AI evidence extraction

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

Matthieu Dubois, François Yvon, Pablo Piantanida · ArXiv.org · 2025

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.48550/arxiv.2510.13681

Methodology & findings

Study design

Comprehensive empirical study combining supervised and unsupervised text detection methods evaluated on a large-scale dataset.

Main result

The study found 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." More specifically, "state-of-the-art detection systems are highly sensitive to generation parameters—revealing drastic performance drops (e.g., AUROC from 0.99 to 0.01)—thus exposing critical blind spots in current benchmarks."

Research paradigm

Empirical-computational positivism

Author conclusions

The authors conclude: "In this work, we have presented a systematic study of how sampling-based decoding strategies affect the detectability of texts produced by large language models. Considering 37 different configurations—varying multiple generation parameters, we have shown that even modest adjustments in the (sub)word-level generation distribution can undermine state-of-the-art detectors, observing AUROC drops from near-perfect down to null simply by tuning the repetition-penalty. Our analysis sheds light on the underlying failure mechanisms: different sampling techniques induce distinctive distributional signatures (e.g., changes in token entropy or frequency tails) that current detectors are not able to recognize. This suggests that robustness in detection is still an open issue, both for supervised and unsupervised methods."

Risk of bias

Single language bias (English only); Single base model bias (Llama-3.2-3B); Limited detector architectures tested (one RoBERTa variant); Training-test split bias: supervised detector trained and tested on same generator model; Domain bias: human texts from specific sources (RAID dataset, BBC News, Common Crawl); Selection bias in decoding parameter choices; Training and test set used texts from same generator model (Llama-3.2-3B), potentially introducing model-specific bias; Overfitting risk in supervised detector to in-domain human training data (BBC News articles); Potential selection bias in choosing specific decoding parameter ranges from continuous spaces; Only English language evaluated, limiting generalizability; Single supervised architecture (RoBERTa-base) tested; Only two unsupervised methods evaluated; Limited to single language (English); Limited to single base model (Llama-3.2-3B); Single supervised architecture evaluated (RoBERTa); Only two unsupervised detectors tested; Domain mismatch in validation experiment (BBC News training vs. CCNews test) showed brittleness; Generation parameters selected based on availability in HuggingFace, not exhaustive coverage

Limitations

"In this empirical study of the impact of the generation parameters and strategies used in text generation, we had to limit ourselves to a small number of parameters, and to a restricted set of values
Likewise, our study only considered one supervised architectures, and two unsupervised models
Finally, only text generation in English was considered, using Llama-3.2-3B
These choices were made for the sake of limiting the computational expenses to a reasonable budget."

Open questions raised

Need for more comprehensive evaluation protocols for detection methods
Robustness of detection systems to advanced generation techniques such as Minimum Bayes Risk (MBR) decoding and Monte Carlo Tree Search (MCTS)
Detection of texts generated across multiple languages
Detection across different base language models beyond Llama-3.2-3B
Understanding why two-model unsupervised methods require detector models to assign very similar probabilities
Out-of-domain generalization for supervised detectors

Data: RAID dataset (subset of 2,000 English texts from human samples); Generated dataset: 37 decoding configurations × 2,000 samples = 74,000 machine-generated texts; Available at: https://github.com/BaggerOfWords/Sampling-and-Detection; RAID dataset (human samples extracted: 2,000 English texts across 11 domains); Generated dataset: 37 decoding configurations with 2,000 samples each; Dataset release URL: https://github.com/BaggerOfWords/Sampling-and-Detection; RAID dataset (human samples): https://github.com/BaggerOfWords/Sampling-and-Detection; Generated dataset with 37 decoding configurations released at: https://github.com/BaggerOfWords/Sampling-and-DetectionCode: https://github.com/BaggerOfWords/Sampling-and-Detection (contains dataset, code, and evaluation framework); https://github.com/BaggerOfWords/Sampling-and-Detection; https://github.com/BaggerOfWords/Sampling-and-Detection (includes code, evaluation framework, and dataset)Extracted from: pdfAgreement 46%

Explore related topics

Hallucination Control Research Integrity Empirical Study