AI text watermarking

The internet is full of AI slop nowadays. I’m writing my blog for humans, so I write myself¹. I’d never randomly prompt an agent and copy-paste the output here without reading.

And so it is no surprise, that people want to see AI-generated text clearly marked. Well, it’s quite easy to tell apart written text from AI slop, but why should I spend my time on this? The problem is that most don’t mark their text in any way, either accidentally or deliberately.

That’s why Anthropic is introducing AI text watermarking. Claude will replace some words in generated text by synonyms to make the entire text have a specific checksum. I see a problem here though. Even without it AI texts are disgusting to read, what would be after it is enabled?

I’ve recently stumbled upon an example of such a technique (albeit for a different purpose) in an old game manual. Heavy use of synonyms was made there to justify monospaced text to fill a specific screen width. Although, it is impressive demonstration of writing skills, I doubt that it improves the quality of the text itself. And the same would happen with AI text watermarking as well.

Subscribe to all blog posts via RSS