Anthropic adds invisible watermarks to Claude text output worldwide under EU AI Act Article 50
Anthropic has begun embedding invisible statistical watermarks in text generated by its Claude models and attaching C2PA metadata to image files, applying the marks worldwide to comply with Article 50 of the EU AI Act that took effect on 2 August 2026.
Anthropic's watermarking commitment
Anthropic has begun attaching machine-readable marks to text and images generated by its Claude models, applying them to users worldwide rather than only in Europe. The company confirmed in a support document that it signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-generated Content, making it one of the first large AI developers to set out operationally how it intends to comply with the EU transparency rules that took effect on 2 August 2026. The marking spans the company's products, including its developer API, the Claude apps, Claude Code, Claude Cowork, and Claude Tag, and is being applied globally. The system also functions wherever Claude is offered, including via the cloud services of Amazon, Google, and Microsoft.
How the watermark works
For text, Claude embeds an imperceptible watermark directly into the generated output through a statistical pattern in how the model selects words or tokens. Anthropic says the watermark does not change the meaning, quality, or readability of a response, and remains in place when text is copied and pasted or lightly edited. The company has not disclosed the precise technical recipe, but the approach works at the token level: the model is nudged to prefer certain tokens, leaving a statistical signature that a detector, knowing the key and algorithm, can reconstruct. The method works better on long, discursive text and worse on code, formulas, citations, numbers, and short responses, where the model has less freedom in word choice.
For files, Anthropic attaches signed provenance metadata to supported image formats including.png,.jpg, and.svg, following the open C2PA standard maintained by the Coalition for Content Provenance and Authenticity. The same standard has been adopted by Adobe, the BBC, and several camera manufacturers to label image origin.
- Article 50 of EU AI Act takes effect, requiring machine-readable marks on AI-generated content
- Anthropic announces invisible text watermarks and C2PA file metadata for Claude models, applied worldwide
- Deadline for AI providers to bring legacy models into compliance with EU transparency rules
Limitations and ambiguities
Anthropic highlighted two key limitations. First, content may trigger a detected mark even if Claude was used solely to proofread, format, or translate human-written copy. The company stated that detecting a Claude watermark means the text may have been processed by Claude but does not confirm the source.
Claude may not be the original author. People often use Claude to proofread, translate, summarise, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.
Second, if text is heavily rewritten, mixed with other copy, translated, or simply too short, the watermark may not be detectable. Anthropic is also developing tools that will allow users and third-party companies to search for the mark, with more technical details promised at a later stage.
EU rules and industry response
The changes are designed to meet Article 50 of the EU AI Act, which since 2 August 2026 requires providers of AI systems to add machine-readable markings to content generated or manipulated by AI. The obligations divide responsibility between providers, who build and supply AI models, and deployers, the larger set of organisations that use those systems to create and publish content. Violations risk drawing stiff fines. AI providers have until 2 December 2026 to bring legacy models into compliance.
Anthropic is among approximately 190 organisations that signed a voluntary EU code of good practice late last month, with others including Google, Meta, Microsoft, Mistral, OpenAI, Black Forest Labs, and Synthesia. Google has developed SynthID, a watermarking and detection technology including SynthID-Text, and is pushing for industry giants like Apple, Nvidia, and OpenAI to adopt this standard. Other platforms have accelerated their own AI identification measures: LinkedIn is testing a "seems like AI slop" button, Substack partnered with Pangram for AI content detection, and Snap stopped promoting AI-generated video in its main feed. Anthropic has not yet released a watermarking detection tool as called for in the EU code of good practices.
