Anthropic adds invisible watermarks to all Claude outputs starting August 2026
Starting August 2, 2026, every Claude model Anthropic releases carries an invisible machine-readable watermark in generated text and signed provenance metadata in generated images. Anthropic confirmed the global rollout on August 11, after joining the EU AI Act Code of Practice. The requirement applies worldwide, not just to EU users.
What changed
Every new Claude model per Anthropic's Transparency Hub embeds an invisible watermark in text outputs at the model level, meaning the signal travels with the text regardless of which Claude product generated it. Per Anthropic, the watermark does not change the meaning, quality, or readability of generated text, survives copy-paste, and "may persist through some editing." Heavy paraphrasing, translation, or format conversion can remove it entirely.
For generated files (.svg, .png, .jpg), Anthropic applies signed provenance metadata following the C2PA standard, developed by the Coalition for Content Provenance and Authenticity. The C2PA signature records that Claude processed the file and can detect later tampering.
The scope covers Claude.ai, the Platform API, Claude Code, Claude Cowork, and Claude Tag. Text watermarks also extend to Claude models accessed through AWS, Google Cloud, and Microsoft Foundry per The Decoder, though Anthropic notes those platforms may not support the signed C2PA metadata for files. Existing Claude models get a transition period under the EU AI Act, and Anthropic says it is already working to retrofit them. A detection API for third parties is planned, with no release date announced.
Why it matters
API operators and enterprise teams using Claude through any channel now receive watermarked outputs by default. For legal and compliance teams, the watermark creates a new provenance signal: rather than relying on self-reported disclosure policies, outputs carry a machine-readable mark from the point of generation.
Two limits shape what the system can and cannot prove. A positive watermark does not confirm Claude was the original author; text can carry the watermark when Claude was used only to proofread or translate human-written content. No watermark is equally inconclusive: the generating model may predate the rollout, the text may have been edited substantially after generation, the passage may be too short for reliable detection, or metadata may have been stripped through format conversion.
Developers building products on top of Claude through the API must assess which Article 50 requirements of the EU AI Act apply to their own services, per Anthropic. That analysis is the developer's responsibility, not Anthropic's.
What to watch next
The practical test is whether these watermarks survive real-world editing workflows. Google DeepMind open-sourced its SynthID text watermarking system and built it into Gemini, but SynthID struggles with text edited after generation per The Decoder. If Anthropic's approach holds up better, it raises pressure on OpenAI, which has held an internal text detector with reported 99.9% accuracy for roughly two years without releasing it. Watch also for whether AWS, Google Cloud, and Microsoft Foundry update their enterprise contract terms to require disclosure of watermarking to downstream customers.