Are AI Detectors Good Enough? A Survey on Quality of Datasets With Machine-Generated Texts

TL;DR

This paper addresses the gap between high benchmark scores for AI-generated text detectors and their real-world robustness by systematically reviewing datasets used for AI-generated-content detection and proposing multi-faceted dataset quality assessments. It introduces three evaluation axes: topological statistics via $PHD$ and the symmetric $\text{KL}_{\text{TTS}}$, perturbation-based robustness via $\Delta_{\text{shift}}$ and $\text{KL}_{\text{shuffle}}$, and a set of strong baselines (DetectGPT/Fast-DetectGPT, Binoculars, and $\text{mDeBERTa}$). The findings reveal substantial variability in data quality across datasets, with detector performance not always reflecting underlying data properties, and highlight issues with short texts and model evaluators. The work argues for using high-quality generated data to both train detectors and clean training datasets, and provides a foundation for a more reliable, multi-faceted benchmarking framework to preserve information integrity in an increasingly automated information landscape.

Abstract

The rapid development of autoregressive Large Language Models (LLMs) has significantly improved the quality of generated texts, necessitating reliable machine-generated text detectors. A huge number of detectors and collections with AI fragments have emerged, and several detection methods even showed recognition quality up to 99.9% according to the target metrics in such collections. However, the quality of such detectors tends to drop dramatically in the wild, posing a question: Are detectors actually highly trustworthy or do their high benchmark scores come from the poor quality of evaluation datasets? In this paper, we emphasise the need for robust and qualitative methods for evaluating generated data to be secure against bias and low generalising ability of future model. We present a systematic review of datasets from competitions dedicated to AI-generated content detection and propose methods for evaluating the quality of datasets containing AI-generated fragments. In addition, we discuss the possibility of using high-quality generated data to achieve two goals: improving the training of detection models and improving the training datasets themselves. Our contribution aims to facilitate a better understanding of the dynamics between human and machine text, which will ultimately support the integrity of information in an increasingly automated world. The code is available at https://github.com/Advacheck-OU/ai-dataset-analysing.

Figures (4)

• Figure 1: Comparison of embedding shifts after two types of modifications for the HC3 dataset.
• Figure 2: Topological Time Series for different datasets. The results for the remaining datasets selected in this paper can be found in Figure \ref{['fig:tts']}.
• Figure 3: PHD values on all datasets, except TweepFake and AuTex23 Spanish, texts from which were too short for proper calculation of PHD.
• Figure 4: Topological Time Series on all datasets.