Anthropic opens Claude watermark detection to regulators and media

Anthropic is giving regulators, journalists, researchers and fact-checkers access to a new API that allows them to determine whether a text contains an invisible watermark from the Claude language model. The company is responding to a concrete regulatory obligation: the EU AI Act requires that text generated by AI models carries a hidden marker.

The API is intended for parties that have a societal or professional role in monitoring AI use. Regular users or companies deploying Claude in their own products will not have access to the detection endpoint. Anthropic is explicitly targeting entities with a verification function, such as newsrooms, regulatory authorities and independent research groups.

Watermarking AI-generated text is technically a different challenge from marking images or videos. With visual content, metadata can be embedded or subtle patterns can be encoded in pixels. With text, the approach relies on statistical patterns in word choice and sentence structure that are invisible to humans but can be recognised by an algorithm. How robust that approach is when text is edited or translated is one of the open questions in the field.

What the EU AI Act requires

The EU AI Act, which is being phased in, obliges providers of generative AI models to label AI-generated content. For text, this means a form of invisible watermarking, so that the origin can in principle be traced without the marker being visibly disruptive to the end user.

The rationale behind the obligation is that recipients of AI-generated text, whether citizens, journalists or public institutions, must be able to verify its origin. The legislator aims in this way to counter misinformation and promote transparency around the use of AI in sensitive contexts, such as news, legal documents or government communications.

With this step, Anthropic is one of the first major providers to actively make a detection API available to external parties. Whether competitors such as OpenAI or Google DeepMind offer or plan similar tools has not been confirmed on the basis of available sources.

Criticism of the technology

Anthropic's move has also drawn critical responses. Opponents of text watermarking point to two objections that are separate but mutually reinforcing.

The first objection is technical. Watermarking works by introducing subtle biases into the text generation process, for example, by favouring certain synonyms more frequently. Critics argue that this can come at the expense of the quality and naturalness of the output. Whether that effect is noticeable in practice depends on how aggressively the marking is applied, but it is a trade-off made by the provider outside the end user's view.

The second objection is legal and ethical. Many employment and service contracts now include a clause that prohibits or restricts the use of AI-generated text. A detection API makes it possible to retrospectively check whether someone has used AI. Critics warn that this creates a form of surveillance in which the burden of proof lies with the user, and where false positives, texts incorrectly flagged as AI-generated, can have serious consequences.

Added to this is the fact that watermarks can be removed or disrupted by paraphrasing, translating or merging text with human-written content. The reliability of detection in such cases is a subject of ongoing research.

Access and scope of the API

Anthropic has not yet fully disclosed what the access procedure for the detection API looks like: who exactly qualifies, whether there is an application process and under what conditions use is permitted. What is clear is that the company intends to restrict access to parties with a demonstrable verification function.

For media companies and fact-checking organisations, the API can serve as a tool for assessing submitted material or investigating automated disinformation campaigns. For regulators, it provides a technical instrument for enforcing the transparency obligations set out in the AI Act.

The broader question is how effective it is for a single provider to make its own watermark available to third parties, while content from other models falls outside the scope of this API. A text generated by a different large language model carries no Claude watermark and will therefore not be recognised by this API. Full coverage requires multiple providers to open up comparable tools, ideally with harmonised standards.

For the Dutch and European AI scene, this is a concrete example of how regulation compels technical choices at major providers. Startups and scale-ups deploying generative AI in their products will need to account for similar obligations once the relevant provisions of the AI Act come fully into force. Investors and policymakers tracking the development of the European AI ecosystem can see here how compliance requirements influence product design and make new infrastructure, such as detection APIs, necessary.