Anthropic has confirmed that it will watermark text generated by its AI models, including the Claude family, as part of its effort to comply with European regulations. The artificial intelligence company updated its support page to explain that watermarking will be applied automatically to new models and gradually to older ones. The move aligns Anthropic with the European Union’s AI Act, which introduces binding transparency obligations for companies that develop and deploy generative AI systems.
The EU AI Act’s Transparency Code came into effect on August 2, and it requires AI providers to ensure that AI-generated or AI-edited content is identifiable by other systems. This means that text, images, audio, and video produced by AI tools need to carry some form of machine-readable marker. Anthropic’s announcement is one of the first clear commitments from a major AI lab regarding text watermarking, a technically difficult problem that many companies have been exploring quietly.
How Anthropic plans to watermark text
According to the updated support page, all models released after August 2 will automatically include technology that watermarks both computer-generated text and files. For files, Anthropic will use the Coalition for Content Provenance and Authenticity, or C2PA, an open technical standard designed to attach cryptographic metadata to digital content. For plain text, the company plans to embed a watermark directly into the generated content.
“Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing,” the support page states. “Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.”
The decision to apply watermarking at the model level is significant. It means the watermark is present even if a user interacts with Anthropic’s API, the consumer chatbot, or one of the company’s more specialized tools. The company listed several products that will be covered, including Claude, Claude Code, Claude Cowork, and Claude Tag, in addition to the Claude API.
Anthropic also said it will extend watermark support to older models as time goes on. However, the company did not provide a specific timeline for when older models would receive the technology. It also acknowledged that the robustness of the watermark varies depending on how much the text is edited. The company said it is not yet clear how much editing a user needs to do in order to remove the watermark, and it has been asked for clarification.
Why the EU AI Act matters
The EU AI Act is a comprehensive piece of legislation that regulates artificial intelligence based on the level of risk each application poses. Most generative AI systems, including large language models such as Claude, are considered to be limited risk, but they still come with transparency responsibilities. Providers must ensure AI-generated content is distinguishable from human-written content when that is technically feasible.
The transparency obligations that took effect on August 2 are part of the act’s broader approach to building trust in AI. Lawmakers in Europe have expressed concern that AI-generated content could be used to spread disinformation, interfere with elections, or manipulate public opinion. Watermarking is seen as one tool among many that can help platforms, journalists, and ordinary users understand the origin of the content they encounter.
Watermarking is not the only requirement in the AI Act. Companies also need to publish detailed summaries of the copyrighted data used to train their models. They must implement cybersecurity protections and follow specific design standards. But the transparency requirements are among the most visible and immediately relevant to consumers, because they directly affect the way AI outputs are labeled and tracked.
The technical challenge of text watermarking
Watermarking images and video is relatively straightforward compared to text. Images can carry invisible pixel patterns or metadata that is difficult to remove without degrading quality. Video and audio can have similar signals embedded in their data streams. Text, however, is a sequence of discrete tokens. There is no natural place to hide a signal that remains intact every time the text is copied, re-encoded, or paraphrased.
Most text watermarking approaches work by altering the statistical pattern of token selection during generation. A language model assigns probabilities to different words; a watermarking system introduces a subtle bias that makes certain words or word sequences more likely to appear. If the bias pattern is known, a detector can look at the text and determine whether it was generated by that particular model. This is often called a statistical watermark.
The main challenge is making the watermark robust to editing. If someone rewrites a sentence, reorders paragraphs, or uses a different language, the statistical signals may be destroyed. Anthropic says its watermark may persist through some editing, but it does not specify how much. In the past, researchers have shown that strong paraphrasing can defeat most statistical watermarks, although recent advances have made them more resilient.
Another challenge is avoiding false positives. A watermark detector must be highly precise so that it does not flag human-written text as machine-generated. This is especially important for journalists, authors, and students who may use digital tools that subtly alter their writing without being fully AI-generated. The EU AI Act requires that transparency mechanisms be reliable and accurate, but it does not define a specific technical standard.
There is also a tension between watermarking and open access to models. Open-source language models can be downloaded and run without any watermarking infrastructure. Users can remove a watermark layer or fine-tune the model until the statistical signal is no longer detectable. Anthropic’s models are proprietary, which makes it easier for the company to enforce watermarking. But other companies that release open weights will face greater difficulties implementing mandatory watermarks.
Industry momentum is building
Anthropic is not alone in moving toward watermarking. The broader industry has been under pressure from governments, regulators, and users to address the risks of AI-generated content. Platforms that host user-generated content have been criticized for spreading AI-generated misinformation, and many are now looking for technical solutions.
Last week, AI music platform Suno said it would begin marking tracks created on its platform after facing a string of legal challenges. The music industry has been especially concerned about AI-generated songs that mimic the voices and styles of real artists. Watermarking can help rights holders identify unauthorized AI music and enforce copyright claims.
Last month, newsletter platform Substack partnered with Pangram to flag AI-generated content on its site. Substack’s CEO, Chris Best, also used the term “Claudefishing” to describe people who use AI to generate content that appears to be from a real person. The term is a play on “catfishing,” which refers to the practice of creating a fake online persona. Best’s remarks highlight the growing concern about authenticity in online writing and social interactions.
Beyond these recent moves, several major AI companies have committed to complying with the EU’s transparency code. Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have all said they would follow the code’s requirements. This suggests that watermarking is becoming a de facto industry standard, at least in Europe.
What this means for users
For everyday users, watermarking can provide a useful signal when trying to evaluate the authenticity of content. If a long text or an important document appears to have been generated by Claude or another AI model, the watermark can help platforms label it accordingly. This could reduce the spread of viral AI-generated hoaxes and make it harder for malicious actors to use AI for impersonation.
At the same time, watermarking is not a silver bullet. It does not reveal the identity of the person who prompted the model, nor does it guarantee that the content is truthful. A watermark simply indicates that the text originated from a particular AI system. Users still need to apply critical thinking and verify facts from reliable sources.
Privacy advocates have also raised concerns about watermarking. A watermark that is embedded at the model level could potentially be used to track a specific user’s outputs if combined with session data. However, Anthropic has not indicated that it plans to use watermarks for surveillance or user profiling. The company’s support page frames watermarking primarily as a compliance and transparency tool.
There is also the risk that watermarks create a false sense of security. Platforms may rely on automated detectors and fail to investigate nuanced cases where AI content has been mixed with human writing. The EU AI Act acknowledges this by framing watermarking as one part of a broader transparency ecosystem that also includes model documentation, data summaries, and human oversight.
Anthropic’s announcement is a notable step forward, but it is also a starting point. The company will need to refine its watermarking technology in response to real-world use, adversarial attacks, and feedback from the research community. As AI models become more capable and more widely used, the ability to certify the origin of digital content will only become more important.
Source: TechCrunch News