Towards the Watermarking of Large Language Models: A Survey

Intellectual property theft through the unlawful sharing of texts is a pressing issue, further intensified by the rapid development of Large Language Models (LLMs). These models, designed for Natural Language Processing tasks, have reached a level of performance and popularity that makes them very vulnerable to misuse, such as the generation of fake text or false authorship attribution of the generated content. Text watermarking is the best tool to safeguard LLMs from such misuse. Because Large Language Models and text watermarking exhibit such complementarity, in this comprehensive survey we describe the state of the art of these two fields and how they can be combined. Our contribution can be summarized in the following four main aspects: (1) an overview of the fundamental characteristics of LLMs and their state of the art; (2) an overview and analysis of the main types of text watermarking methods that target existing text, and how they compare to each other; (3) an overview and analysis of the main LLM watermarking methods; (4) identification of underexplored areas through the formulation of targeted research questions, which are then addressed after the comprehensive study on LLMs and text watermarking.