The ramifications of LLM watermarks for authorship and groups

Abstract

Recently, watermarking schemes for large language models (LLMs) have been proposed to distinguish between text generated by machines and by humans. This paper explores philosophical, political, and ethical ramifications of developing and deploying watermarking schemes. A definition of authorship attribution that encompasses both machines (LLMs) and humans is proposed as a backdrop. It is argued that private watermarks may provide private companies with sweeping rights to determine authorship, which is incompatible with traditional standards of authorship attribution. Then, possible ramifications of the so-called “entropy dependence” of watermarking mechanisms are explored. It is argued that entropy levels may vary across different, socially salient groups. This could lead to group-dependent disparities in the detection rates of machine-generated text. Specifically, groups with a greater interest in low-entropy text may face challenges in detecting machine-generated text that is of interest to them.