Anthropic explained how watermarks will work in Claude's texts
Anthropic has released details regarding the implementation of watermarks in text generated by its chatbot, Claude. The company is implementing this mechanism to comply with the Transparency Code of the EU’s Artificial Intelligence Act, which requires systems to identify content created by AI, according to TechCrunch.
According to Anthropic, Claude will generate a hidden pattern in its responses by choosing between equivalent word options. For example, the model might select different words to describe cloudy weather. To the reader, text marked in this way will look no different from ordinary text, but it can be detected by anyone with the key to decode the watermark.
The company stated that it will use the SynthID-Text approach, which the Google DeepMind team introduced in 2024, and also plans to release an API for detecting watermarks. Anthropic emphasized that this technology differs from AI text detection systems, which look for characteristic linguistic patterns and other stylistic features.
For more breaking news, follow the UA.News Telegram channel.
According to Anthropic, minor edits are unlikely to remove the watermark completely. However, completely rewriting the text—replacing every word—can eliminate it. If Claude has only proofread or made minor edits to human-written text, the likelihood of the watermark being detected will depend on the extent and intensity of those edits. With light editing, almost all words will remain human-written, so the watermark will have little to latch onto.
In program code, the opportunities for flagging will be more limited, since the model must produce a functional result and cannot always freely choose between different options. At the same time, the technology can be applied to arbitrary elements of the code, including comments, with minimal impact on the code itself. Anthropic also noted that other major model developers who have signed the same code of conduct will be implementing their own watermarks.
