Claude 与 Grok 本周起为文本加 AI 水印，视频讲解水印原理与识别方法

以下是视频口播的自动转写，可能有错字。

More ais are adding watermarks in their text this week. So let me give you the full breakdown on the 3rd of november anthropic start rolling out.

Claude's invisible text watermark to four more models. Now this already exists in their most advanced and expensive models like opus 5.5.

Then on the 7th of november grok start watermarking text by default through their api and i'd expect that to roll out in the chat experience pretty soon but they haven't announced that yet.

So this table shows you which of the big labs are watermarking text, basically all of them. And it's happening in all the newer models and some of the older ones too.

Now, this is driven by eu european law, but OpenAI are the only lab that have an eu only rollout. Everyone is just rolling this out globally.

So google and microsoft haven't explicitly said that they haven't said anything about geography, so we can only assume it's global, which makes sense from a transparency point of view.

Now, as we've talked about this text watermark across most of these systems is about word choices. So models like gpt and claude break text down into tokens.

A token can be a whole word or part of one, and each token has a number. So in gpt, my name will with a capital w is token 17,886.

Now imagine we were making our own watermark, we could favor token ids where i don't know, let's say the digits add up to a multiple of three and where there's a choice will always have a bias towards those token ids.

Now you wouldn't see much of a pattern over a short sentence but over a longer piece of text. You'd start seeing those token ids more often than you'd expect just by chance.

Now, claude goes a bit further and uses a secret key, plus the last few tokens that it wrote, to help pick the next one.

And so that stops it from getting hung up on choosing the same words again and again and becoming repetitive. And to detect that kind of watermark, you need a detector, and you need that secret key.

So that you can look for that pattern, and we know that anthropic and OpenAI are issuing a detector at some point that will have that key.

So that means that rearranging or slightly rewriting some of the text doesn't remove the watermark. The best thing to do is write most of your text yourself, which is doing your brain a favor.