Anthropic is now putting an invisible watermark on text generated by its Claude models. The change, which applies to models launched on or after August 2, comes in response to new AI transparency rules that took effect in Europe and California this week. The watermark itself may not matter much to most companies. The regulatory requirements behind it could.

Bottom Line

• Laws generally do not require disclosing every use of AI. Under the EU AI Act, the disclosure duty for AI-generated text generally applies only when that text is published, informs the public, and concerns a defined list of public interest topics, and even then it disappears where the content underwent genuine, documented human review or editorial control by someone holding ultimate legal responsibility for the publication.
• Anthropic's compliance with the AI Act does not necessarily make your company compliant. The watermark is designed to satisfy its own obligation as a provider under the AI Act. It does not, by itself, satisfy your company's separate obligations as a “deployer” of AI, and if you have embedded a vendor's model into your own branded product, you may be the “provider” for that system rather than a deployer.
• The EU's list of "public interest" topics is specific and does not automatically include ordinary marketing or one-to-one client communications. Public-facing reports, commentary, and thought leadership touching consumer safety, public health, or financial and economic developments deserve a closer look.

What Changed and Why

Every Claude model launched on or after August 2 carries an invisible statistical watermark in its generated text, along with signed provenance metadata for supported file types. There is no visible label and no way to turn it off. Anthropic disclosed the change through a help center update rather than a product announcement, a fair signal that this is compliance, not a new feature. 

The watermark is designed to satisfy Anthropic's obligation as a provider under the EU AI Act's Article 50, the same framework roughly 190 companies, including top tech companies, have committed to. What it does not do is answer the separate question that any company whose AI-generated content reaches people in the EU has to answer for itself.

The Test That Actually Determines Your Exposure

Article 50 does not create one blanket duty to disclose AI use. It creates four distinct obligations, split between providers and deployers, and only one is likely to reach most companies' day-to-day content.

• Chatbots and AI assistants that interact directly with people: the provider, meaning whoever develops the system and puts it into service under its own name, must design it so people know they are interacting with AI, unless that is obvious. If your company embeds a vendor's model into its own branded chatbot or assistant, you are very likely the provider for that specific system, not a deployer relying on someone else's compliance, and that changes what you are responsible for building in.
• Emotion recognition and biometric categorization systems: deployers must inform anyone exposed to the system. Few companies will trigger this one.
• Deepfakes: deployers must disclose, with a lighter touch for content that is evidently artistic, satirical, or fictional.
• AI-generated or manipulated text: deployers must label it, but only where three things are true together. The text is published. It informs the public. And it concerns a matter of public interest, a category the Commission defines specifically: political and democratic processes, public administration and public services, the administration of justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific, or cultural developments open to public debate.

The labeling duty does not apply where the text underwent genuine human review or editorial control, meaning substantive examination by someone with real judgment over the content, and a person or organization holds ultimate legal responsibility for the publication. A grammar check does not count. Real editorial judgment does, and it needs to be something you can point to, since an informal or undocumented review is hard to rely on if the question is ever asked.

That is the actual question for your company. Whether the content you publish that could plausibly touch one of those public interest categories goes through a review process that would meet that standard, and whether you could show it if asked.

What the Watermark Does and Does Not Do for You

A watermark shows that Claude touched a piece of text somewhere along the way. It does not show that Claude wrote it from scratch, and it does not show that a human did not substantially contribute. The marking duty itself belongs to Anthropic as the provider, not to your company as the deployer, so you do not need to add your own watermark on top of it. 

What it does not do is stand in for your own disclosure obligation. 

The Commission has been explicit on this point for deepfakes: a provider's invisible, machine-readable mark does not satisfy a deployer's separate duty to give people a clear, visible disclosure. That principle should extend to text disclosure as well, even though the Commission's guidance addresses it specifically in the deepfake context. An invisible watermark is not a substitute for your own review and disclosure process.

Two Other Regimes to Watch

 Five Practical Steps Related to Your Company's AI Content

1. Inventory What You Publish 

Identify AI-generated or AI-assisted text your company puts out externally: marketing, reports, blog content, public-facing commentary. Set aside anything that could plausibly touch public health, consumer safety, or financial and economic developments, since that is the category the EU rule actually reaches.

2. Test That Content Against the Three-Part Rule

For anything flagged in step one, check whether it is published, informs the public, and concerns a listed public interest topic. If all three are true, confirm whether someone with real subject matter judgment substantively reviewed it and holds responsibility for the publication, and whether that review is actually documented.

3. Close the Gaps That Matter

Where the review does not meet that standard, either build a genuine review step or add a clear, visible disclosure that the content was AI generated.

4. Check Your Chatbots and Your Files 

If you built or commissioned an AI assistant using a vendor's model, confirm whether that makes you the provider responsible for the AI disclosure rather than just the deployer, and make sure the disclosure happens in the interaction itself, not buried in terms of service. Do not treat an invisible vendor watermark as satisfying that duty.

5. Assign Ownership

Someone, legal, compliance, marketing, or a cross-functional group, needs to own tracking this across the organization. Most gaps trace back to no one being responsible for the question.

Providers are already marking their output. The next wave of enforcement will turn on whether the companies using that output can show real human judgment stood behind what they published, not just a policy that says so.