Original title: The ramifications of LLM watermarks for authorship and groups
Tim Räz · September 2026
People want to know if a human or a computer wrote a text. To solve this, researchers use watermarks, which are hidden signals placed inside text generated by large language models to help identify it. This paper examines the philosophical, political, and ethical problems this creates. The authors propose a shared definition of authorship for both machines and humans. They warn that private watermarks could let private companies control who gets credit as an author, clashing with traditional rules. They also find that watermarks rely on entropyA measure of how unpredictable or varied text is; low entropy means more predictable., which refers to how predictable text is. Because predictable writing has low entropy, groups interested in low-entropy text may have a harder time detecting machine-generated content, creating unfair differences across social groups.
Imagine you're at a party and everyone has to wear a name tag. The tags are designed to be easy to read, but some people have handwriting that's messy, so their tags are harder to read. Now suppose the party host decides that anyone whose tag is hard to read must be a spy. That's unfair because the messiness isn't about spying, it's just how they write. In this paper, watermarks are like those name tags, and the messiness is like the predictability of text. Some groups write in a more predictable way, so their AI-generated text might be easier to spot, while others might get away with it, leading to unfair differences in who gets flagged.
If AI text watermarks are deployed, they could give companies power to decide who wrote what, and they might work worse for some groups than others. This affects anyone who reads or writes online, especially those whose writing style is more predictable. The stakes are about fairness and who controls authorship.
The abstract does not provide empirical data or specific examples of watermarking schemes. It focuses on theoretical arguments and does not quantify the disparities in detection rates. The paper does not propose a solution or policy recommendation.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
The ramifications of LLM watermarks for authorship and groups — Tim Räz — AI and Ethics — doi:10.1007/s43681-026-01389-5
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.