Anthropic has confirmed that it will begin watermarking text generated by its AI models, including its flagship Claude assistant, citing the need to comply with European transparency regulations. The company updated its support documentation to explain the new policy, which applies to all models released after August 2.
The move comes as the European Union's AI Act begins to take effect, placing new obligations on developers of general-purpose AI systems. Under the EU AI Act's Transparency Code, companies must clearly mark AI-generated or AI-edited content in a way that other systems can identify. The rules took effect on August 2, and Anthropic has now clarified how it will meet those requirements.
What Anthropic has said
Anthropic said that all models released after August 2 will automatically include technology designed to watermark both computer-generated text and files. For files, the company will use the C2PA open standard, a technical protocol used to trace the origin and history of digital content. For text, the watermark is embedded at the model level, meaning it will be present regardless of which Claude product or surface the text comes from.
"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 reads. This addresses a key concern about AI-generated content: once text is copied or lightly edited, provenance information is often lost. Anthropic says its approach is designed to survive those common actions.
The company said it will also extend support for older models, though it did not specify exactly how far back that support will go or when it will be added.
The EU AI Act and transparency requirements
The EU AI Act is a comprehensive piece of legislation that classifies AI systems according to risk. For general-purpose models, transparency obligations are a core part of the framework. The Transparency Code specifically requires providers to make it possible to identify machine-generated content, especially when it is shared publicly.
The requirement is intended to help people and automated systems understand when they are interacting with AI-produced material. This includes not only text but also images, audio, and video. The rules align with broader efforts to prevent deepfakes, misinformation, and scams that use AI-generated content to deceive people.
Anthropic is not alone. Several other major AI companies have committed to adhering to the EU's code, including Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia. The commitments vary in detail, but they all involve some form of content provenance or watermarking.
How watermarking works
Text watermarking is a technical challenge. Unlike images or video, which can carry hidden markers in pixels or compression artifacts, text is symbolic and easily altered. Common approaches include embedding statistical patterns into word choice, sentence structure, or punctuation. These patterns are designed to be invisible to most readers but detectable by algorithms.
Anthropic has not revealed the exact method it will use for text watermarking. The company said only that the watermark will be applied at the model level. That means any output produced by a Claude model, whether through a web interface, API, or third-party application, will carry the mark.
For files, C2PA (Coalition for Content Provenance and Authenticity) provides a more standardized method. C2PA metadata is cryptographically signed and can include information about how a file was created, by what software, and what modifications were made along the way. This makes it harder for malicious actors to strip away provenance without leaving signs of tampering.
Which products are affected
The watermarking requirement will apply to a range of Anthropic products. The company listed the Claude platform API, Claude, Claude Code, Claude Cowork, and Claude Tag as areas where watermarking will be present. Claude Code is Anthropic's agentic coding assistant, while Claude Cowork and Claude Tag appear to be newer tools or features designed for different use cases. The key point is that the watermark is not tied to any single interface; it is a property of the underlying model.
This is significant for developers and businesses that build on Anthropic's APIs. If the watermark is embedded at the generation stage, it will appear in any application that uses a Claude model, unless the provider offers a way to disable it. Anthropic has not indicated whether there will be options to opt out for legitimate uses, such as creative writing where provenance might be less critical.
Open questions and limitations
One important question is how much editing a person must do to remove the watermark. Anthropic acknowledged that the watermark "may persist through some editing," but it did not define what "some editing" means. If the watermark relies on statistical patterns distributed across a piece of text, then heavy rewriting, translation, or paraphrase could potentially wash it out. The company said it has been asked for clarification, but has not yet provided additional details.
There are also concerns about false positives. Watermarking schemes can sometimes misidentify text that was not generated by AI, especially if the detector is overly sensitive. That could have consequences for writers, students, and journalists who produce natural language and might find their work flagged as machine-generated.
Another issue is compatibility. If Anthropic's watermark is only detectable by Anthropic's own tools, other platforms and researchers may not be able to verify it. The C2PA standard helps for files, but text watermarking has no universal standard. This could limit the practical usefulness of the system for independent oversight.
Industry context and reactions
The announcement comes amid a broader push by platforms to label AI-generated content. An increase in public backlash against undisclosed AI material has led companies to take action. AI music platform Suno said last week that it will mark tracks created on its platform after a series of legal challenges. The company now includes metadata to identify AI-made music, making it easier for streaming services and consumers to know what they are listening to.
In the publishing space, newsletter service Substack has teamed up with the firm Pangram to flag AI-generated content. Substack CEO Chris Best used the term "Claudefishing" to describe people using AI to produce content that appears to come from a human with a particular perspective or style. The term is a play on "catfishing," and it highlights the growing concern that AI can be used to impersonate credible authors or manufacture authenticity.
Why this matters
Watermarking alone will not solve every problem created by generative AI. The technology is evolving quickly, and so are techniques for spoofing or stripping content provenance. Still, the adoption of watermarking at the model level is an important step toward accountability.
When AI-generated content can be reliably identified, it becomes easier for social media platforms, search engines, and news organizations to handle it appropriately. It also gives users the information they need to make judgments about what they are reading. The EU AI Act's transparency rules are designed to push the entire industry toward that baseline.
Anthropic's decision to apply watermarking across all its models after August 2 is one of the clearest examples yet of a major AI company operationalizing those requirements. The company's explicit acknowledgment that the watermark will travel with copied text addresses one of the most common ways content spreads across the internet. Whether it can withstand more aggressive edits, however, remains to be seen.
Developers and enterprises that rely on Claude models should begin to assess how the watermark might appear in their products and what implications it could have for user trust, content moderation, and compliance. For everyday users, the change may be invisible at first, but it could eventually become a routine part of how AI-related content is handled online.
Source: TechCrunch News