NPR's Michel Martin speaks with Fortune Magazine AI reporter Beatrice Nolan about the invisible watermark embedded in anything new models of Anthropic's AI assistant Claude processes.
Anthropic is adding an invisible watermark to new models of its AI assistant Claude. The change brings Claude into compliance with the European Union's AI Act. It will be applied worldwide. Here to explain how this will work is Beatrice Nolan. She covers AI for Fortune Magazine. Good morning.
MARTIN: So a lot of times, when I – if I get an early release of a film that I need to watch in order to prepare for an interview, it'll have a watermark. So I think of it as something that goes over a picture. Like, how does it work when we're talking about text that an AI chatbot is generating?
NOLAN: Yeah. It's interesting 'cause AI-generated text has been traditionally pretty hard to watermark, especially comparatively to images. Most AI-generated or AI-modified images already carry a type of watermark. Previously, the best we've had for text is these kind of probability checkers that make an assessment on how likely something is to be AI-generated based on style and word choice.
But what Anthropic is introducing is a little different. They explained over the weekend how it's going to work. It's essentially adding an invisible mark. And the way it does this is when AI writes a sentence, it's not just spitting out words. It's making a tiny decision based on pattern prediction after every single word about what word comes next. A lot of the time, there are several words that could work equally well. Like, if you think of a sentence like – the weather today was cold and – the AI could say gray or it could say overcast. Either works. Whenever the AI makes a decision like this, Anthropic is going to be adding a mark. And anyone with Anthropic's tool to identify this can see if Claude was involved in the text.
MARTIN: Interesting. So can it be erased?
NOLAN: It's not clear yet because they haven't really released the details of how the tool is going to check. Guys, essentially, there's a theory that you could probably get rid of it if you're heavily editing the text. But, like, a light edit of the text probably wouldn't get rid of it.
MARTIN: So is this happening because of the EU law?
NOLAN: Yes. This is happening because of the EU's AI Act coming into force. It came into force on August 2. This requires AI companies to mark AI-generated or edited content in a way that other systems can identify. I mean, notably, other companies, including OpenAI, Meta, Google, Microsoft, have signed this code of conduct, so they will likely be introducing similar watermarks as well.
MARTIN: But this is the first we're hearing of this with Anthropic. You do expect the other companies to follow.
NOLAN: Yes. To be compliant with the EU AI Act, they'll probably likely have to do something similar. The only company that hasn't signed the code of conduct is SpaceX AI, which makes the AI chatbot Grok. So we'll probably be seeing a lot more of this. Anthropic's just the first to kind of lay out exactly how they're going to go about it.
MARTIN: So the supporters of the EU's AI Act said they wanted more protections from harm. They wanted greater guarantees of honesty and transparency. Do you have a sense yet – I understand this is very new – that they are satisfied by Anthropic's approach?
NOLAN: Yeah. I mean, the EU AI's transparency rules are just kind of built around that idea. Like, if something is AI-written or AI-altered, people should be able to find out. This is, like, worrying people. As AI content gets harder to spot, it gets easier to pass off as human work. So it's a bit too early to see if people are satisfied by it until we get to see the results of it. But it seems like it is going to work on a lot of stuff, so I would imagine so.
MARTIN: That's Beatrice Nolan. She covers artificial intelligence for Fortune Magazine. Thank you so much.
Copyright © 2026 NPR. All rights reserved. Visit our website terms of use and permissions pages at www.npr.org for further information.