Are AI Detectors Good Enough? A Survey on Quality of Datasets With Machine-Generated Texts
German Gritsai, Anastasia Voznyuk, Andrey Grabovoy, Yury Chekhovich
2024-10-21
Summary
This paper examines the effectiveness of AI detectors that identify machine-generated texts, questioning whether their high accuracy rates are truly reliable or just a result of poor-quality evaluation datasets.
What's the problem?
As AI language models become better at generating human-like text, there is a growing need for reliable detectors that can distinguish between human-written and AI-generated content. While some detectors claim to achieve up to 99.9% accuracy in controlled tests, they often perform poorly in real-world situations. This raises concerns about whether these detectors can be trusted or if their success is due to the low quality of the datasets used for testing.
What's the solution?
The authors conducted a systematic review of various datasets used in competitions for detecting AI-generated content. They proposed methods for evaluating the quality of these datasets to ensure they are robust and free from bias. Additionally, they discussed the potential of using high-quality generated data to improve both the training of detection models and the datasets themselves. This approach aims to enhance the understanding of how human and machine texts interact, ultimately supporting more reliable detection methods.
Why it matters?
This research is important because as AI-generated content becomes more common, ensuring the integrity of information is crucial. By improving how we evaluate AI detectors and the datasets they rely on, we can create more trustworthy tools that help identify machine-generated texts accurately, which is essential for maintaining quality in journalism, education, and other fields.
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.