Anthropic is watermarking Claude's text to comply with EU law, here's how it works

Key points

• Anthropic announced on 17 August 2026 that Claude will embed invisible watermarks in AI-generated text to meet EU legal requirements.
• The system uses a version of SynthID-Text, an open-source watermarking technology built by Google DeepMind.
• Anthropic says the watermarks will not change the quality of Claude's responses or increase costs for users.
• Google's Gemini chatbot has used the same SynthID-Text approach since 2024; OpenAI has not yet detailed its watermarking plans.

When Claude writes you a paragraph, a hidden pattern is now woven into the words you read. You cannot see it. A detection tool with the right key can.

On Friday, Anthropic revealed that its AI assistant Claude will add invisible watermarks to the text it generates. The company says it is using a version of SynthID-Text, an open-source watermarking system built by Google DeepMind that encodes detectable patterns into AI writing by nudging which words get chosen at certain moments.

How does invisible watermarking actually work?

It works by exploiting moments where the choice of word barely matters. Think of a sentence like "The weather today was cold and..." The next word could be "overcast" or "grey" and the meaning stays identical. Normally, the AI picks between these near-equivalent options using a random number generator. With watermarking switched on, it instead uses a secret key combined with the surrounding words to make that choice, leaving a faint but consistent pattern across hundreds of such decisions in a single piece of text. A reader never notices; a detector holding the matching key does.

The system carries no cost to users. Anthropic is clear that the watermarks will not alter the quality or content of Claude's outputs.

Why is this happening now?

Europe's AI Act, the bloc's sweeping new law covering artificial intelligence products, requires that AI-generated audio, images, video, and text all carry machine-readable marks so people and tools can identify synthetic content. Anthropic must comply because it operates in the EU.

As first reported by The Verge AI, Claude is not alone here. Google's Gemini already uses SynthID-Text for the same purpose, having adopted it in 2024. OpenAI has not yet published a text-watermarking plan for ChatGPT, though it falls under the same legal obligations.

Anthropics's announcement also covers images: the company is adding C2PA support, a technical standard that embeds verifiable origin data, to images Claude processes.

What does this mean for people using Claude?

Practically speaking, nothing changes in how Claude feels to use. The watermark is silent, invisible, and free. The benefit is societal rather than personal: journalists, platforms, and regulators gain a tool to verify whether a piece of text came from an AI, which matters when detecting misinformation or enforcing disclosure rules.

Common questions

Can I remove the watermark from Claude's text?

Anthropically hasn't published details on how resistant the watermark is to editing, but SynthID-Text is designed to survive minor rewording. Heavy rewriting could degrade it.

Does this mean all AI text will be detectable from now on?

Not yet. Detection requires access to the right key, and coverage depends on which AI companies implement watermarking. OpenAI has not yet announced a comparable system for ChatGPT.