Anthropic adds invisible watermarks to Claude AI output

Anthropic has begun embedding invisible, machine-readable watermarks directly into text generated by its Claude AI models, a move driven by new transparency requirements under the European Union's AI Act. The company detailed the system in a support page published this month, confirming that Claude models launched on or after August 2, 2026 support machine-readable marking from launch.ndtv+1

How the Watermarks Work

The system uses two complementary techniques. For text, Claude weaves an imperceptible watermark into the generated output at the model level — not as metadata but as part of the text itself. According to Anthropic, the watermark "doesn't change the meaning, quality, or readability of Claude's response" and can travel with the text when it is copied and pasted, potentially surviving light editing. For supported file types including PNG, JPG, and SVG, Claude attaches digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity (C2PA) standard, an industry-backed framework for recording a digital asset's origin.cyberpress+2

The marking applies worldwide — not just within the EU — across Claude's API, Claude Code, Claude Cowork, and Claude Tag, as well as through cloud partners including Amazon Web Services Amazon.com, Inc. ↘ 2.24%, Google Cloud Alphabet Inc. ↘ 3.58%, and Microsoft ↗ 0.52% Foundry. Anthropic said it is working to retrofit older models released before August 2 during a transition period that extends through December 2, 2026.claude+3

What the Watermarks Cannot Do

Anthropic has been careful to note what the watermarks do not prove. A detected mark signals that "content was processed by Claude, but is not fully conclusive," the company said. Text submitted to Claude for proofreading, translation, or summarization can also carry a watermark even if a human wrote the original material. Conversely, absence of a mark does not mean AI was uninvolved — heavy editing, paraphrasing, short passages, or output from older models may lack a detectable signal. For files, provenance metadata can be stripped through format conversion, re-saving, or screenshots.aiweekly+4

Anthropic said it is still developing tools to allow users and third parties to detect the watermarks, with technical documentation described as "forthcoming".ndtv+1

EU AI Act Compliance

The watermarking rollout is part of Anthropic's commitments under Article 50 of the EU AI Act, which became enforceable on August 2, 2026. The provision requires providers of generative AI systems to mark outputs in a machine-readable format that is detectable as artificially generated. Non-compliance carries fines of up to €15 million or 3% of worldwide annual turnover. AI systems already on the market before August 2 have until December 2, 2026 to meet the marking requirement.cloudsecurityalliance+2