Anthropic has confirmed that it will begin watermarking text generated by its AI models, including Claude, as part of its compliance with European Union regulations. The AI company updated its support documentation to disclose the new watermarking measures, which will apply across its range of products and services.
The move follows the entry into force of the EU AI Act’s Transparency Code on August 2. Under this regulation, AI companies must mark AI-generated or AI-edited content in a way that other systems can identify. The requirement is part of a broader push by the European Union to establish clear rules for transparency and accountability in artificial intelligence.
EU AI Act transparency requirements
The EU AI Act is a landmark piece of legislation designed to govern the development and use of AI across a wide range of applications. The Transparency Code specifically targets the need for machine-readable labels on AI-generated content, ensuring that such content is distinguishable from human-created material. This is particularly important for text, which can be easily copied, modified, and redistributed without obvious signs of machine origin.
Anthropic said that all models released after August 2 will automatically include technology that watermarks both computer-generated text and files. For files, the company is using the C2PA open standard, which is a widely adopted framework for content authenticity. C2PA provides cryptographic evidence of the origin and history of digital content, allowing downstream systems to verify whether a file was produced by AI.
The company also said it will extend support to older models, ensuring that watermarking is applied retroactively where possible. This means users interacting with previous versions of Claude may also encounter watermarks, depending on how the models are deployed.
How the watermark works
According to Anthropic’s support page, the watermark is embedded directly in the text itself. This design choice means the watermark will travel with the text when it is copied and pasted elsewhere. It may also persist through some forms of editing, although the company did not specify how much editing would be required to remove it.
The watermarking is applied at the model level, which means it will be present regardless of which Claude product or surface the text comes from. This includes the Claude platform API, the standalone Claude chat interface, Claude Code, Claude Cowork, and Claude Tag. By applying watermarking at the model level, Anthropic aims to create a consistent and reliable mechanism for identifying AI-generated text across all user interactions.
The company acknowledged that there are still unanswered questions about the durability of the watermark. For example, it is not yet clear how much editing, paraphrasing, or translation might strip the watermark from the text. Anthropic said it will provide clarification if requested, but the lack of detailed technical specifications has left some users wondering about the practical implications.
Industry moves toward watermarking
Anthropic is not alone in moving toward watermarking AI-generated content. Platforms across the technology industry are rushing to implement similar measures after backlash from users and concerns about regulatory scrutiny. The ability to generate convincing text, images, and audio at scale has raised concerns about misinformation, spam, and fraud, prompting both legislative bodies and private companies to seek solutions.
Last week, AI music platform Suno said it will mark tracks created on its platform after a spate of legal challenges. The move is seen as an attempt to address concerns about copyright infringement and the unlabeled use of AI-generated music. Suno’s decision reflects a broader trend among generative AI companies to be more transparent about the origins of their outputs.
Last month, newsletter service Substack teamed up with Pangram to flag AI-generated content. Substack’s CEO, Chris Best, called out “Claudefishing,” a term used for people using AI to generate content that appears to be written by a human expert. This collaboration is designed to help readers understand when they are engaging with AI-produced material, especially in a context where trust and authenticity are critical.
Apart from Anthropic, other companies including Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have committed to adhering to the EU’s code. This collective commitment suggests that watermarking and content provenance are becoming baseline expectations for AI companies operating in the European market.
Technical considerations and challenges
Watermarking text is a complex technical challenge. Unlike images, which can be embedded with invisible watermarks, text offers fewer opportunities for unobtrusive marking. Approaches to text watermarking often rely on subtle statistical patterns in word choice or sentence structure. These patterns are designed to be imperceptible to human readers but detectable by algorithms.
The use of C2PA for files provides a different mechanism. C2PA relies on cryptographic metadata that is securely bound to the content. This metadata can include information about the AI system that generated or edited the file, along with a timestamp and other verification details. However, C2PA metadata can be stripped if the file is converted or re-encoded, which is why embedding a watermark directly into text is considered an important complementary approach.
One of the key challenges is balancing watermark reliability with usability. If the watermark is too strong, it may affect the quality or readability of the generated text. If it is too weak, it may be easily removed by even basic editing. Anthropic’s decision to embed the watermark at the model level suggests it is attempting to achieve a level of robustness that can survive common user actions without compromising the user experience.
There are also concerns about how watermarking might interact with accessibility and innovation. Some researchers argue that watermarking could unfairly penalize non-native speakers or people who use AI to improve their writing. Others worry that watermarking systems could be bypassed by advanced attackers, making the technology only partially effective. These concerns are not unique to Anthropic, and the industry as a whole is still grappling with the implications.
Regulatory backdrop
The EU AI Act is one of the most comprehensive AI regulatory frameworks in the world. It takes a risk-based approach, imposing stricter requirements on AI systems considered high-risk. The transparency provisions apply to a wide range of generative AI systems, including those that create text, audio, and visual content. Companies that fail to comply with the act face significant fines, which creates a strong incentive to adopt watermarking and other compliance technologies.
The act’s focus on transparency is intended to give people the ability to make informed decisions about the content they encounter. By requiring machine-readable labels, the EU aims to allow platforms and regulators to automatically detect AI-generated content and apply appropriate policies. For example, social media platforms could use these labels to flag or down-rank AI-generated misinformation, while e-commerce platforms could use them to ensure product reviews are authentic.
Anthropic’s confirmation that it will watermark text is therefore a significant step, not only for the company but also for the wider adoption of transparency standards. The company’s position as a major AI developer means its approach could influence how other companies implement their own watermarking systems. It also sets a precedent for how AI-generated text might be handled in regulated environments.
Product coverage and user impact
Anthropic’s watermarking will apply to a broad range of surfaces. The mention of Claude Code, Claude Cowork, and Claude Tag indicates that the company is integrating watermarking into its newer and more specialized tools. This includes coding assistance, collaborative work features, and tagging or annotation systems. By covering all these products, Anthropic is ensuring that watermarks are not limited to a single chat interface.
For users, the practical impact will depend on how they use Claude and other models. Those who copy text from one document to another may not notice the watermark, as it does not alter the visible content. However, tools that analyze text for AI provenance may be able to detect the embedded pattern. This could have implications for writers, marketers, and researchers who rely on AI assistance and need to prove the authenticity or origin of their work.
The company has not disclosed whether there is a way for users to opt out of watermarking. Based on the support page, watermarking is an automatic and mandatory feature for models released after the cutoff date. This suggests that users will not have a choice in the matter, at least for models covered by the EU regulation.
Looking ahead
The announcement from Anthropic is part of a rapidly evolving landscape. As more companies adopt watermarking, the technology is likely to improve in both robustness and subtlety. We may also see the development of standards that allow different AI systems to recognize each other’s watermarks, creating a common framework for content authenticity.
At the same time, the effectiveness of watermarking will rely on the cooperation of the entire ecosystem, including content platforms, browser makers, and detection tool vendors. Without widespread adoption, watermarks may not fulfill their intended purpose of reducing misinformation and building trust. Anthropic’s commitment to adhere to the EU code is one piece of a larger puzzle, as regulators, companies, and civil society continue to shape the future of AI transparency.
Source: TechCrunch News