Starting August 2, 2026, every Claude model released carries an imperceptible, machine-readable watermark woven directly into generated text and code. The watermark survives copy-paste and moderate editing, embedded in the statistical patterns of word choice rather than visible formatting.
Files generated or processed by Claude receive signed C2PA provenance metadata, the same industry standard already used by Adobe, Microsoft, and others to label AI-generated media. Anthropic says older models will be retrofitted during a transition period, while detection tools will be published for third parties to verify content.
The regulatory trigger is EU AI Act Article 50, which became enforceable August 2 for newly launched systems. Anthropic applied the requirement globally rather than creating separate inference paths for different regions.
Why It Matters for Your Stack
The watermark lives at the model level, so it appears whether you call Claude via API, claude.ai, Claude Code, AWS Bedrock, Google Cloud, or Microsoft. You cannot opt out. Anthropic is careful to note the mark indicates content was "processed by Claude," not authored—it confirms the model touched something, not who wrote it or how much was AI-generated.
OpenAI, Google, and Meta signed the same EU Code of Practice. xAI notably did not. For teams that need private, unmarked output, the case for self-hosted open models just became significantly stronger.