Anthropic adds invisible watermarks to text and images generated by Claude

Anthropic will add invisible, machine-readable watermarks to all text and images generated by Claude. The company announced this on a new support page for Claude, explicitly linking the measure to European transparency requirements for AI systems. For users, nothing changes visually: the markings are not visible in the output itself.

Generated text will carry embedded watermarks, while generated files will where possible be provided with digitally signed provenance metadata. The latter is done via the C2PA standard (Coalition for Content Provenance and Authenticity), an open specification also used by Adobe, Microsoft and other technology companies to record the origin of digital files.

The announcement fits within a broader movement in which AI providers are preparing for the requirements of the European AI Act, which stipulates, among other things, that AI-generated content must be recognisable to systems and, in certain cases, to people.

How the watermarks work

The two techniques Anthropic employs target different types of output. For text, a watermark is encoded directly into the generated text in a way that remains invisible to human readers but can be read by specialised detection software. For images and other files, C2PA metadata is added: a digitally signed certificate indicating that the file was created by an AI system and specifying which system.

C2PA is now an established industry standard. Files containing C2PA metadata can be verified using tools such as Adobe's Content Credentials verification service. The standard not only describes the provenance of AI-generated content but is also used to record the editing history of photographs.

Anthropic does not specify how much of the watermarking technique holds up against editing or file conversion. This is a well-known concern with such systems: copying and pasting text into another application, or saving and re-uploading an image, can result in the loss of metadata. Anthropic implicitly acknowledges this by referring to metadata support 'where supported'.

European regulation as the driver

Anthropic ties the measure to European regulation. The EU AI Act, which is being phased in, requires providers of so-called general-purpose AI models to take measures to mark AI-generated content in a machine-readable way. Additional obligations apply to synthetic audio and video, requiring that content also be recognisable to people.

Many major AI providers are currently working on similar implementations. Google has added C2PA support to some of its image generation tools, and OpenAI already applies the standard to images created via DALL-E. For text, broad industry adoption of watermarking standards remains more limited, partly because text watermarks are technically more vulnerable than metadata embedded in files.

The European AI Act follows a layered compliance timeline. The most stringent obligations for providers of general-purpose models take effect in August 2025. With this announcement, Anthropic is getting ahead of that deadline.

What this means for developers and businesses using Claude

For businesses and developers working with Claude via the API, the measure has practical implications. Those using Claude to generate content that is subsequently published can in principle rely on the output already carrying the required markings. This reduces their own compliance burden, but at the same time makes it relevant to understand exactly what metadata appears in the output and what may become visible through external verification tools.

For end users working through Claude.ai, the interface does not change. The watermarks are designed to provide transparency to systems and regulators, not to be visible in everyday workflows.

Anthropic has not yet announced an exact date on which the watermarks will be activated for all users. The support page describes the functionality as an upcoming measure.

For the Dutch and broader European AI landscape, this is a signal that the AI Act is compelling major international providers to make concrete product adjustments. Companies and government bodies that deploy AI-generated content and must comply with transparency obligations will increasingly be able to rely on built-in markings from the model provider itself. At the same time, the question remains relevant as to how robust those markings are in practice, and whether enforcement is feasible if metadata can easily be lost when files are edited or forwarded. That question is also pertinent for regulators such as the Autoriteit Persoonsgegevens and the future AI supervisory body in the Netherlands.