittleness by capturing the distributional shape of this low-probability region via Rényi entropy. We further extend the detector to $\textbf{Uncertainty++}$ via conditional independent sampling, yielding a more stable uncertainty estimation. Experiments across seven datasets and sixteen LLMs demonstrate high effectiveness, generalization, and robustness. Our code is available at github.com/guoyikai2000/Uncertainty-AIGT.

How can we tell whether a piece of text was written by a person or by an AI, especially when both can use the same common phrases? We wanted to answer this question because unreliable detection can affect misinformation, academic integrity, and the quality of text collections used to train future models. Our paper presents the idea that the most useful clues often appear in words that a reference language model finds unusually surprising in a passage. This is important because many existing detectors average over all words, so routine phrases can drown out the evidence that actually separates human and AI writing. We built two methods, Uncertainty and Uncertainty++, that focus on these surprising words, technically called low-probability tokens, and also measure how uncertain the model is around them, rather than depending on a single probability score. Across many datasets and language models, our methods detect AI-generated text more accurately and remain more stable when the text style, topic, or generation method changes. Our findings suggest a more reliable way to support AI-generated text detection, while still requiring human judgment in high-stakes decisions.