It could end up being your ideas with their watermark. #claudeai #anthropic #aiwatermark #LearnOnTikTok #aiprivacy

I'm going to explain what this Claude watermarking thing is and whether you should be concerned, spoiler alert, you should be a little concerned. Okay, Claude is watermarking by changing the distribution of tokens. So your tokens are drawn from a distribution. I don't know if you knew this. It's not a linear thing where they pick one token off the assembly line and that's the next token. It's actually a distribution. And so Claude pulls the next token from a probability distribution around what the next token should be based on tokens so far. That's how it's always worked. Now, when you are trying to put a watermark on something, what you're doing is you're shifting that distribution and you're shifting it a secret amount. And so there's a cryptographic hash. It shifts the mask of the distribution a little bit. And so it basically looks like the curve goes from here to here and you would never know because it doesn't change the semantic meaning. It's still inside the picture that you're trying to paint when you want the AI to come back with a story. And so it's still going to sound like the same story. The narrative is going to sound the same. You won't really be able to read it and tell, but very subtly, the distribution of which tokens have been picked has changed. And so the concern that people have is that when you do that, you open up a lot of other downstream risks that aren't well thought through. And by the way, the reason they say they're doing this is the EU regulation, but the EU regulation doesn't require them to go as far as they're going. And so I find that interesting. So anyway, they are doing next token prediction masking, which allows them to watermark something they can see absolutely everything the cloud has generated. What does this allow them to do? This allows them to one know exactly where cloud is being used anywhere in the world. Two, it allows them to know when companies are using cloud for documents, et cetera, and make the case to companies, maybe you don't need as many people. Three, it allows them to sort of have a piece of IP conversations if they want to have them and if the ULA allows them to. Now, that last one and Thropic may not be interested in, but the IP conversation is going to be really relevant for code bases everywhere because code bases are suddenly going to be like, OK, well, how much of this code was written by cloud? Well, now we can tell. Let's apply apply this cryptographic hash and let's see how much was written by cloud. And now you're going to have conversations about IP rights. You're going to have a conversation about how much of the code base is something that we want to use. Can we just recreate it? It's going to open up a whole mess. Now, is anthropic doing some of those things now? They are not. They say they're just doing this to comply with the U regulation. I am not saying that they are going to companies and saying, look at all of your documents that are using cloud. But it is certainly on the table for them to do that. It is certainly on the table for anyone to do that who has the cryptographic key, not just anthropic. They can go and sort of make that claim. And what's problematic about that is anyone who has worked in knowledge work will tell you if everything that goes through cloud, including my ideas, ends up with this cryptographic hash, it may look like it was cloud's work, but it was me, it was my ideas, it was my thinking, it was not clouds. And cloud may have put some semantic meaning on it or maybe arrange it in the sense, but it was my thinking was my ideas that got there. And so I think that the other risk for this is that people will start to miss perceive a cryptographic hash as authorship. It's not authorship. It never was authorship. This is a writing tool. It's a token prediction tool. That's what's going on. And as a bonus risk, one more thing, if you're technical, if you're shifting the cryptographic hash, why would your code be as clean and as elegant as it as it could be? It might not be. It might not be the cleanest, most elegant solution because you're shifting the cryptographic, you're using the cryptographic hash to shift the distribution. It might not be the cleanest solution technically to the problem. So there's a lot of issues with it. And yes, the EU did require watermarking. So some of it is required in the EU specifically. That doesn't mean it has to be worldwide. It doesn't mean it needs to be on every single thing that anthropic puts out. This was a choice by that. And it's a choice everyone should know about. Let me know what you think in the comments.