Forum Discussion
Claude Watermarks
With Claude now including watermarks on the text it generates, are those watermarks being passed into Copilot?
Since we do not know when Copilot uses Anthropic or OpenAI LLMs, how can we tell if our content is watermarked?
3 Replies
Claude can be available in Microsoft 365 Copilot when the tenant permits Anthropic as a Microsoft subprocessor, and Auto mode may route prompts between eligible models. However, Microsoft’s documented watermark feature does not describe invisible watermarks in generated text. Microsoft documents visual or audio watermarks for generated video and audio, optional visual watermarks for images, and provenance metadata for images. It provides no admin or user method to inspect plain text and determine that Claude produced it. If model identity matters, select Claude explicitly instead of Auto and preserve the conversation record, but do not treat copied text as carrying a detectable signature. An administrator can review whether Anthropic access is enabled in the Microsoft 365 admin center. For compliance, classify and review output based on its content and business use, not an assumed watermark. Ask Microsoft Support for written confirmation if a regulatory process requires model-level attribution.
- CloudAdvisorsCopper Contributor
Thank you for the response but that didn't answer the question. I'm talking about the Claude watermarks, not the Microsoft watermark feature. It isn’t hidden characters or attached metadata. It is a statistical pattern created through Claude’s word choices, so Microsoft doesn’t need to “pass” a separate watermark field.
All of my research indicates yes —when Copilot directly returns text generated by a watermark-enabled Claude model, Claude’s watermark should travel into the Copilot response. This is separate from Microsoft’s own watermarking and provenance features for images, audio, and video.
- CloudAdvisorsCopper Contributor
It's important to note that Anthropic made these change to be compliant with a new GDPR requirement, so we'll soon see this same "feature" come from all the LLM vendors.