Briefing Chat: Anthropic rolls out new AI watermark — will it make a difference?

**Benjamin Thompson** (0:04)
Hi, listeners, Benjamin here. Welcome to the Friday Show, where we talk about a couple of stories that have been in the Nature Briefing, which is, of course, Nature's daily email roundup of the latest science news. Here with me today is Lizzie Gibney. Lizzie, how are you doing?

**Lizzie Gibney** (0:17)
Hello, Ben, I'm good, thank you. We've just been yabbering about how much we enjoyed the eclipse.

**Benjamin Thompson** (0:21)
Oh, I enjoyed the eclipse so much. My colander got a lot of action, I tell you what. Right, well, you've got a story that couldn't be more different today to start off with, and it's about AI and AI-generated techs. Now, we live in a world, of course, where this is becoming a lot more prevalent, certainly in academia as well, that evidence suggests that it's getting used more and more. And there's often a question that people have like, oh, is this just AI-generated? And it can be quite hard to tell, but there may be ways to do so, and that's what you've been looking at.

**Lizzie Gibney** (0:51)
Yes, absolutely. So Anthropic, which of course makes the model Claude that many people use, has announced that for future versions of the model that come out after the 2nd of August, they will have an invisibly embedded watermark. So that is something that doesn't affect the actual quality of the text, like the user will not notice this at all, but it's effectively constructed in such a way that, at least for now, Anthropic and potentially in the future, anybody can tell if that was made by Claude or not.

**Benjamin Thompson** (1:22)
Right. So an invisible watermark, I'm going to guess, given our discussions on AI in the past, this is going to be statistics-based.

**Lizzie Gibney** (1:28)
Absolutely. So we don't know the exact details of how they did this, but we do know the details of how Google created its synth ID. They actually published a paper about it and made it open for anyone to use. And I am almost certain this is something very similar. So, I mean, with text, you don't have that many variables anyway. You can change, right? So it all comes down to word choice.
And in essence, they use an algorithm that will tweak the selection of the words in such a way that some words are going to, for that particular sequence, be more likely to be picked than others. You can then reverse engineer that from the text and see if there's the statistically observable kind of trace of that algorithm left in the word choice. And it is pretty reliable in terms of it can tell you, okay, this was generated by Claude. But I suppose there are lots of things that we'll talk about where it's not reliable in terms of, well, what does that mean? What exactly did you use Claude for? That it can't say.

**Benjamin Thompson** (2:18)
Well, you've read my mind. We certainly will talk about that. Let's talk about why this is being done. So, Anthropic and Claude then, and you mentioned Google too. There are reasons that they're introducing these things.

**Lizzie Gibney** (2:28)
Absolutely. So, this is quite a hard and fast reason, which is that the European Union demands it. So, the EU AI Act, which was drafted two years ago, but has become enforceable now. So, fines can be enforced from the 2nd of August. They say that it should be possible for people to know if something was generated by AI. And when it comes to text, that means a watermark like this. So, the companies are all trying to comply with the EU, basically. Google, as I say, have this Synth ID, and they've actually been using it for a few years now, but no one would necessarily know or bat an eyelid because there's no way of the general public to be able to detect it. So, Google themselves will know, but as anybody else, general public or user, you would not know. They have means to do that for images and audio, and OpenAI actually uses Synth ID from Google as well.
So, for video images and audio, that is possible, but for text, it's not yet.

**Benjamin Thompson** (3:18)
Okay. So, the difference then is that they can know if it's been made by a robot, but with the clawed version, that you potentially, as a user, could identify if something's AI-created.

**Lizzie Gibney** (3:27)
Yes. They have not yet created that, but they said they are working on it. So, they've signaled their intention to do that, to make some kind of tool, which again, is what the law asks for. So, presumably, the others will have to follow suit.