Anthropic Will Add Invisible Watermarks to Claude-Generated Text

Anthropic is preparing to add invisible AI watermarks to text generated or processed by Claude, including code produced with Claude Code. The move is tied to new European Union requirements around disclosure of AI-generated content and is intended to make AI-assisted writing easier to identify.

Claude Will Watermark Generated Text and Code

Anthropic’s approach goes beyond what the EU AI Act requires. The company plans to watermark not only text generated by Claude, but also content that has been processed or edited by the AI, including code.

The policy has received criticism from some Claude users, with complaints that the approach is unnecessary or unethical. However, supporters argue that identifying AI-generated content can improve transparency and encourage people to disclose when AI has played a role in creating their work.

For example, someone using Claude to polish a job application could disclose that assistance to an employer. The same principle could apply to businesses using Claude to produce internal documents or other communications.

The EU Rules Have Several Exceptions

The EU AI Act provides a broader framework for AI transparency, but its watermarking requirements come with several important exceptions.

Computer code is exempt from the watermarking provisions, even though Anthropic plans to watermark code generated by Claude. The rules also include an exemption for “standard editing” and require AI providers to apply watermarking only “as far as this is technically feasible.”

Those provisions could leave considerable room for interpretation when determining which AI-assisted content needs to carry an identifier.

AI Watermarks Cannot Prove Human Authorship

There are also significant technical limitations.

Claude’s watermarks are designed to survive simple copy-and-paste operations and light editing, but they aren’t impossible to circumvent. AI-generated text could potentially be passed through an open-weight model that does not apply its own watermark, removing the original identifier.

That means the absence of a Claude watermark does not prove that text was written entirely by a human. Anthropic itself acknowledges this limitation.

Watermarking Won’t End AI Slop

The EU’s push toward greater transparency could make AI use easier to disclose and verify, but watermarking alone is unlikely to solve the broader problem of low-quality AI-generated content.

AI watermarks can provide another signal about how content was produced, but they cannot reliably establish human authorship when a piece of text has been processed through multiple systems.

The result is likely to be a useful step toward AI transparency and accountability, rather than a definitive solution to the growing problem of undisclosed AI-generated content.