How Sampling Affects the Detectability of Machine-written texts: A Comprehensive Study

TL;DR

This work critically assesses how decoding-based sampling strategies affect the detectability of machine-generated text, showing that modest changes in generation parameters can drastically reduce detector performance. By introducing a large-scale benchmark with 37 decoding configurations and evaluating both supervised RoBERTa-based detectors and unsupervised detectors (Binoculars, FastDetectGPT), it reveals systematic robustness gaps in current ATD methods. The study finds that AUROC can plummet from near-perfect to near-random under certain settings and demonstrates that training detectors on diverse, mixed-generation data improves generalization, while domain shifts remain a challenge. These results highlight the need for broader, more realistic evaluation protocols and drive toward more trustworthy and robust artificial-text detection approaches.

Abstract

As texts generated by Large Language Models (LLMs) are ever more common and often indistinguishable from human-written content, research on automatic text detection has attracted growing attention. Many recent detectors report near-perfect accuracy, often boasting AUROC scores above 99\%. However, these claims typically assume fixed generation settings, leaving open the question of how robust such systems are to changes in decoding strategies. In this work, we systematically examine how sampling-based decoding impacts detectability, with a focus on how subtle variations in a model's (sub)word-level distribution affect detection performance. We find that even minor adjustments to decoding parameters - such as temperature, top-p, or nucleus sampling - can severely impair detector accuracy, with AUROC dropping from near-perfect levels to 1\% in some settings. Our findings expose critical blind spots in current detection methods and emphasize the need for more comprehensive evaluation protocols. To facilitate future research, we release a large-scale dataset encompassing 37 decoding configurations, along with our code and evaluation framework https://github.com/BaggerOfWords/Sampling-and-Detection

Figures (5)

• Figure 1: Changes in the decoding strategy can lead to errors in detection.
• Figure 2: Heatmap of accuracy detection rates when training using data generated with the row parameter and testing on the column one. For further details, this heatmap is displayed in full in Appendix \ref{['app:supervised-details']} Figure \ref{['fig:roberta-accuracy-details']}.
• Figure 3: Scores histograms for top-$p=0.9$ (left) and repetition penalty $=1.05$ (right). For top-$p$, both detectors behave as intended, clearly separating human and machine scores. On the right, Binoculars switches the scores, going over the threshold value that was around 0.9 (this can be fixed by introducing a second theshold for high scores). For FastDetectGPT, the scores histograms are completely blended together, showing that repetition-penalty fully breaks that method.
• Figure 4: Heatmap of accuracy detection rates when training using data generated with the row parameter and testing on the column one.
• Figure 5: Heatmap of AUC detection rates when training using data generated with the row parameter and testing on the column one.These values can be directly compared to the ones in Table \ref{['tab:auroc-subtables']}.