Anthropic announced that all Claude models launched on or after August 2, 2026, will embed imperceptible watermarks directly into generated text as part of compliance with Article 50 of the EU AI Act, which took effect on August 2, 2026. The watermark is embedded at the model level using a statistical technique based on Google DeepMind's SynthID-Text approach: during token generation, when Claude has multiple reasonable next-word choices, the watermarking system uses a secret key to bias selection in a way that creates a detectable statistical pattern across the text without changing meaning, readability, or word choice. Additionally, Claude will attach cryptographically signed C2PA provenance metadata to generated files. Anthropic says the watermark carries no identifying information about the user and survives copying/pasting and minor edits.
The marking applies globally, not just in the EU, because Anthropic states it lacks a durable way to scope watermarking by region. The company also committed to adding watermarking to older Claude models within a transition period (through December 2). The watermark will be present everywhere Claude is offered: the Claude app, Claude Code, Claude Platform (API), AWS, Google Cloud, and Microsoft Foundry. Anthropic has not yet released the detection API; it will share technical details in forthcoming documentation. The watermark adds no extra tokens and has negligible performance overhead. Major competitors have taken different approaches: OpenAI has long had text-watermarking technology but has chosen not to deploy it due to concerns about false positives and user flight; Google has watermarked images since 2023 and expanded to text and audio.
For developers and architects, this marks a regulatory inflection point where EU transparency rules are reshaping global AI product defaults. Enterprises using Claude for summary, translation, or proofreading of human-written content will now face a detection problem: even if Claude was used only to refine human work, the output may carry a Claude watermark, forcing teams to disclose AI involvement in contexts where they might not disclose it otherwise. Third-party AI-detection tools (which lack Anthropic's key) cannot verify watermarks; they rely on linguistic pattern analysis instead. Fine-tuning, rephrasing, or heavy editing can weaken the watermark. Fines for non-compliance with Article 50 range up to €15 million or 3% of global annual turnover.