Anthropic has announced its intention to implement an invisible watermarking system for text generated by its Claude large language models, initially rolling out globally. This move is in response to the European Union's Code of Practice, which mandates that AI companies operating in its market identify AI-generated content. While the EU AI Act is the catalyst, Anthropic states the global deployment is due to current limitations in regional scoping.
The watermarking mechanism is based on Google DeepMind's SynthID-Text approach and operates during the text generation process. Unlike methods that add hidden characters or modify text post-generation, Anthropic's system subtly alters the source of randomness used when Claude selects subsequent tokens (words or sub-word units). When Claude has multiple plausible choices for the next word, the watermarking system uses a secret key and preceding words to influence the selection, leaving a statistical pattern that is imperceptible to human readers.
This statistical pattern allows a detector, equipped with Anthropic's key, to analyze a sequence of words and assess its consistency with choices Claude would make when watermarking is active. This process can estimate the probability that Claude generated the text. Anthropic confirms that internal testing indicates no practical impact on the quality, creativity, or readability of Claude's output. The company also states that the watermarking process does not require additional tokens and has a negligible effect on generation speed.
Certain types of content will carry less watermarking. For factual statements where there is only one correct answer, or for code where altering terms could break functionality, the watermark will not interfere with the model's choice. However, in parts of code where arbitrary choices exist, such as comments, watermarking may still be applied. The effectiveness of detection is influenced by text length and the entropy of the model's distribution; longer texts provide more evidence, and watermarking is less effective when the model almost always returns the same response.
Anthropic plans to offer a watermark detection API, which will estimate the likelihood of Claude's involvement in generating a piece of text. The company emphasizes that this API will indicate the probability of Claude's authorship, not definitive proof, and cannot identify text generated by other AI models that may use different watermarking methods or keys.
Models launched before August 2, 2026, fall under the EU's transition period, and Anthropic is working to integrate watermarking into these older models in the coming months. The company notes that light proofreading or minor grammatical changes to human-written text may leave insufficient Claude-generated material for reliable detection. However, a translation produced by Claude would carry a watermark, as Claude selects every word in the translated output.
