News Summary
- Anthropic confirmed on August 11, 2026, that it will watermark Claude-generated text and mark files to comply with the EU AI Act’s Article 50(2) Code of Practice.
- All Claude models released on or after August 2, 2026, support machine-readable marking at launch. Older models are being updated during the EU’s transition period, which runs until December 2, 2026.
- Text watermarks are imperceptible to readers and travel with content when copied and pasted. Files receive cryptographically signed C2PA metadata.
- The marking applies worldwide at the model level, covering Claude, Claude Code, Claude Cowork, Claude Tag, and the Claude Platform API, including through cloud partners AWS, Google Cloud, and Microsoft Foundry.
- Anthropic explicitly states that a detected mark is not conclusive proof that Claude authored a piece of content, and that marks may be absent after heavy editing or paraphrasing.
Anthropic announced on Monday that it will embed invisible watermarks in text generated by Claude models and attach signed provenance metadata to files, a move driven by the European Union’s new transparency requirements for AI-generated content.
The company updated its support documentation to confirm that all Claude models launched on or after August 2, 2026 , the date the EU AI Act’s Article 50(2) obligations took effect , will carry machine-readable marks from launch. Anthropic said it is also working to retrofit older models during the transition period allowed under EU law.
Two Marking Systems
Anthropic is deploying two distinct techniques. For text, the company says it weaves an imperceptible watermark directly into the generated output itself. According to the updated support page, users will not see the mark, and it will not change the meaning, quality, or readability of Claude’s responses. Because the watermark is embedded in the text rather than stored as metadata, it will travel with the content when copied and pasted, and may persist through some editing.
For files, Anthropic is using the Coalition for Content Provenance and Authenticity (C2PA) open standard. Generated files in supported formats , specifically .svg, .png, and .jpg , will carry cryptographically signed provenance metadata. Unlike the text watermark, C2PA metadata can be stripped through format conversion, screenshots, or re-saving.
Global Scope, Model-Level Enforcement
The marking is applied at the model level, meaning it is present regardless of which Claude product or surface generates the output. Coverage extends beyond Anthropic’s direct consumer interface to include Claude Code, Claude Cowork, Claude Tag, and the Claude Platform API. Cloud customers accessing supported models through third-party platforms , specifically AWS, Google Cloud, and Microsoft Foundry , will also receive marked output.
Notably, Anthropic chose to deploy the system worldwide rather than limiting it to the European Union. The company said marking will apply “wherever Claude is offered, worldwide.” The EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content does not explicitly require non-EU deployment.
Limitations and Caveats
Anthropic was explicit about the limitations of its approach. The support page states that “a detected mark provides a signal that content was processed by Claude, but is not fully conclusive.” Marks may be absent if content was heavily edited, paraphrased, or generated by a model released before the August 2 cutoff. Short passages may also provide too little material for reliable detection.
The company has not yet published detailed technical documentation explaining how the text watermark is embedded or detected. It said it will share detection mechanisms in forthcoming documentation, but did not provide a release date. The support page also does not disclose the watermarking algorithm, false-positive rates, or specific passage-length thresholds.
Anthropic additionally cautioned that a detected mark does not necessarily mean Claude authored the underlying material. If a user pastes human-written text into Claude for translation, summarization, or editing, the output could still carry a Claude mark.
Regulatory Context
The EU AI Act’s Article 50(2) requires providers of generative AI systems to mark synthetic outputs in a machine-readable format so they are detectable as artificially generated. The obligations took effect on August 2, 2026. Generative AI systems already on the market before that date have until December 2, 2026, to meet the machine-readable marking requirement under a transition agreement.
Anthropic has signed the EU’s Code of Practice on transparency of AI-generated content as both a provider of generative AI models and a provider of generative AI systems. According to the European Commission, approximately 190 companies and organizations had signed the code by the end of July 2026. Other signatories include Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia.
For businesses and developers building on top of Claude, Anthropic noted that they must “independently assess what Article 50 requires of your products and services.” The company said it will share technical guidance on marking and detection as it becomes available.