Hey there,

One of the most commonly mentioned downsides of AI is that it can be nearly impossible to detect when something is AI-generated. This week, Anthropic has introduced a new watermark on all Claude created content to help address this concern. Unlike watermarked images or videos, this new system is invisible to users and meant to be a machine-readable watermark. Anthropic claims that this new watermarked text doesn’t impact quality or readability. This week we look at what Claude is actually doing, why this is happening, and what this means for you. 

What Claude is actually doing

This update brought two distinct changes to Claude: watermarked text and metadata for AI created files.

The text watermarks are invisible patterns within the writing itself. This means that it isn’t simply a hidden tag stuck on the text when you copy and paste it. Instead the words themselves are the watermark that signifies something as “AI-generated”. That ensures that the only way to remove the watermark is to heavily edit and personalize the text and make it your own. The watermark is applied at the model level itself so it doesn’t matter if you’re in the Claude app, the API, Claude Code or Cowork. All models released after August 2nd have this new feature, and there doesn’t seem to be a way to disable it. 

The second change is that Claude adds unique metadata to all the visuals it creates. This new tag shows that the file was processed by Claude and can indicate whether it has been altered since. Unlike the text watermark that can’t easily be removed, this is only a label attached to a file so it can be stripped away much easier. Currently this only impacts certain visual file types and Anthropic hasn’t stated whether this will be expanded to things like documents or spreadsheets. 

Why now: the EU AI Act

The short answer is because of regulations. These new watermarks are a direct result of the EU’s new regulations around AI transparency. The law requires AI companies to mark their content in a way that other systems can identify it. This law will impact all AI companies operating in Europe, but doesn’t necessarily guarantee that these changes will carry over into other regions. It’s important to note that Anthropic decided to roll out this feature globally. This decision also seems to be part of a larger trend within the AI space where companies are creating ways to easily identify their AI-generated content. Google has made a big push this year with their SynthID that marks and identifies all Gemini content. 

What it actually means for you

This is the most important aspect of this update and the answer depends on how you use Claude.  

Because the watermark only shows up on something that Claude writes itself, you can still have it look over your work without needing to worry about a watermark appearing. If you paste in an email and ask “does this sound okay”, that email will stay unmarked unless you have Claude rewrite it. It's completely different if you just take an email Claude writes and send it out because that will be watermarked. 

There’s also a length threshold for the watermark. Anthropic explains that short passages may not carry enough of the pattern to be detected. This makes sense given that the watermark probably needs to be applied across a large chunk of text before it can be reliably identified. And while we don’t know the exact lengths yet, one sentence probably isn’t enough to create a valid watermark, while a full page will. 

My Thoughts

AI-generated content having a watermark has been something that both supporters and skeptics of AI have been requesting for a while. We are constantly surrounded by media of all sorts so having a way to tell if something is created by AI makes a lot of sense. I don’t really see any downside to having watermarked text or visuals, especially because we currently don’t have a public tool to detect these watermarks yet. When those detection tools do arrive, the real questions start. Like does it matter if your colleague sent you an AI-generated email? Personally I don’t think it does. I believe that using some level of AI is and will continue to be a common practice going forward. I think this new watermark will encourage people to be more honest about what is AI-generated and what are their own thoughts and work.