Claude will watermark AI-generated text: What it means

Anthropic, the company behind AI chatbot Claude, has started adding an invisible watermark to text generated by its AI models. The move is intended to make it easier to establish whether Claude was involved in producing a piece of text, as AI-generated content becomes increasingly difficult to distinguish from human writing.

The company says the watermark is being introduced globally to comply with the European Union’s AI Act, which requires AI-generated content to be marked in a machine-readable way. Anthropic said it is applying the system globally because it does not yet have “a durable way to scope it by region.”

How does the watermark work?

The watermark is not a visible label or a string of hidden characters. It is built into the way Claude selects words while generating a response. Large language models generate text one word at a time, choosing between several possible options. When there are multiple words that would make sense in a sentence, Claude’s watermarking system uses a key and some of the preceding words to determine which option is selected.

Anthropic said these choices leave behind a pattern that is invisible to a reader but can be identified by someone with the relevant key.

The company said the watermark “does not have any practical impact on the quality or content of Claude’s outputs”. It also said that the difference between watermarked and unwatermarked text “will not be distinguishable to readers”. There are no extra characters or tokens added to the response, meaning the watermark does not make Claude more expensive to use or noticeably slower.

Can the watermark survive editing?

The watermark is designed to remain in the text when it is copied and pasted and may survive some editing. It is not, however, impossible to remove. Anthropic says a complete rewrite in which every word is replaced will remove the watermark.

The company also makes an important distinction about what the watermark can actually prove. It said, “A watermark can only determine that Claude was likely involved with the content at some point.” That means a watermark does not necessarily prove that Claude wrote an entire article, essay or document.

If a person writes something themselves and asks Claude to make substantial edits, translate it or summarise it, the resulting text could contain a watermark. Anthropic says the more text Claude generates itself, the more decisions it has to make and therefore the more scope there is for the watermark to appear.

Why does this matter?

The distinction could become important for schools, workplaces and publishers that use AI detection to assess whether someone has relied on generative AI. Anthropic's system is different from conventional AI detectors. Those tools generally examine the writing itself for patterns associated with AI-generated text. Anthropic's system puts a signal into the text while Claude is generating it.

The company is also developing a detection tool. Anthropic said it will “soon be offering a watermark detection API”, although details of how it will work have not yet been released. 

What does it mean for users?

For most Claude users, there will be no obvious change. Text will look and read the same, and users can continue to copy, paste and edit it. A piece of Claude-generated writing could carry a machine-readable signal that allows its origin to be checked later.

Anthropic also says the watermark does not contain information that can identify the user, their organisation or their conversation. “There’s nothing in the watermark, or its key, that would allow anyone to recover any information about the user, their organization, or their chats with Claude,” the company said.

The move is part of a wider push to make AI-generated content traceable. Anthropic says several other major AI providers have also signed the EU's Code of Practice on Transparency of AI-Generated Content and are working on their own watermarking systems.