Claude’s invisible watermark
You won’t see it. Copy-paste still carries it
Anthropic signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content. Models launched in the EU on or after 2 Aug 2026 mark from day one. Marks apply worldwide. Detection docs are still forthcoming.
Four facts first
Code of Practice signed
Anthropic signed as a provider of generative models and systems. This page restates the help center. It is not legal advice.
New models mark on day one
Claude models launched in the EU on or after this date support machine-readable marking at launch. Older models sit in a transition window; Anthropic says it is adding support.
Text watermark + file metadata
Generated text weaves in an imperceptible watermark that travels with copy-paste and may survive some edits. Supported .svg / .png / .jpg get signed C2PA provenance metadata.
A mark ≠ Claude wrote it
Official wording: a detected mark means the content may have been processed by Claude. Proofreading, translation and summaries still mark. No mark does not mean “not AI.”
What Claude actually stamps
Two complementary techniques. (1) A text watermark woven into the generated string. It does not change meaning, quality or readability. Because it is part of the text, it travels on copy-paste and may persist through some edits. (2) Signed provenance on supported files, using the C2PA open standard, so you can tell a file was processed by Claude and whether it was tampered with later.
Where it applies — and where it does not
Official: marks cover output from supported models on Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, worldwide. Cloud partners (AWS / GCP / Microsoft Foundry) get the text watermark; signed C2PA may not be on every platform. Detection details are “forthcoming.” If you ship a product on Claude, you still assess your own Article 50 duties.
Timeline
Anthropic signs the EU AI Act Art. 50(2) Code of Practice. The help-center article is how they say they will implement it.
New models launched in the EU on or after this date mark from day one. Older models are in the statutory transition; marking is being added.
Anthropic will help users and third parties detect marks. Technical documentation is still forthcoming as of this page.
Confirmed vs do not invent
Confirmed
New EU-launched models on/after 2 Aug 2026 mark from day one; marks apply worldwide; API / Claude / Claude Code / Cowork / Tag; text watermark + C2PA on supported files; cloud watermark via AWS / GCP / Foundry; C2PA is not promised on every platform. Source: support.claude.com.
Do not invent
Do not write that every older model is already marked. A mark is not proof Claude is the original author. Do not assume C2PA survives a screenshot, re-save or format conversion. Do not claim QCode strips or adds watermarks.
How the two marks split the job
Invisible text watermark
Woven into generated text. Travels with copy-paste; may survive some edits. No change to meaning or quality. Applied at the model, whichever Claude surface you used.
C2PA file metadata
Signed provenance on supported .svg / .png / .jpg. Signals Claude processed the file and whether it was tampered with. Screenshots, conversions and stripped metadata drop it.
If you build a product, assess yourself
If you deploy Claude inside your own product, Anthropic says you must independently assess what Article 50 requires of you. Their CoP goal is to help you meet transparency duties; detection docs are not out yet. Short passages, heavy edits, older models and stripped metadata can all fail a check.
On QCode
This page is about how Claude marks output. It is not a new model SKU. QCode Claude ids (for example claude-sonnet-5) follow Anthropic model behavior. We do not claim we strip watermarks or add our own. Whether a call is marked is whatever Anthropic implemented for that model.
Claude watermark FAQ
What is the invisible watermark?
When a supported Claude model generates text, it weaves an imperceptible watermark into the text itself. You will not see it. It does not change meaning or quality. Copy-paste carries it; some edits may leave it intact.
Does it change the words I read?
Officially no — meaning, quality and readability stay the same. Marking is at the model level, whether you used the API, Claude, Claude Code, Cowork or Tag.
No mark means it is not AI?
No. Older models, heavy edits, very short passages, stripped file metadata, or a platform that does not support that mark type can all come back clean. No mark ≠ not AI.
A mark proves Claude authored it?
No. A mark only says the content may have been processed by Claude. Proofreading, translation, summaries and file conversion still mark. Claude may not be the original author.
Does this hit the API and Claude Code?
Official coverage is supported models on the API, Claude, Claude Code, Cowork and Tag, worldwide. Cloud text watermarks go through AWS / GCP / Foundry. C2PA is not guaranteed on every platform.
Does QCode add or strip watermarks?
No. We do not claim to strip marks or add our own. Listed ids such as claude-sonnet-5 follow Anthropic’s model behavior. This is not a new SKU.
Sources
Anthropic help center: How Claude marks AI-generated content (support.claude.com/en/articles/16266773). EU AI Act Art. 50(2) Code of Practice. This page is not compliance advice.
The mark follows the model, not the reseller
Listed Claude ids mark the official way. QCode does not add or strip.
Related
Claude Sonnet 5
The current mid-tier workhorse. Output is marked per official rules.
Claude Code guide
Terminal Claude is marked at the model, same as the API.
Claude Opus 5
Flagship tier. Marking follows that model’s launch date.
Not affiliated with Anthropic. This page restates the official help center and is not legal advice. Marking rules may change as docs land.