How do AI watermarks work?

Anthropic's move to introduce an invisible watermark on content processed by its models, like the popular Claude, is a significant development - it's got major implications for both AI practitioners and users. This watermark, which will be rolled out globally across all new models, is essentially a response to the European Union's AI Act, requiring AI system providers to watermark any AI-generated or manipulated content. But here's the thing: this approach has raised some eyebrows, with concerns about its effectiveness and potential for misuse.

The way the watermark works is pretty clever - it biases the model's word choices in a pattern that's spread out across the entire document, so it's only detectable in aggregate, and you need the right tool to spot it. Anthropic says the marks will stick with the text even when it's copied and pasted elsewhere, and they might even persist through some editing - but they can be destroyed if the watermarked text is fed into another chatbot system that edits the text. And let's not forget about image and video content - those watermarks can be removed using screenshotting, recording, or metadata editing tools, which kind of undermines their purpose.

It's also worth noting that there's a lot of potential for misinterpretation here - the watermark isn't exactly informative, so a detected mark only tells you that the content was processed by Claude, but it's not conclusive proof. This lack of clarity could lead to some confusion among the general public, who might not fully grasp the difference between processed text and wholly generated text. As Anthropic plans to share more details about how to detect these marks, it's crucial for AI practitioners and users to understand the limitations and potential drawbacks of these watermarks.

To make sense of this complex issue, it's a good idea to stay informed about the development and implementation of AI watermarks - and be cautious when using AI-generated content. By getting a handle on the tech behind these watermarks and their potential limitations, individuals can make more informed decisions about their use and potential consequences. And since the EU's law requires AI system providers to offer technical support for detecting watermarks, we can expect more information and guidance on this topic in the future - which should help clear up some of the confusion.

Source: Ars Technica