In recent years, the way we consume content online has changed dramatically. Thanks to the rise of AI, it's harder than ever to know if what you're reading has been written by a person, or generated by a Large Language Model (or LLM) like ChatGPT or Claude.
While there are certain tells – from turns of phrase to grammatical conventions – that indicate AI's involvement, Anthropic is about to draw clear lines around Claude's output. Moving forward, text generated by Claude will be watermarked. But if the word watermark brings to mind images of a large, easy-to-see logo appearing across Claude text, think again. This watermark will actually be subtly embedded in text, imperceptible to the human eye. Here's what you need to know.
What is the Claude watermark?
According to Anthropic, in this context, watermarking refers to certain patterns in Claude's responses. Importantly, the company says "that pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it."
The pattern relates to the way Claude chooses words within a sentence. Anthropic explained that, "instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick."
In other words, while the words Claude generates are still random, the watermarking system introduces a pattern into those choices. Anyone with the key can then analyse a sequence of words to determine whether it is consistent with the choices Claude would have made using that key, therefore making it likely that the text was generated by Claude.
What does the watermark prove?
In short, whether Claude was involved in the development of the text in question. It doesn't distinguish between whether the text was entirely written by Claude, or just heavily edited. By the same token, if a user edits a piece of text significantly, the watermarking may be removed.
Can I use it to determine if a text was written by AI?
According to Anthropic, this is coming soon. Per its website, Anthropic "will soon be offering a watermark detection API. We’re in the process of working out the details of its implementation."
The watermark is a response to the European Union's AI Act
Claude's watermarking initiative is in response to the EU Code of Practice on Transparency of AI-Generated Content. 190 AI providers signed the act in July 2026, agreeing to mark and label AI-generated content.
The Code of Practice was developed under the EU AI Act, which aims to establish rules around the development and use of artificial intelligence, including greater transparency around AI-generated content, in order to address risks of deception and manipulation.
Will other generative AIs include watermarks too?
At this stage, other major LLMs like ChatGPT and Gemini have signed the agreement but have yet to respond on whether they will be implementing similar tactics.
