Content Credentials
EU AI Act · Article 50(2) · Code of Practice
How Claude marks
what it makes
Anthropic has signed the European Union's Code of Practice on transparency of AI-generated content. From August 2, 2026, Claude models launched in the EU mark what they generate — an invisible watermark woven into text, a signed provenance label bonded to files — wherever Claude is used, worldwide.
- Text · embedded watermark
- Files · C2PA provenance
- Every Claude surface, worldwide
Claude drafted this passage. Somewhere inside these ordinary-looking words, a quiet watermark is already at work — invisible as you read, yet ready to be found by anyone who knows how to look.
The commitments
Four promises,
signed and dated
Anthropic signed the Code of Practice as a provider of both generative AI models and generative AI systems. Signing was optional; the commitments are concrete.
-
New models mark from day one
Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. Generated text carries an embedded watermark, and generated files include digitally signed provenance metadata where supported.
-
Marking everywhere Claude appears
The mark applies across Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag — worldwide, including when models are accessed through cloud partners AWS, Google Cloud and Microsoft Foundry.
-
Detection you can build on
Anthropic will support users and third parties in detecting Claude's marks — so a watermark isn't a signal only Anthropic can read. Details will arrive in forthcoming documentation.
-
Existing models are in transition
Claude models released before August 2, 2026 are in a transition period, with marking support in progress. The date applies to models launched in the EU on or after it.
The milestone
August 2, 2026 — marking at launch
—
days
Checking the calendar…
Every Claude model launched in the EU on or after this date ships with machine-readable marking.
-
2026
The Code is signed
The Article 50(2) Code of Practice on Transparency of AI-Generated Content is published by the European Commission. Signing is optional — Anthropic signs as a provider of generative AI models and generative AI systems.
-
August 2, 2026
Marking at launch
New Claude models launched in the EU support machine-readable marking from their first day: embedded watermarks in generated text, signed provenance metadata in supported generated files.
-
Transition period
Earlier models catch up
Models released before August 2, 2026 gain marking support as work continues. Until then, their output may carry no mark at all.
-
Forthcoming
Detection docs & guidance
Technical documentation for detecting Claude's marks, and guidance to help builders meet their own transparency obligations, are on the way.
Where you meet Claude
- Claude Platform API
- Claude
- Claude Code
- Claude Cowork
- Claude Tag
Supported Claude models
Marks are applied at the model level — the same marks, on every surface.
Worldwide — wherever Claude is offered, not only in the EU
Cloud partners
- AWS
- Google Cloud
- Microsoft Foundry
Embedded watermarks apply through cloud partners. Signed provenance metadata depends on each platform's supported features.
How the marks work
Two marks.
One idea: show your work.
Text and files fail differently, so they are marked differently. Claude pairs an invisible watermark woven into generated words with a signed provenance label bonded to generated files. Both are machine-readable. Both travel with the content.
Technique A
Embedded watermarks
A quiet pattern applied at the model level as words are generated. Invisible to readers, unchanged in meaning, carried along by every copy and paste.
Explore the demoTechnique B
Signed provenance
A C2PA Content Credentials label embedded in generated files such as .svg, .png and .jpg — a cryptographic note that says “processed by Claude,” and flags tampering.
Inspect a fileEmbedded watermarks
An invisible watermark,
woven into the words
When Claude generates text, a watermark is embedded directly in the output at the model level. It is imperceptible — it does not change the meaning, quality or readability of the writing. And because the signal lives in the text itself, it follows the text wherever it goes.
- Invisibleto readers, in normal reading
- Unchangedmeaning, quality & readability
- Portabletravels with copy & paste
- Durablecan persist through some editing
- Model-levelpresent on every Claude surface
Transparency works best when it travels with the work itself. When Claude generates text, an invisible watermark is woven into the words as they are produced — a quiet pattern that never changes the meaning, quality, or readability of the writing. You can read the passage, quote it, or share it, and the signal stays hidden inside. Copy and paste it anywhere, and it makes the journey too. Edit a sentence here or there, and enough of the pattern usually survives to still be detected. The mark is not a judgment about who deserves credit, or whether the ideas are good. It is a simple, durable sign that this text passed through Claude — a courtesy to readers, and a small foundation of trust between people and the tools they use.
Illustrative simulation. Claude's real watermark is applied at the model level and is not a visible highlight — this overlay simply makes the concept tangible.
