Anthropic explains how Claude marks AI-generated content

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content and has published how it is putting those commitments into practice, as a provider of both generative AI models and systems.

Two kinds of marks

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text. The company says this does not change the meaning, quality or readability of a response. Because the mark is part of the text itself, it travels with the text when it is copied and pasted elsewhere and may persist through some editing; watermarking is applied at model level, so it is present whichever Claude product produced the text.

For files, Claude attaches Content Credentials — signed provenance metadata that follows the open C2PA standard (Coalition for Content Provenance and Authenticity), which the industry uses to record content provenance. A signed label signals that a file was processed by Claude; supported types include PNG and JPEG.

Which models, and where

Claude models launched on or after 2 August 2026 support machine-readable marking at launch. Marking applies across Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, and wherever Claude is offered worldwide. Output accessed through AWS, Google Cloud or Microsoft Foundry also carries watermarks. For Claude Opus 5, text watermarking becomes gradually available on cloud partner surfaces starting 14 September 2026 and fully within one week.

Models released before 2 August 2026 fall under a transition period: Anthropic is adding watermarking support to them, with all of them covered by 2 December 2026.

Detection, and its limits

Anyone can check whether a file contains a Claude-issued Content Credential with the free Claude Content Checker. Watermark detection is in private preview for organizations covered by EU law — regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups — as well as enterprises with similar verification duties. Anthropic says access to the detection API will expand over time.

The company stresses the limits of these signals. A detected mark means content may have been processed by Claude, not that Claude wrote it: people often use the model to proofread, translate, summarize or convert files, and marked content can be modified afterwards. A missing mark proves nothing either — heavy editing or paraphrasing, very short passages, metadata stripped by conversion or screenshots, or an unsupported platform or file type can all remove it. Anthropic adds that developers deploying Claude in their own products should assess independently what Article 50 requires of them.