Claude heads take note

I got a bunch of DMs and text messages asking me to explain the AI watermarking thing. We think that we can spot AI writing, there's the obvious tales, the M-dash, much maligned, also load-bearing, the format, it's not X, it's Y. And you may be familiar with sites like Pangram, which claimed to be able to detect AI-generated writing. But the news dropped that Anthropic have started watermarking generated text, so that detection tools can more reliably know if something was touched by Claude. It works in this crazy way that really took me a minute to understand, so I'm going to try and break it down as simply as I can, so that perhaps it can click for you too. As context, I wrote a newsletter on tech, I've been doing it for five years, I'm also doing a research project on AI and shopping, and I'm just generally spending a lot of time messing around with these AI models to see how they work. Okay, so first on how watermarking does not work. It isn't adding any hidden characters or spaces, it's not going to be detectable by the human eye, and also critically it's not in the metadata. So if you were to say copy the text that's generated by AI and paste it into a plain text file or into a Google sheet, the mark will remain. But if you were to heavily paraphrase or rewrite the text, the mark may disappear because the words themselves are the watermark. Roll with me, how LLM's work is that they generate sentences one word at a time. They're using very complex statistics basically to predict what the next most likely word is in a sentence. And sometimes several words will work equally well, like there's effectively a tie between the words. And with watermarking, when that tie exists, the model will subtly lean towards one particular option. And this leaning over the course of many words and sentences and paragraphs becomes sort of pattern that a detector with a key can pick up. So you can think about it as sort of like an invisible signature or fingerprint. When I first read about this, I was like, "Oh, there's going to be certain words that become triggers for these detectors." And that's not the case. It's more like the model is following this invisible thread that gets woven through the passage. And that becomes the mark. So two final details on this. And so I think I've not released exactly how there's works. Google has, so I'm drawing on that. And if I'm getting it wrong, let me know. And critically, the tools to detect the watermarking are not available yet. However, the watermarking is rolling out now. So assume anything that you're generating with Claude today can be checked later. And that's today's lesson in air literacy. I think it's important that we all understand how these things work. Much to debate about this one, so let me know what you think in the comments.