Anthropic is preparing to add invisible AI watermarks to content created or edited by Claude, with the change tied to transparency requirements under the European Union’s AI Act.
The policy will apply not only to text written from scratch by Claude, but also to material that Claude modifies. That means documents, code, reports and other content could receive a watermark even when a human originally created the material.
Anthropic says new Claude models launched in the European Union on or after August 2, 2026 will support watermarking, while older models are expected to gain the capability over time.
Claude watermarking will apply across products
| Area | Detail |
|---|---|
| Main requirement | Invisible watermarking for AI processed content |
| Regulatory basis | EU AI Act transparency rules |
| New EU models covered | Models launched on or after August 2, 2026 |
| Older Claude models | Support planned later |
| Products affected | Claude, Claude Cowork, Claude Code and Claude API |
| Text watermark | Embedded invisibly in the text |
| File marking | Provenance metadata for some files |
| Detection tools | Planned for public release |
| Global rollout | Planned |
Anthropic says the system will be deployed worldwide rather than limited to the European Union.
That means Claude generated or edited content could carry identifying markers regardless of where the person using the service is located.
Watermarks will be embedded into text itself
For supported models, Anthropic says the watermark will be woven directly into generated text in a way that is not visible to the reader.
The company says the marker should survive copying and pasting and may remain detectable even after some editing.
That could make it easier to identify content that has passed through Claude, but the system will not be perfect.
A watermark does not necessarily mean Claude originally wrote the material. Human written content that is later edited by Claude may also be marked.
At the same time, Claude processed content may sometimes lose its watermark, especially if it is heavily rewritten, too short or generated by a model that does not yet support the system.
Files will use provenance metadata
Anthropic is also planning a different approach for some file formats.
Certain files created by Claude, including SVG, PNG and JPG images, will receive provenance metadata rather than relying solely on the text watermarking method.
The goal is similar: provide a way to trace whether AI played a role in creating or processing the content.
Anthropic also plans to release tools that can detect these markers, although details on how the detection system will work have not yet been announced.
The EU AI Act is driving the change
The policy is being introduced as part of compliance with the European Union’s AI Act.
The law includes transparency requirements for AI generated or AI processed content, pushing providers to make that involvement easier to identify.
Anthropic is applying the system across its full product range rather than restricting it to its consumer chatbot.
That includes Claude Code, where generated or modified source code could also contain the watermark.
Claude Cowork and API generated content are expected to follow the same policy.
Other AI companies also face transparency requirements
The EU rules do not apply only to Anthropic.
Other major AI companies operating in Europe also need to comply with the transparency provisions.
Google and Meta have said they intend to follow the relevant requirements, while OpenAI has described a layered approach to compliance.
However, the companies have not all provided the same level of detail about whether they will use invisible text watermarks, metadata, detection systems or a combination of methods.
Anthropic’s approach is therefore one of the clearer examples so far of how AI content transparency may work in practice.
Watermarking may change how AI assisted writing is identified
The biggest practical effect is that content touched by Claude may become easier to identify, even when the AI contribution is relatively small.
A cover letter written by a person and polished by Claude could be marked. The same could apply to business reports, programming code or other professional material.
That does not automatically show how much of the content Claude created, only that Claude processed it.
The distinction matters because AI editing and AI authorship are not the same thing.
Anthropic’s planned detection tools should provide more clarity once they become available, but watermarking will still have limitations. Heavy editing, short content and unsupported models may prevent reliable detection.
Even so, the change signals a broader shift toward making AI involvement in digital content more visible as new transparency rules begin to take effect.
Discussion (0)
Be the first to comment.