Starting with models launched on or after August 2, 2026, Claude automatically embeds marks that are invisible to people into the text and images it generates. Based on Anthropic's official support documentation, this guide walks beginners through what this marking is, which products it covers, how you can check for it, and what its limitations are.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
Why does this marking exist?
It exists because Anthropic signed onto an EU transparency commitment for AI-generated content. According to the official documentation, Anthropic has signed the EU AI Act's (EU AI Act — European Union regulation that places several obligations on AI providers) Article 50(2) Code of Practice on Transparency of AI-Generated Content, which requires making AI-generated content identifiable. As part of that commitment, Claude began adding machine-readable marks to the content it produces.
A few terms, explained simply
- Watermark — normally the faint pattern visible in paper currency; here it means a digital mark embedded in text that people cannot perceive.
- C2PA — short for Coalition for Content Provenance and Authenticity, an international standard for recording, in files like images and video, when and how a piece of content was created.
- Provenance metadata — extra information about a piece of content's origin, i.e., who or what created it, attached invisibly inside the file itself.
- API (Application Programming Interface) — a channel that lets programs exchange data with each other; developers use it to connect Claude to their own products.
Which models and products does this cover?
Coverage starts with Claude models launched on or after August 2, 2026, and spans nearly every Claude product across all of Claude's worldwide service regions. Here is what the official documentation states about the scope:
| Category | Details |
|---|---|
| Models covered | Claude models launched on or after August 2, 2026. Support for earlier models is being added over time. |
| Products covered | Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag |
| Cloud partners | Includes access through AWS, Google Cloud, or Microsoft Foundry |
| Regions | Everywhere Claude is offered, worldwide |
Note: content made with older models released before August 2, 2026 may not carry a mark yet. Anthropic is rolling out support gradually, so the absence of a mark on output from an older model isn't unusual.
How exactly does Claude mark content?
The official documentation describes two marking methods. The first is an embedded text watermark: Claude inserts a mark directly into generated text that people cannot perceive. It's invisible and doesn't affect the quality of the writing, and the documentation notes this mark may be preserved even after the text is copied and pasted elsewhere. The second is signed provenance metadata attached to images: for supported file formats such as .svg, .png, and .jpg, Claude attaches signed metadata that follows the C2PA standard. For example, opening an image Claude generated won't show any visible mark on screen, but information showing the content passed through Claude is recorded in the file's metadata.
How can I check whether content is marked?
There isn't yet an official tool for everyday users to check this themselves. According to the official documentation, Anthropic plans to support users and third parties in detecting both the text watermark and image provenance metadata, and says it will share the specific methods in separate technical documentation at a later date. If you're looking for a way to verify right now whether a given piece of content came from Claude, be aware that no such official tool has been released yet.
A missing mark doesn't mean Claude didn't make it
Marking alone can't fully confirm a piece of content's origin. The official documentation lays out these limitations:
- Finding a mark only indicates the content may have passed through Claude — it isn't complete proof of provenance.
- Claude isn't always the original author. It may have only proofread, translated, summarized, or reformatted text someone else wrote.
- Even marked content can lose its mark if it's later modified by a person or another program.
Conversely, the absence of a mark doesn't prove Claude wasn't involved. The documentation notes content can be Claude-generated without a detectable mark when:
- It was created with an older model launched before August 2, 2026, which doesn't support marking
- It was heavily edited, paraphrased, translated, or mixed with other content
- The text is too short to carry a mark
- Metadata was stripped through format conversion, resaving, or screenshotting
- It came from a platform, feature, or file format that doesn't yet support marking
Caution: don't treat this marking feature as a perfect authenticity detector. The documentation itself states that finding a mark only indicates a possibility, so making important decisions based solely on whether a mark is present or absent isn't recommended.
If you're building your own product with Claude
In that case, you need to independently review what Article 50 requires of your own product. The official documentation states that if you deploy Claude within your own product, you should independently assess what Article 50 requires of your products and services. Anthropic says that, consistent with its commitments under the EU Code, its goal is to support such businesses in meeting their own transparency obligations.
Frequently asked questions
Q. Does everything I generate with Claude right now carry a mark?
No. Marking is supported starting with models launched on or after August 2, 2026, and the official documentation says support for earlier models is being extended gradually.
Q. If I copy text with a watermark and paste it into another document, does the mark disappear?
According to the official documentation, the watermark may be preserved even after copying and pasting. That said, as the documentation also notes, heavy editing or paraphrasing can cause the mark to disappear.
Q. Does every image file get signed metadata attached?
No. The official documentation states that C2PA signed metadata is attached only to supported file formats such as .svg, .png, and .jpg. Support for other formats isn't specified in the documentation.
Q. If a mark is found, is that definitive proof Claude made the content?
No. The official documentation states that finding a mark indicates a possibility rather than complete proof of provenance, and notes that Claude may not be the original author — it may have only proofread, translated, or otherwise transformed the content.