Anthropic is adding hidden watermarks to Claude-generated text: Can they be removed?

Anthropic’s move to add watermarks and other provenance signals to its AI-generated content is aimed at complying with the European Union’s AI Act and the voluntary ‘Code of Practice on Transparency of AI-Generated Content’.

Under the EU’s AI Act, Anthropic is classified as a provider of generative AI models and generative AI systems. As a result, it has to comply with legal requirements pertaining to the marking and detection of AI-generated content under Article 50(2) of the Act, which went into effect from August 2, 2026.

The rise of AI-generated, synthetic content has threatened to disrupt online information ecosystems and cause tangible harm such as fraud, misrepresentation, and harassment. This has pushed policymakers worldwide to search for solutions that could mitigate such risks and empower users.

But how does Anthropic plan to comply with the obligations under the EU’s Transparency Code? How will watermarking work for AI-generated content? And how effective are these measures likely to be in detecting such content?

How will Claude-generated content be marked?

Anthropic has broadly outlined two complementary steps to mark content generated and processed by Claude: watermarks embedded in text and signed provenance metadata attached to files.

As part of the first step, Anthropic said that the watermarks will be imperceptibly woven directly into the text itself. This means that human users will not be able to see it. The watermarks will be applied at the model layer. Text generated by any of Anthropic’s Claude family of models will carry these watermarks, as will content produced through various Claude-powered tools and interfaces.

Since the machine-readable watermark will be embedded directly into the text, it will be carried over when users copy and paste the text elsewhere. Anthropic has only said that the text-based watermark “may persist” after the content has been edited.

Under the second technique, Anthropic said it will attach signed provenance metadata for various file types such as .svg, .png, or .jpg. Essentially, this will help indicate whether content such as images, graphics, and documents was generated by Claude.

The cryptographically signed metadata, which includes information such as the origin, context, and the history of a piece of text or image, will appear in accordance with the open standards established by the Coalition for Content Provenance and Authenticity (C2PA), a cross-industry initiative to record information about content provenance through Content Credentials.

The signed metadata label will not only lets users know whether a file has been generated or processed using Claude, but also makes it easier to tell whether the file has been tampered with. Anthropic further said it will be rolling out AI detection tools to help users and third parties detect embedded watermarks and signed metadata on their own.

What are the limitations of watermarks?

While machine-readable watermarks provide important signals about content, they also carry limitations across all content types. In the case of images, visual labels can be cropped or edited out while watermarks embedded in metadata can easily be circumvented by taking a screenshot of the image. Note, Google’s SynthID watermarks are harder to tamper with because the tool subtly changes the pixels of an image.

Coming to invisible watermarks in text, research studies in the last few years suggest that they can be manipulated far more easily than in audio-visual content. Other AI models can be used to paraphrase AI-generated text, which may pass scrutiny given that AI detection tools themselves are not 100 per cent reliable. It is also possible to rewrite the text manually, although doing so means that the final version is no longer purely AI-generated.

The difficulty of reliably distinguishing AI-generated text from human writing is perhaps best illustrated by the debate around the use of the em dash. Online posts, along with a few AI detection tools, have treated the frequent use of the punctuation mark as a telltale sign of AI-generated writing, even though em dashes are also commonly used by human writers.

What is the EU Transparency Code?

The Transparency Code makes labelling of AI-generated content mandatory to ensure transparency and enable human users to distinguish manipulated content online. Anthropic suggested that developers who have deployed Claude within their own products should “independently assess what Article 50 requires of your products and services”.

The transparency obligations fall under the purview of Article 50(2) of the EU Artificial Intelligence Act (AIA), which is a framework for the use and supply of AI systems within the bloc with a risk-based categorisation. The Transparency Code lays down specific obligations for two types of actors: providers and deployers

AI providers such as OpenAI and Anthropic are required to undertake marking of AI-generated or manipulated content by  rolling out machine-readable marking techniques such as digitally signed metadata and imperceptible watermarks, with an optional method of fingerprinting or logging of AI-generated content.

Companies designated as AI providers are also required to provide users access to detection mechanisms while ensuring that the detection solutions are compliant with EU privacy and data protection law. Non-compliance could lead to administrative fines of up to €15 million.

Several tech giants besides Anthropic such as Google, OpenAI, Meta, and Microsoft have signed the EU’s voluntary Transparency Code so far. Note, adherence to the Code is voluntary but the transparency requirements under Article 50 of the EU’s AI Act are legal obligations.

In February this year, India’s IT Ministry also notified an amendment to the IT Rules, 2021, that requires users and social media platforms to label synthetically generated information (SGI) prominently for users to see.

A key difference between the EU Transparency Code and the IT Rules amendment is that the Code requires developers of AI systems to include a label or watermark on AI-generated content, and the IT Rules put the onus of disclosure on the users while requiring social media platforms to enforce it.