How Claude’s Text Watermark Works: What It Detects, What Survives and Who Can Check It

TL;DR

• What: since August 2026, Anthropic embeds an invisible statistical watermark in text generated by Claude, worldwide and across all its products.
• Why: to comply with the EU AI Act and the EU Code of Practice on transparency of AI-generated content.
• How: the method, SynthID-Text from Google DeepMind, uses a secret key to steer which of several equivalent words Claude picks.
• Limits: it survives copy-paste and light edits, but not complete rewrites; it shows Claude’s involvement, not human authorship.
• Detection: watermark checking is in private preview for eligible organizations; images and files use C2PA Content Credentials.

Contents

1. What Anthropic announced
2. How the text watermark works
3. What survives, and what doesn’t
4. What the watermark can and can’t tell you
5. Who can detect it
6. Images and files: C2PA Content Credentials
7. What it means for businesses and creators
8. Vyara’s take

Text written with Claude now carries an invisible mark that can reveal the model was involved. Here is how the system works, how robust it is, and what it can and can’t prove.

What Anthropic announced

In August 2026, Anthropic announced that text generated by Claude models will carry a watermark. The move implements requirements of the EU AI Act and the EU Code of Practice on transparency of AI-generated content, signed in July 2026.

Anthropic applies it globally rather than only in the EU. According to Claude’s help center, marking covers the Claude apps, Claude Code, the Claude Platform API and Anthropic’s other products, as well as supported cloud deployments. Models released before August 2, 2026 are being updated over the following months under the law’s transition period.

Anthropic says the watermark adds no extra tokens, doesn’t change pricing and has a negligible impact on speed.

How the text watermark works

When a language model writes, it repeatedly chooses between words with similar meaning, such as “overcast” and “grey”. These choices are normally random.

Anthropic’s watermark changes the source of that randomness. A cryptographic key subtly steers the pattern of word choices. As Anthropic puts it: “the words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key.”

The method is SynthID-Text, published by Google DeepMind in Nature in 2024. It belongs to the same family as earlier academic schemes, such as the “green list / red list” approach, in which a detector counts how often a text uses words from a secret subset of the vocabulary.

The watermark is invisible. It doesn’t change meaning, quality or formatting, and it contains no information about the user.

What survives, and what doesn’t

What the watermark can and can’t tell you

According to Anthropic, a positive result means that “Claude was likely involved with the content at some point.” The watermark:

• can’t distinguish text Claude wrote from text Claude heavily edited;
• doesn’t confirm whether a text was written by a person;
• can’t tell whether a text came from a different AI model;
• can’t prove a text is human if no mark is found, because heavy editing, translation or mixing with other content can remove it.

Who can detect it

Watermark detection is in private preview for eligible organizations, including regulators, law enforcement, media, fact-checkers, researchers, educators and EU civil society groups, plus enterprises that need compliance verification. Anthropic plans wider access later.

Until then, no public tool can check Claude’s text watermark. Third-party “AI detectors” work differently: they look for statistical and stylistic patterns, which leads to false positives on human writing and misses on edited AI text.

Images and files: C2PA Content Credentials

For supported files such as PNG and JPEG, Claude attaches Content Credentials based on C2PA, an open provenance standard. The C2PA coalition was founded in 2021 by Adobe, Arm, BBC, Intel, Microsoft and Truepic and has since been joined by AI companies including OpenAI.

C2PA works like a passport for a file: signed metadata that records where the content came from and how it was edited. Anthropic offers a free Claude Content Checker for these marks.

Social platforms read the same signals. Meta labels images on Facebook, Instagram and Threads using C2PA and IPTC metadata added by tools from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock. That’s why an edited photo can receive an “AI info” label even when the original photo is real.

What it means for businesses and creators

• Disclosure beats concealment. Content can be traced back to AI involvement, so an open policy on AI use protects credibility better than hiding it.
• Agree on internal rules. Define when AI use is acceptable, for which tasks, what data never goes into AI tools and who reviews the output.
• Keep a human in the loop. The real risk in organizations is AI-written documents passed along without anyone reading them.
• Treat detector results with caution. A pattern-based detector flag is not proof, and neither is a clean result.
• Write in your own voice. Generic phrasing, repetitive rhythm and grandiose statements make any text weaker, whether it’s flagged or not.

Vyara’s take

“The watermark shows that Claude was involved, not that a person didn’t think. I use AI openly for research and drafts, and I review and sign everything I publish. In a world where anything can be generated, the most valuable currency is trust: in the person behind the tools, not in the tools themselves.”

Further reading: Claude now watermarks its text: should you worry, and what actually matters – Vyara Stefcheva’s analysis on stefcheva.com.

Sources: Anthropic – How Claude’s text watermarking works · Claude Help Center – How Claude marks AI-generated content · Microsoft – C2PA founding announcement · Meta – Labeling AI-generated images on Facebook, Instagram and Threads

About the Author

Vyara Stefcheva

Vyara Stefcheva is the founder and editor of Egorithms and an independent marketing expert. She follows how technology, AI and algorithms change the way people find, choose and buy, and writes about what that means for businesses and for the people who use them.

She has spent her career building marketing and business strategies for more than 300 brands across three continents. She built and exited a marketing agency and now works independently. She has taught more than 20,000 people through lectures, courses and corporate training, and speaks at conferences about AI and the future of the internet.

On her YouTube channel she tests new AI and marketing tools as they launch and gives each an independent verdict: adopt, watch or ignore.

See author's posts