# Anthropic Adds Invisible Watermarks to Claude's AI-Generated Text

Page: https://hirantha.me/blog/anthropic-adds-invisible-watermarks-to-claudes-ai-generated-text
Published: 2026-08-15
Written by: hirantha.me Newsroom (AI-generated). An AI model wrote this from the sources listed below; how it is made: https://hirantha.me/about/newsroom
Topics: anthropic, claude, ai watermarking, eu ai act, c2pa

> Anthropic now embeds an invisible watermark in text from newer Claude models to meet EU AI Act transparency rules.

## Two marks, not one

Anthropic has started embedding an imperceptible watermark directly into text generated by Claude, its family of artificial intelligence (AI) models, alongside a separate system for marking files. According to Anthropic's own documentation, the two mechanisms work differently. The text watermark is woven into the output itself and does not change its meaning, quality or readability; it travels with the text when copied and pasted, and can survive some editing. For files, such as .svg, .png and .jpg images that Claude generates, Anthropic instead attaches signed provenance metadata following the Coalition for Content Provenance and Authenticity (C2PA) standard, the same open specification Google and Adobe use to record how a piece of content was created or edited.

## What triggered it

The change is a compliance move. Anthropic signed the European Union's (EU) Code of Practice on Transparency of AI-Generated Content, a voluntary framework built around Article 50(2) of the EU AI Act, the bloc's law governing artificial intelligence. That article requires providers of AI systems that generate synthetic content to make outputs identifiable as AI-generated. Anthropic is not alone in signing on: more than 200 companies, including Meta, Google and Microsoft, have agreed to the same code of conduct.

The rollout is tied to model release dates rather than being retrofitted everywhere at once. Claude models launched in the EU on or after August 2, 2026 support marking from the moment they ship. Anthropic says it is separately working to add the capability to models that launched before that date. Once a model supports marking, the marks apply to its output globally, not only to EU users, and cover Claude's consumer app, the Claude API (application programming interface, the interface developers use to integrate Claude into their own software), Claude Code, Claude Cowork, Claude Tag, and Claude access sold through cloud partners Amazon Web Services, Google Cloud and Microsoft Foundry.

## What the marks can and cannot tell you

Anthropic is explicit that the system is not a definitive authorship test. A detected watermark means the text was processed by a supported Claude model at some point, not necessarily that Claude wrote it unaided. The company gives its own example: a document a person wrote by hand, then ran through Claude for proofreading, translation or summarizing, would carry the same watermark as text Claude generated from scratch. The marks can also disappear. Anthropic says text that is heavily edited, paraphrased, translated or mixed with other content may lose its watermark, and file metadata can be stripped by converting formats or taking a screenshot. Anthropic has not published the detection method itself, saying only that it plans to release technical documentation to let users and third parties check for its marks.

## How this compares to what else is out there

Anthropic is a relative latecomer among major AI labs on this specific feature. Google DeepMind's SynthID, launched in 2023, embeds similar imperceptible watermarks into AI-generated text, images, audio and video, and Google says the technology has now been used to watermark more than 20 billion pieces of content. SynthID has also spread beyond Google's own products: companies including OpenAI, Kakao and ElevenLabs have adopted it for parts of their own generated content. Separately, streaming service Spotify has taken a more manual approach for one narrow case, labeling AI-generated artist personas on its platform and excluding them from algorithmic recommendations.

What sets Anthropic's approach apart is less the watermarking technique, which is broadly similar in concept to SynthID, and more the explicit legal hook: Anthropic is framing this specifically as an EU AI Act compliance obligation rather than a voluntary trust-and-safety feature, and applying it worldwide once a model qualifies rather than geofencing it to the EU.

## What it means in practice

For everyday Claude users, nothing about the writing experience changes: outputs read the same, and there is no opt-out mentioned in Anthropic's documentation, meaning the mark is not something a user can turn off. For people trying to detect AI-written text, whether teachers, editors or platform moderators, the watermark is a new signal but not a reliable one on its own. Because Anthropic has not yet published its detection tooling, third parties currently have no independent way to check for the mark, and even once that tooling ships, the company's own caveats mean a negative result never proves a document is human-written, only that no detectable mark was found. The more durable effect is regulatory: this is one of the first concrete, dated examples of the EU AI Act's content-transparency rules changing how a major AI provider ships text output by default, and it is a preview of what other signatories to the same code of practice, including Google, Meta and Microsoft, are also committed to rolling out.

## Questions

**Can I turn off Claude's watermark?**
Anthropic's documentation does not describe an opt-out; marking applies automatically to output from supported models.

**Does a Claude watermark prove text was written entirely by AI?**
No. Anthropic says the mark only shows the text was processed by Claude at some point, which includes cases where a person wrote the text and had Claude proofread, translate or summarize it.

**Which Claude models are watermarked?**
Models launched in the EU on or after August 2, 2026 support marking from launch; Anthropic says it is working to extend support to earlier models.

**Why is Anthropic doing this now?**
It is fulfilling commitments under the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, which Anthropic has signed along with more than 200 other companies.

**Is this similar to Google's SynthID?**
It is conceptually similar, an imperceptible watermark embedded in generated content, though Anthropic has not published its detection method, while Google's SynthID has been used to watermark over 20 billion pieces of content since 2023 and is adopted by other companies including OpenAI.

## Sources

- [How Claude marks AI-generated content](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content), Claude Help Center (Anthropic)
- [Anthropic's Claude will watermark AI-generated text. Here's how it works](https://globalnews.ca/news/12018450/ai-anthropic-claude-watermark-text/), Global News
- [Anthropic adds invisible watermarks to Claude-generated text](https://cyberinsider.com/anthropic-adds-invisible-watermarks-to-claude-generated-text/), CyberInsider
- [SynthID - Google DeepMind](https://deepmind.google/models/synthid/), Google DeepMind