Anthropic Reveals How Claude’s Invisible Watermark Works

Anthropic is using statistical patterns in Claude-generated text to create a machine-detectable signal without changing how the text appears to readers.

Aug. 17, 2026: Last week Anthropic disclosed that it is using statistical patterns in Claude-generated text to create a machine-detectable watermark, rolling out globally as new EU AI Act transparency rules take effect.

How Does Claude’s Invisible Watermark Work?

The watermark is not a visible label, hidden character or separate piece of metadata. Instead, it is created during text generation by subtly influencing Claude’s statistical token choices. Across a sufficiently long passage, those choices form a pattern that specialized detection software can analyze. Anthropic says the watermark does not change the meaning, quality or readability of Claude’s responses. “The watermark is imperceptible to humans and does not change the text.”

Anthropic says the signal is embedded directly into the generated text and is designed to travel with the content through copying and some forms of editing.

The approach is similar to statistical watermarking demonstrated by Google DeepMind’s SynthID-Text. Google DeepMind’s research describes a system that modifies token sampling during generation to create a detectable statistical pattern while maintaining text quality.

Anthropic’s watermark is not intended to prove that Claude authored an entire piece of content. If a user supplies human-written material and asks Claude to translate, edit or summarize it, the resulting output can still carry a Claude watermark.

That distinction makes the technology a provenance signal rather than definitive proof of authorship.

The Watermark Has Limitations

The signal can survive copying and some light editing, but extensive rewriting can weaken or remove it. Short passages can also be difficult to identify reliably because statistical watermarking requires enough generated text to establish a detectable pattern.

The official statement said, “Anthropic is developing detection tools for users and third parties, with a detection API planned as part of the broader rollout.”

For images and supported files, Anthropic is using a different system: Coalition for Content Provenance and Authenticity (C2PA) metadata. The metadata can provide information about a file’s origin and subsequent changes, but it can also be removed through processes such as screenshots, conversions or platforms that strip metadata.

EU Rules Are Driving the Shift

The rollout comes as transparency requirements under Article 50 of the EU AI Act take effect. The regulation requires providers of AI systems generating synthetic audio, images, video or text to ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, where technically feasible. 

“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”

AI systems placed on the market before August 2 have a transition period for the Article 50(2) marking obligation until December 2, 2026

The requirement does not prescribe Anthropic’s specific watermarking technology. Instead, the law establishes the transparency objective while allowing providers to implement technically appropriate solutions.

Anthropic is applying its marking system globally, meaning users outside Europe can also receive Claude content carrying the signal.

Users Raise Concerns

The rollout has also triggered criticism from some Claude users. Business Insider reported that dozens of users said they were cancelling their Claude subscriptions, with developers, consultants and researchers raising concerns about how watermarked content could be treated in professional or academic settings. Some users reportedly said they were moving to other AI tools.

Anthropic, however, told Business Insider that it had not observed an increase in cancellations attributable to the watermarking rollout.

The development marks another step toward AI content provenance. But the central limitation remains: detecting Claude’s involvement is not the same as proving that Claude authored the underlying work. As AI becomes more deeply integrated into writing, translation, editing and software development, that distinction could become increasingly important for businesses, publishers and educators.

Mansi Hake