Anthropic now watermarks all Claude text globally

Anthropic will embed invisible watermarks in all Claude-generated text starting this month, a change that applies globally and covers every Claude product. The company signed the EU AI Act Code of Practice, but the labeling requirement extends far beyond Europe to include the API, Claude Code, Claude Cowork, and Claude Tag.

Claude models launching on or after August 2, 2026 ship with watermarking built in. Generated text carries an invisible mark that survives copy-paste and, according to Anthropic, "may persist through some editing." Files like images (.svg, .png, .jpg) get signed provenance metadata using the C2PA standard, which can reveal later tampering.

How the watermarking actually works

Anthropic uses two labeling methods. Text watermarks are embedded at the model level, so the product you use does not matter. The watermark does not change meaning, quality, or readability. It persists when you copy text into another document.

For supported file types, Anthropic applies C2PA-standard provenance metadata. The Coalition for Content Provenance and Authenticity developed this open standard, and the digital signature indicates that Claude processed the file. Cloud partners like AWS, Google Cloud, and Microsoft Foundry should support text watermarks, though Anthropic notes those platforms may not carry the signed metadata.

Existing Claude models get a transition period under the EU law, but Anthropic says it is already retrofitting them. The company plans to release verification tools for users and third parties, though no date has been set.

What watermarks cannot prove

Anthropic is direct about the limits. A detected watermark does not prove Claude wrote the content. People use Claude for proofreading, translation, and summarizing their own writing. Text with a watermark could still contain entirely human ideas.

No watermark does not clear things up either. The model might predate the watermarking rollout. The text might be too short for reliable detection. Heavy editing, translation, format conversion, or a simple screenshot can strip the marks entirely.

“A detected watermark doesn't mean Claude actually wrote the content. People use Claude all the time for proofreading, translating, or summarizing, so the output might carry a watermark even though the ideas came from a human.”

— Anthropic

The practical test is durability. If these watermarks hold up through editing and reformatting, third-party detection tools could integrate Anthropic's signal. That would be more reliable than proprietary detectors like Pangram, which do not disclose what triggers their results.

Where this fits in the AI detection landscape

Anthropic is not the first to move here. Google DeepMind open-sourced its SynthID system and built it into Gemini. SynthID tweaks probability values during token prediction, creating a watermark without degrading text quality. It works across languages but struggles once text is edited.

OpenAI has reportedly sat on a text detector with 99.9 percent accuracy for about two years. The company has not released it, citing concerns about circumvention through translation or rewriting, the risk of stigmatizing certain groups, and likely worries about the impact on its own business.

Detection remains especially fraught in education. Unreliable detectors have led to false cheating allegations. But there are real reasons schools want to know when AI was used. Studies suggest heavy reliance on AI tools weakens critical thinking and writing skills, particularly among students who treat them as shortcuts. Scammers have enrolled fake students at US colleges, used AI to complete coursework, and collected financial aid.

What developers need to do

Developers integrating Claude into their products must determine which Article 50 requirements apply to their services. Anthropic is explicit that the burden falls on downstream builders, not just the model provider.

For teams building on Claude via API, the watermark is automatic. But disclosure obligations, user-facing labeling, and compliance documentation remain the integrator's responsibility under EU law.

Claude is popular for knowledge work, including among students. Watermarking could affect that user base if schools start checking. Whether this shifts usage patterns or just pushes users toward models without detectable marks remains an open question.

Source: The Decoder / Matthias Bastian