A watermark on AI text sounds like a signature. Anthropic's own description is narrower. Future Claude models will generate text that contains a watermark so someone with the right key can estimate whether Claude was involved in writing it. That is a likelihood, not a byline.
What Anthropic has published
On August 14, 2026, Anthropic explained the method on its news page. The company says it is using a version of Google DeepMind's SynthID-Text approach. The watermark sits in low-stakes word choices. Readers are not meant to see it. Anthropic says it does not add hidden characters, extra tokens, or a practical cost increase, and that internal testing showed no impact on quality or readability.
The company is doing this to comply with the EU AI Act. Anthropic says it signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, along with other major providers and around 190 signatories in total. It is applying the mark worldwide at launch because it does not yet have a durable way to scope the change by region.
A separate help-centre note says models launched on or after August 2, 2026 support machine-readable marking at launch, including on the API, Claude apps, Claude Code, Claude Cowork, and Claude Tag, and when supported models are reached through AWS, Google Cloud, or Microsoft Foundry. Older models are in a transition period. Anthropic says a detection API is coming and has not published the finished details.
What the mark does not prove
Anthropic is explicit about the limits. A detected mark can suggest Claude was involved. It cannot tell you Claude was the original author. It cannot tell you a different model wrote the text. Short passages, heavy edits, grammar-only proofreading, and exact outputs such as code all leave less room for a reliable signal. A full rewrite can remove it. Files such as PNG or SVG get a different treatment: signed C2PA provenance metadata in the file, not a hidden pattern in the pixels.
That matters for anyone treating a detector as a staff policy. A watermark is not a confession, a copyright stamp, or a privacy leak. Anthropic says it does not identify the user.
A practical test for Canadian teams
If Claude is already in your shop, assume new-model output may carry a mark that travels when someone pastes the text into a proposal, a tender reply, or a client memo. Write the policy before the first awkward screenshot. Decide which work may leave on a public model, which work stays on a private AI path, and who approves the final document. Do not wait for the detection API to invent a process. The process is the same as any other outbound draft: a person owns the claims.
If you sell or file work in Europe, or you sit in a Canadian firm that does, treat this as a transparency rule landing in the tools you already pay for. Anthropic also tells builders who deploy Claude in their own products to assess Article 50 themselves. That is the AI consulting question in plain language: what does your stack disclose, and what does it keep?