Anthropic has announced that future Claude models will include an invisible text watermark. This four-point overview covers what the watermark does, why it’s being added, and what it means for your use of Claude at UT.

• What is watermarking?
  A statistical signal embedded in Claude’s writing that helps indicate whether Claude likely generated or contributed to a piece of text. It does not add hidden characters, change wording, or affect the quality of Claude’s output.
• Why is Anthropic adding this now?
  The watermark supports compliance with the EU AI Act’s transparency requirements and reflects a broader industry shift toward accountable AI. Anthropic is applying it globally, not only for users covered by EU regulations.
• How does watermarking actually work behind the scenes?: Claude generates text one word at a time, often choosing among several equally valid options. Watermarking slightly adjusts that decision-making process, leaving a pattern that is undetectable to readers but identifiable with Anthropic’s detection key. Longer passages carry a stronger signal; short responses, heavily edited text, code, and constrained writing (like grammar fixes) carry little or none.
• What should you know as a campus user?
  Watermarking does not identify individual users, organizations, or conversation content, and it does not change ownership or usage rights for Claude-generated text. It has no effect on speed, cost, or model performance. Anthropic will also offer a detection API, and supported image files will carry separate transparency data through C2PA metadata.

Read the complete Claude watermarking announcement on Anthropic’s website.

UT AI Available tools, support, and Policies

Faculty, staff, and students can continue using Claude as usual under existing UT AI guidance. Visit the OIT Artificial Intelligence service page for information on approved AI tools, data classification guidance, and responsible use policies.

Anthropic, Claude models are available in the UT AI hub—oit.utk.edu/ai/aihub.