Anthropic is now implementing text watermarking across all future Claude models. This update is a direct response to the EU AI Act, which mandates that developers label AI-generated content. The system uses a specific method known as SynthID-Text to identify patterns in the generated output. This process ensures that the model can be identified as the source of a text sample without altering the quality or readability of the content itself.
The technical approach is subtle. Large language models choose words based on probability. Claude typically selects from a range of valid options when constructing sentences. The watermarking system uses a key to influence these random selections in a way that creates an undetectable pattern. To a human reader, the text appears normal. There are no hidden characters or added tokens. The watermark only becomes visible when someone with the correct detection key analyzes the sequence of words to check for consistency with the model's decision-making process.
This implementation maintains the current performance and cost of the model. Because the watermark relies on low-stakes word choices where multiple options might be valid, it does not impact the accuracy of factual information or the logic of code. In instances where specific output is required, such as factual statements or precise programming commands, the watermark is not applied. This protects the utility of the model while meeting transparency requirements.
Anthropic intends to provide a detection API in the near future to assist with identification. It is important to note that this system does not track users or collect personal information. The watermark exists solely to verify the involvement of the model in generating a piece of text. As other major AI developers adopt similar practices under the EU Code of Practice, this standard will become a common feature of large language models operating in European markets.
While the technology is effective, it is not foolproof. Significant manual edits or complete rewrites of AI-generated content can remove the watermark pattern entirely. Additionally, the system functions more accurately as the length of the text increases. Shorter passages provide less data for the detection system to process, which limits its effectiveness for snippets or brief responses.