Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sure! Here's a short and concise title for your paper: "Contamination in Generated Text Detection Benchmarks"

Philipp Dingfelder, Christian Riess

TL;DR

The paper investigates contamination artifacts in the DetectRL benchmark for AI-generated text detectors, revealing that common generation patterns can be exploited as shortcuts by detectors. It analyzes the types and prevalence of contamination, then implements a regex-based data-cleansing pipeline and a secondary GPT-4.1-mini re-cleaning step to produce a higher-quality dataset. Through SHAP explainability and adversarial prompt attacks, it demonstrates that cleaning reduces susceptibility to spoofing while preserving cross-domain generalization. The authors provide a cleaned DetectRL dataset and discuss implications for benchmark reliability and future improvements in data quality practices for generated-text detection research.

Abstract

Large language models are increasingly used for many applications. To prevent illicit use, it is desirable to be able to detect AI-generated text. Training and evaluation of such detectors critically depend on suitable benchmark datasets. Several groups took on the tedious work of collecting, curating, and publishing large and diverse datasets for this task. However, it remains an open challenge to ensure high quality in all relevant aspects of such a dataset. For example, the DetectRL benchmark exhibits relatively simple patterns of AI-generation in 98.5% of the Claude-LLM data. These patterns may include introductory words such as "Sure! Here is the academic article abstract:", or instances where the LLM rejects the prompted task. In this work, we demonstrate that detectors trained on such data use such patterns as shortcuts, which facilitates spoofing attacks on the trained detectors. We consequently reprocessed the DetectRL dataset with several cleansing operations. Experiments show that such data cleansing makes direct attacks more difficult. The reprocessed dataset is publicly available.

Sure! Here's a short and concise title for your paper: "Contamination in Generated Text Detection Benchmarks"

TL;DR

The paper investigates contamination artifacts in the DetectRL benchmark for AI-generated text detectors, revealing that common generation patterns can be exploited as shortcuts by detectors. It analyzes the types and prevalence of contamination, then implements a regex-based data-cleansing pipeline and a secondary GPT-4.1-mini re-cleaning step to produce a higher-quality dataset. Through SHAP explainability and adversarial prompt attacks, it demonstrates that cleaning reduces susceptibility to spoofing while preserving cross-domain generalization. The authors provide a cleaned DetectRL dataset and discuss implications for benchmark reliability and future improvements in data quality practices for generated-text detection research.

Abstract

Large language models are increasingly used for many applications. To prevent illicit use, it is desirable to be able to detect AI-generated text. Training and evaluation of such detectors critically depend on suitable benchmark datasets. Several groups took on the tedious work of collecting, curating, and publishing large and diverse datasets for this task. However, it remains an open challenge to ensure high quality in all relevant aspects of such a dataset. For example, the DetectRL benchmark exhibits relatively simple patterns of AI-generation in 98.5% of the Claude-LLM data. These patterns may include introductory words such as "Sure! Here is the academic article abstract:", or instances where the LLM rejects the prompted task. In this work, we demonstrate that detectors trained on such data use such patterns as shortcuts, which facilitates spoofing attacks on the trained detectors. We consequently reprocessed the DetectRL dataset with several cleansing operations. Experiments show that such data cleansing makes direct attacks more difficult. The reprocessed dataset is publicly available.

Paper Structure

This paper contains 10 sections, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Analysing the Level of Contamination
  4. Methodology and Data Cleansing
  5. Impact of the Contamination
  6. Removal of the Identified Contaminations
  7. Results
  8. Summary and Outlook
  9. Acknowledgments.
  10. Disclosure of Interests.

Figures (2)

  • Figure 1: Comparison of SHAP token importance of a text for a model trained on Claude-generated, multi-domain data. The sample is labeled as LLM-generated. The classifier is trained to predict whether text is human-generated or AI-generated. Values close to 1 indicate human-generated text, while values below 0.5 indicate AI-generated text.
  • Figure 2: Comparison of SHAP token importance in a waterfall plot of a text for a model trained on Claude-generated, multi-domain data. The sample is human-written and an adversarial attack is applied.