Impact of Claude's Watermarks on the Legal Sector
Anthropic, the developer of Claude, is set to implement watermarks in compliance with the EU AI Act, which mandates that AI generated outputs must be identifiable. This requirement applies to all large language model (LLM) creators operating within the European Union.
The watermarking mechanism employed by Claude involves the insertion of specific words in a manner that aligns with a key designed to indicate whether the text has been generated by AI. Rather than a visible watermark, the system camouflages signals within the normal text, making detection possible only for those with the appropriate key.
Claude generates text by selecting one word at a time from a list of potential candidates, guided by the preceding context. The watermarking process utilises low stakes choices made throughout the text to create a discernible pattern in the responses. While the randomness of word selection remains, the source of that randomness shifts to incorporate the key, allowing for the identification of AI generated content.
This innovative approach builds upon previous work by Google DeepMind and is expected to include a watermark detection API in the near future. Importantly, the watermarking does not disclose sensitive information or identify the creator of the document.
A critical consideration is whether the watermark alters legal responsibility for documents. According to Anthropic, the watermark does not affect ownership or authorship rights, nor does it change the user's rights under their terms of service. It serves solely as an indicator of AI involvement in content processing.
As the EU AI Act applies to all LLMs, including those from smaller or international developers, compliance will be essential for any AI outputs used within Europe. This is particularly relevant for law firms utilising open weight models that may not yet have adopted similar watermarking systems.
It is also noteworthy that the presence of a watermark does not guarantee the accuracy of the content. Instead, it merely indicates the method of production. Perspectives on AI generated work are evolving, with some viewing watermarks as a potential cause for concern, while others may see them as a standard practice in modern document creation.
Ultimately, the responsibility for implementing watermarks lies with LLM developers, not law firms or legal technology companies. However, legal professionals must be prepared to address any inquiries regarding watermarks from clients or courts, particularly in sensitive situations where the extent of AI involvement may be questioned.