Erkennungsverfahren für KI-generierte Texte: Überblick und Architekturentwurf

Transformer-based AI systems have made great progress in areas such as text processing and comprehension. These deep learning models enable the generation of texts and form the basis of modern language models. The rapid development of recent years has produced large language models such as ChatGPT, Bard and VICUNA-13B.

This article presents the development of language models towards large language models. The progressive development of language models has resulted in unimagined possibilities and, at the same time, a variety of problems, which is why it is important to recognize LLM-generated textual content. This article presents the different approaches of known recognition methods. In addition to statistical classification methods, deep learning-based and zero-shot methods are also discussed. In addition, a compression-oriented approach is presented as well as labeling methods. After a tabular comparison of the methods presented in the literature, implemented software detectors are presented. Subsequently, considerations for the design of a training dataset are presented, creating the basis for an own approach for the recognition of AI-generated texts in German. In addition, the architecture and design of our own approach, the AI content detector, is presented and described and the limitations are shown.