Anthropic has begun embedding invisible watermarks into everything Claude writes, a move the company says is designed to meet new European Union transparency rules that came into force this month. The change applies to every Claude model launched from 2 August 2026 onwards, across the consumer app, the developer API, Claude Code and Claude Cowork, as well as versions of the assistant offered through AWS, Google Cloud and Microsoft Foundry.
The company confirmed the marking applies globally, not just in Europe. Anthropic said all Claude models launched from 2 August 2026 onwards will include the feature at release, and that it is working to add it to older models.
How the Watermark Works
The system uses two separate techniques depending on what Claude produces. Generated text carries an imperceptible pattern woven directly into the wording itself, which can survive copying and pasting but does not change the meaning or readability of the response.
For image formats such as .svg, .png or .jpg, the system adds signed provenance metadata similar in principle to EXIF data on a photograph, which signals that the file passed through Claude and can help detect tampering. Anthropic has also acknowledged the system's limitations. The file metadata can be stripped through format conversion, screenshots or re-saving, and heavy editing or translation of text can mean genuine Claude output goes undetected entirely.
We’ve written an FAQ to answer some of the questions we've received about watermarking.
— Anthropic (@AnthropicAI) August 14, 2026
In summary:
• We’re implementing watermarking to comply with the EU AI Act. Other major model developers have signed the same Code of Practice and will also be implementing watermarking;…
Why the EU Forced the Issue
The trigger is Article 50 of the EU's AI Act, which requires providers of generative AI systems to mark synthetic content in machine-readable form. Those transparency obligations took effect on 2 August 2026.
Non-compliance carries serious financial risk. According to the EU AI Act text, fines can reach up to €15 million or 3% of a company's total worldwide annual turnover, whichever figure is higher.
Anthropic signed the bloc's Code of Practice on Transparency of AI-Generated Content, a voluntary framework that grants a presumption of meeting the Article 50 standard. Almost 200 companies had signed by late July, including Microsoft, Google, Meta and OpenAI.
Anthropic chose to apply the watermark globally rather than only to EU users, a decision that effectively exports a Brussels regulation into product behaviour used worldwide.
What It Means for Claude's Millions of Users
The practical fallout has already sparked frustration among users, many of whom argue the system misreads how people actually use AI tools. Writers who ask Claude only to proofread or lightly edit their own work have raised concerns that their human-written text could still pick up a mark meant for AI-generated content.
Anthropic maintains that a detected watermark is a signal, not proof of full authorship, since text that has been heavily edited, translated or run through an older, unmarked model may show no trace of AI involvement at all. The company is not acting alone. Google DeepMind has already deployed SynthID to mark AI-generated text, images and audio.
OpenAI has also signed the same EU code but has held off on watermarking ChatGPT text, attributing the delay to concerns over false positives and competitive risk. The watermark rollout signals a broader shift in how AI-generated content will be tracked across the internet, not just within the EU.
For readers, students, employers and publishers, it means the line between AI-assisted and AI-authored text is about to become far more visible, and far more contested, as detection tools mature and other major labs follow Anthropic's lead under the same regulatory pressure.
