AI watermarking is a chocolate teapot

Anthropic’s content-marking documentation, with my own annotation.

Is it me, or is this watermarking text stuff really quite useless?

“Claude, write this for me”…

“Now ChatGPT, rewrite it… without the watermark” 😅

So it’s not for helping people detect AI content themselves. Perhaps useful for labelling AI-generated slop so it’s not used for training data? Maybe.

Will it help block lazy content farms at scale? Meh, only the laziest.

Your watermark will only survive if people want it to… which is exactly where it’s useless surely.

If people are willing to attribute to AI, that’s not the danger.

It’s where people want to hide it, and this won’t stop them.

The bigger problem is if people start believing that this is useful, it becomes a shortcut for the opposite. A lack of watermark will be taken as human made… and so people will remove the watermark so they aren’t seeming like cheaters (weirdly)

There’s nuance between AI written completely and AI essentially as the modern spell and grammar check.

Overall, this seems like a chocolate teapot.

Am I missing something here?