The signal lives in the words, not in a file — so it survives the clipboard. Copy the passage, or pick a scenario, then run the detector. You can also edit the text yourself and scan again: delete a phrase, rewrite a sentence, paste in your own writing.
Try:
Awaiting scan
- Words analyzed
- —
- Signal matches
- —
- Unfamiliar words
- —
Run a scan to see how confidently a detector could recognize this text's hidden signal.
How this demo works: the passage above carries a toy “watermark” — a fixed, hidden selection of signal words. The detector re-derives that selection from whatever text it is given and reports how much of the original signal survives. Claude's real system operates at the model level; the behaviors — copy/paste persistence, graceful decay under editing, weak signals on short excerpts — are the same ideas made visible.
Signed provenance metadata
A signed label,
bonded to the file
When Claude produces a file — an .svg, .png or .jpg — it can embed digitally signed provenance metadata following the C2PA open standard from the Coalition for Content Provenance and Authenticity. The label says “this file was processed by Claude” — and because it is signed, any change to the file afterward can be detected.
- Open standardC2PA Content Credentials
- Signeda cryptographic seal
- Tamper-evidentdetect changes after signing
- File types.svg · .png · .jpg
- Where supportedwhen Claude processes files
Content Credentials
C2PA manifest c2pa.standard · v2
- Claim generator
anthropic.claude- Actions
processed — "Content processed by Claude"- Signer
- Anthropic
- Signing time
- 2026-08-02T09:00:00Z
- Hash algorithm
- SHA-256
- Signed file hash
…- Computed hash
—- Signature
- Not yet verified
The manifest is illustrative, but the verification is real: “Verify signature” hashes the file's actual bytes in your browser and compares them against the signed hash above.
C2PA is the same open standard behind “Content Credentials” across the photography and media industries — a common, interoperable way for files to carry their own history. A mark here tells you the file was processed by Claude; it is not, by itself, a verdict on authorship. What the signature does prove is whether the file changed after signing.
Detecting Claude's marks
Can you check
content yourself?
Detection means checking whether text or a file carries a supported Claude mark. Anthropic will support users and third parties in running exactly that check, with details in forthcoming technical documentation.
Text
Scan a passage for Claude's embedded watermark. A found mark means the text may have been processed by Claude — nothing more, nothing less.
Files
Inspect a file's Content Credentials for Claude's signed label — and check whether the signature still matches the bytes on disk.
What a detected mark means — and what the absence of one doesn't — is where honesty matters most. Read the limitations →
Read the signal honestly
What a mark proves —
and what it doesn't
Claude's marks are designed to be a durable signal, not a verdict. Both directions of the signal deserve equal weight, so neither is softened here.
If a mark is detected
A signal — not a verdict
A detected mark signals that the content may have been processed by Claude. It is not fully conclusive, for two reasons:
- Claude may not be the original author. People use Claude to proofread, translate, summarize, and convert files. The output can carry a mark even when the underlying ideas originated elsewhere.
- The content may have changed since. Text can be modified, excerpted, or combined with other writing after Claude processed it. Files can be edited after they were signed.
If no mark is detected
Not a clean bill of health
The absence of a mark does not mean content wasn't AI-generated or processed. Five common reasons:
An older model made it
Models released before August 2, 2026 are in a transition period — their output may carry no mark while marking support is still in progress.
Heavy editing eroded it
Text that was heavily edited, paraphrased, translated, or mixed into other writing can lose too much of the pattern to detect. Try the “Heavily rewritten” preset in Demo 2 →
The passage is too short
A sentence or two offers too little signal to analyze reliably. Try the “Very short excerpt” preset →
The file's metadata was stripped
Format conversion, re-saving, and screenshots can all remove provenance metadata from a file that originally carried it. Try “Strip metadata” in Demo 3 →
The surface doesn't support it
The content may come through a platform, feature, or file type where that kind of marking isn't supported — for example, provenance metadata depends on each cloud platform's supported features.
No mark detectedStill not proof a human made it.
Mark detectedStill not proof Claude wrote it.
For builders
Shipping Claude
inside your product?
If you deploy Claude in your own product, the commitments here don't automatically cover you. You should independently assess what Article 50 requires of your products — and Anthropic's stated goal is to support you in meeting your own transparency obligations.
- Assess your obligations. Article 50 applies to what your product does with generated content — review it independently.
- Watch for the guidance. Technical guidance on marking and detection is forthcoming from Anthropic.
- Plan for detection. Support for users and third parties detecting Claude's marks is part of the commitments — build with that in mind.