Anthropic is set to introduce text watermarking in future Claude models to enable the detection of AI-generated content, as the company moves to comply with requirements under the European Union’s AI Act.
The watermarking technology is designed to identify the likelihood that a piece of text was generated by Claude without visibly altering the content or adding hidden characters. Anthropic said the method is intended to have no practical impact on the quality, readability or creativity of Claude's responses.
The company said several other major AI providers are also implementing similar measures to comply with EU requirements for marking AI-generated content.
Unlike visible watermarks used on physical documents or images, Anthropic's approach does not add text, symbols or hidden characters to Claude's output. Instead, the watermark is embedded through the model's existing process of selecting the next word.
Large language models generate text sequentially, choosing from a range of possible words based on the preceding context. Where several choices are similarly plausible, the model can use randomness to determine which word is selected.
Anthropic's watermarking method changes the source of that randomness. It uses a secret key together with preceding words to influence these low-stakes word-selection decisions. Over a sufficiently long piece of text, the resulting pattern can be analysed using the corresponding key to determine whether the sequence is statistically consistent with text generated by Claude.
The company said the method does not force Claude to select words it would not ordinarily consider, but instead uses choices already available within the model's normal generation process.
Anthropic said its internal testing found no measurable impact on the content, creativity or readability of watermarked Claude responses.
The company also referenced research on SynthID-Text, the watermarking technique developed by Google DeepMind, which found no statistically significant difference in user ratings between watermarked and unwatermarked AI-generated responses. Anthropic said controlled testing with human evaluators similarly found no discernible difference in quality.
Watermarking will also not require additional tokens or increase the cost of generating Claude responses, according to the company.
Anthropic said the watermark will not contain identifying information and will not be traceable to a particular person, organisation or conversation. The system is intended to indicate whether content is statistically consistent with generation by Claude rather than identify the individual who generated it.
The move comes as regulatory requirements around transparency and AI-generated content develop globally. Anthropic said that, from August 2, the EU requires AI providers serving its market to mark AI-generated content, with other major model developers that have signed the relevant Code of Practice also expected to implement their own watermarking approaches.
The introduction of watermarking reflects a broader effort across the AI industry to develop technical mechanisms for distinguishing AI-generated content while preserving the quality and usability of generative AI systems.
For Claude users, the change is expected to remain largely invisible: the generated text will look and read the same, while specialised detection using the appropriate key can assess whether the underlying word-selection pattern is consistent with Claude-generated content.
