The short answer
Anthropic announced on August 11, 2026 that all Claude models released from August 2 onward embed invisible, machine-readable watermarks in every piece of generated text. Supported image files receive C2PA-standard digital signatures. The regulatory trigger is EU AI Act Article 50(2), which requires generative AI providers to mark synthetic output so it can be identified as machine-made. The scope is global — every Claude access point is affected, not just EU users.
For teams building products on the EU AI Act compliance path: the watermark satisfies the provider-side obligation, but your product still needs its own human-visible disclosure to end users. Watermarking is a backend signal; it is not a substitute for the transparency requirement placed on deployers.
What changed on August 2
Anthropic signed the EU AI Act's Code of Practice on Transparency of AI-Generated Content, committing to embed machine-readable marks in Claude output starting with models launched on or after August 2, 2026. Existing models will be retrofitted during a transition period.
The change covers every access point: Claude.ai, the Claude Platform API, Claude Code, Claude Cowork, Claude Tag, and Claude models accessed through AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. There is no regional carve-out — a user in California making an API call gets watermarked text just as a user in Frankfurt does.
The regulatory backdrop matters for understanding the scope. EU AI Act Article 50(2) entered into force on August 2, 2026. It places two obligations: providers of general-purpose AI must mark synthetic output at the model level, and deployers — meaning companies that embed Claude in their own products — must ensure end users receive a clear disclosure that they are interacting with AI-generated content. Anthropic's watermark satisfies the first obligation. Your product must satisfy the second independently.
How the two systems work: text vs images
Anthropic uses different techniques depending on content type.
Text watermarks are imperceptible signals embedded at the model level. They do not alter the words, tone, or formatting of the response your application receives. The mark is designed to survive copying and pasting — meaning text extracted from a chat interface or an API response and pasted into a document or email retains the signal. Anthropic says the watermark "may persist through some editing," which means light rewrites keep it, but heavy rewrites, translations, or format conversions (such as rendering to PDF and re-OCRing) may degrade or remove it.
Image watermarks follow a different path. Supported image files — SVG, PNG, and JPEG — receive digitally signed provenance metadata following the C2PA standard (Coalition for Content Provenance and Authenticity). C2PA metadata travels with the file as a signed manifest. Platforms and tools that support C2PA can read this manifest and confirm that the image was processed by Claude. Notably, many image editors and social media upload pipelines strip C2PA metadata, so the signal does not survive all distribution paths — a known limitation of the standard, not specific to Anthropic's implementation.
The Anthropic API documentation for the new Claude models includes details on how to query watermark detection, and Anthropic has indicated it will publish a verification API so third parties can check content provenance programmatically.
What the mark detects — and what it misses
The most important thing teams need to understand: the watermark signals that Claude had involvement in the content, not that Claude wrote it entirely.
This distinction matters operationally. A product that uses Claude only to proofread or translate human-written text will still leave a watermark on the result. A content moderation pipeline that uses Claude to flag inappropriate submissions will mark its outputs. Even a summarization step that passes human-authored text through Claude to produce a shorter version will result in a watermarked output.
This creates two practical edge cases:
- False positives for authorship claims: If a user submits content to a platform that runs a Claude-based watermark detector, the presence of a mark does not prove the user's submission was AI-written — it proves that Claude touched it at some point in its history. Academic integrity tools and plagiarism detectors that adopt this standard need clear policies on this limitation.
- Incomplete disclosure for AI-assisted workflows: If your editorial or content pipeline uses Claude for any step — including grammar checking, tone adjustment, or summarization — and you have not disclosed this to end readers, your disclosure gap is now technically detectable. For EU-facing media products, this has Article 50 implications.
Conversely, the absence of a watermark does not confirm that content is human-authored. Older Claude models, models from other providers, and heavily rewritten Claude output will not carry the signal.
What it means for US & EU software teams
The watermarking change sits at the intersection of three real workstreams for software teams: compliance, content pipeline architecture, and product disclosure design.
For compliance teams: If your product operates in the EU, Article 50(2) requires a visible disclosure to users that content is AI-generated or AI-assisted. The watermark enables machine detection but does not replace the user-facing notice. Review your product's disclosure language to confirm it covers Claude output, particularly if you use newer Claude models post-August 2. If you are working toward EU AI Act compliance more broadly, this is a good moment to audit which pipeline steps involve Claude and which outputs reach end users.
For content pipeline architects: C2PA metadata in images adds a new variable to your pipeline. If you process Claude-generated images through any tool that strips EXIF or proprietary metadata — resize utilities, CDN transformation layers, social media upload APIs — you will lose the C2PA provenance record. If your product depends on provenance tracking for compliance or user trust signals, verify that your image processing chain preserves C2PA manifests. Many do not by default.
For US product teams: There is no current US federal equivalent to Article 50, but the FTC's AI guidance on deceptive practices and California's AI disclosure bills are converging toward similar requirements. Anthropic's global implementation means your content is watermarked regardless of your users' jurisdiction. Getting the disclosure habit right now for EU compliance also positions you well for the US regulatory environment that is forming.
Industry context: Google DeepMind open-sourced its SynthID watermarking system in 2024, and YouTube and Substack have implemented similar AI content detection mechanisms. OpenAI has developed a watermark detector with a reported 99.9% accuracy rate but has not released it publicly. Anthropic's decision to implement globally under a formal regulatory commitment is the most explicit provider-level commitment to date.
Frequently asked questions
Which Claude models carry the watermark?
All Claude models released on or after August 2, 2026 include watermarking. Models released before that date do not carry the mark by default, though Anthropic says it will retrofit existing models during a transition period. The change covers claude.ai, the Claude API, Claude Code, and Claude accessed through AWS Bedrock, Google Cloud, and Microsoft Foundry.
Does the watermark affect the text my application receives from the API?
No. The watermark is imperceptible and does not change the words, formatting, or quality of the text your application receives. It is an invisible signal embedded in the content. Anthropic states that text quality is unchanged.
Is my EU-facing product automatically compliant with EU AI Act Article 50 because of this?
Not automatically. EU AI Act Article 50(2) places the disclosure obligation on the deployer — meaning your company — not only on the model provider. Anthropic's watermark satisfies the provider-level transparency requirement, but your product still needs a visible disclosure to end users that they are interacting with AI-generated content. The watermark is a backend signal; human-facing disclosure is a separate legal obligation on your organization.
Can the watermark be removed?
The text watermark is designed to survive copying and pasting and some editing. Heavy rewrites, format conversion, or translation may remove or degrade the signal. C2PA metadata on images can be stripped by many image editors and social media upload pipelines — a known limitation of the C2PA standard. Anthropic acknowledges both constraints.
Does AI-assisted editing — such as proofreading by Claude — also leave a watermark?
Yes. The watermark indicates that Claude had a role in processing the content, not that it authored the entire piece. Even tasks like proofreading, translation, or summarization performed by Claude will leave the mark. The presence of a watermark does not confirm that Claude wrote the full piece, and the absence of one does not confirm human authorship.
Sources
Fortune — Anthropic plans to add an invisible mark to AI text, August 11, 2026
The Decoder — Anthropic watermarks all Claude outputs globally, August 2026
Business Standard — Claude AI Watermark: How Anthropic Marks AI-Generated Text, August 11, 2026