Watermarking AI Output Can't Work in Practice

Machine-readable marks provide important signals about content, but it’s worth understanding their limitations across all content types.

I am afraid this is purely performative. While I can’t judge the exact technical limitations of the approach, Anthropic themselves are already pretty cautious on what to promise.

Consider how people are actually working with LLMs. It is a spectrum. At one end, workslop: “Hey ChatGPT, write the report/presentation/email for me.” At the other, an interactive session where the user works from a structural outline or a draft and edits that. Or creates the draft themselves, manually so to speak, then has the LLM proofread it and come up with a few suggestions to improve it. A mark is one bit. The practice is a dial.

Since that whole range is my experience and daily practice, I have simply put a footer on my blog linking to this AI use note.

Plenty of AI-generated content is meant to mislead. That is a different thing, and it’s not good at all. The new spam, and there will be a ton of it. Whether it takes down the internet (Dead internet theory) remains to be seen. But we seem to be on course, are we not?

P.S.: All the more reason for me to write and publish. That’s the silver lining.