# Anthropic will watermark text from future Claude models worldwide > Source: > Published: 2026-08-10 23:33:05+00:00 Anthropic will embed invisible, machine-readable watermarks in text generated by new Claude models, applying the marking worldwide across its consumer products, developer platform and cloud distribution channels. The policy covers Claude models launched in the European Union on or after August 2nd, 2026, according to [Anthropic's help-center notice](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content). Models released before that date are subject to a transition period, and Anthropic says it is working to add marking support to them. The plan drew wider attention Monday after [M1 (@M1Astra)](https://x.com/M1Astra/status/2086898041882030353) summarized it in a thread on X. Anthropic's implementation turns a European legal obligation into a global product default. Supported models will mark output from Claude, Claude Code, Claude Cowork, Claude Tag and the Claude Platform API wherever those services are offered. The watermark will also apply when customers access supported models through Amazon Web Services, Google Cloud or Microsoft Foundry. That scope matters for developers building products on top of Claude. The mark is applied at the model level rather than added by the Claude website or another interface. A developer cannot avoid the marking simply by switching from Anthropic's API to a supported cloud provider. ### The watermark travels with copied text Anthropic describes the watermark as an imperceptible pattern woven directly into generated text. It is separate from file metadata, remains invisible to the reader and is designed to travel when text is copied and pasted. Anthropic says the pattern may also survive some editing. The announcement does not specify the encoding method, detection thresholds or how developers and outside parties will access detection tools. Anthropic says it will publish technical documentation covering detection. Those details will determine whether the system can reliably identify Claude output after common workflows such as editing, formatting, translation and combining generated passages with human writing. Anthropic already identifies several limits. Heavy editing, paraphrasing, translation or mixing a Claude passage into other text can make the mark undetectable. Very short passages may contain too little material for a reliable signal. Text generated by older Claude models may remain unmarked until those models receive support. Detection also cannot establish authorship. A document may carry a Claude mark after someone uses the model to proofread, translate or summarize human-written material. Anthropic says a positive result indicates that Claude may have processed the content, rather than proving that Claude originated every idea or sentence. A negative result similarly cannot prove that a person wrote the text. ### Files get a separate provenance layer Anthropic is taking a different approach to generated files. Supported SVG, PNG and JPG files will include digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, standard. That record can indicate that Claude processed a file and whether the signed file was subsequently altered. File provenance depends on the surrounding platform and format. Anthropic says signed metadata may not be available through every cloud partner or product feature, and metadata can disappear when a file is converted, re-saved or captured in a screenshot. Text watermarking is intended to cover the gaps created when metadata is stripped from copied prose. ### An EU mandate becomes a worldwide policy Anthropic adopted the system after signing Section 1 of the European Union's [Code of Practice on Transparency of AI-generated Content](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content). Article 50 transparency requirements under the EU AI Act took effect on August 2nd, requiring covered providers to make generated audio, images, video and text detectable in a machine-readable form where technically feasible. The code is voluntary, while the underlying Article 50 requirements are legal obligations. Signing gives providers a recognized route for demonstrating compliance. The European Commission's [current signatory list](https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content) also includes Google, Meta, Microsoft, Mistral and OpenAI among the organizations committed to the provider section. Anthropic could have limited its implementation to European users. Instead, every supported Claude deployment will carry the same model-level marking. That decision avoids maintaining separate marked and unmarked versions across regions and cloud platforms, while making the EU's transparency rules part of Claude's worldwide technical architecture.