Authorship and the Politics and Ethics of LLM Watermarks

Private companies could gain unprecedented control over authorship determination: The paper argues that if watermarking schemes are privately controlled, companies could claim sweeping rights to decide what counts as machine-generated versus human-authored text, undermining traditional standards for how authorship has historically been determined.

Watermark detection may work differently across social groups: The paper identifies that watermarking mechanisms depend on "entropy" (text randomness), and this entropy varies across different groups. This means machine-generated text could be easier to detect for some communities and harder for others, potentially creating unequal detection rates based on social factors.

Low-entropy text poses a detection challenge: Groups that prefer or produce lower-entropy text (more predictable, formulaic writing) may find it significantly harder to detect machine-generated content, since watermarking schemes are less effective on this type of text. This could disadvantage certain communities or use cases.

Authorship definitions need updating for the AI era: The paper proposes that our traditional understanding of authorship must evolve to account for machines as potential contributors, not just humans, raising fundamental questions about how we should legally and philosophically define who or what "authored" a piece of text.