SAN FRANCISCO, August 16, 2026, 17:13 PDT
- Supported Claude models now embed machine-readable marks in generated text worldwide.
- Four users told Business Insider they canceled subscriptions over the change.
- Anthropic says a detected mark signals processing, not authorship.
Anthropic’s invisible Claude watermarks are drawing early user resistance. The dispute puts a practical limit on what AI provenance can prove.
The privately held AI company applies model-level marks to text from supported models. The policy covers Claude products worldwide, not only European users.
That scope matters for developers, writers and consultants. A mark may survive copying and light editing, yet it cannot establish who wrote the underlying work.
| Output | Claude marking method | What survives | What Anthropic has confirmed |
|---|---|---|---|
| Generated text | Imperceptible embedded watermark | Copy and paste; some editing | Meaning, quality and readability should not change |
| Supported files | Digitally signed C2PA provenance metadata | Normal handling when metadata remains attached | Examples include SVG, PNG and JPG |
| Unsupported file or platform case | May be unavailable | Not applicable | Coverage depends on platform features |
Anthropic says the text mark “will travel with the text when it’s copied and pasted elsewhere.” Detection tools are still forthcoming, and the technical method remains undisclosed.
The rollout follows Article 50 of the European Union’s AI Act. Its transparency duties took effect on August 2 and require machine-readable marks for generated or manipulated content.
| Claude access route | Embedded text watermark | Signed file metadata |
|---|---|---|
| Claude app and website | Yes, for supported models | Where supported |
| Claude Platform API | Yes, for supported models | Where supported |
| Claude Code, Cowork and Tag | Yes, for supported models | Feature-dependent |
| Amazon (NASDAQ:AMZN) Web Services, Alphabet (NASDAQ:GOOGL) Google Cloud and Microsoft (NASDAQ:MSFT) Foundry | Yes, for supported models | May vary by platform |
Anthropic is among 82 signatories to the code’s provider section. The European Commission listed 152 signatories in its deployer section as of August 12.
The commercial reaction is real but small so far. Business Insider interviewed four users who said they canceled Claude subscriptions, while dozens posted similar claims on X.
Anthropic told the publication it had not seen a broader cancellation trend. One user feared that marked code could create authorship questions or contract penalties.
| Detection result | Reasonable conclusion | Conclusion not supported |
|---|---|---|
| Claude mark detected | Claude may have processed the content | Claude wrote every word or originated every idea |
| No Claude mark detected | No supported mark was found | The content is human-made |
| File metadata missing | The provenance record is unavailable | The file never passed through Claude |
False certainty is the central problem. Claude may mark proofreading, translation or summarization of human work. Heavy edits, short passages and stripped metadata can remove a detectable signal.
For employers and schools, that makes the mark an investigative clue. It is not a plagiarism verdict, authorship test or substitute for documented editorial review.
Risks: Undisclosed detection accuracy could produce overconfident decisions. Users may also move work to unmarked models, weakening the transparency goal.
Anthropic plans free detection access for users and third parties. Until technical documentation arrives, buyers cannot independently compare robustness, false positives or edit resistance.
Related articles: News
