What Anthropic announced, exactly
Starting August 14, 2026, new Claude models mark the text they generate with a technique called SynthID-Text. Models from before August 2, 2026 will get it in phases over the coming months. The main reason is regulatory: the European Union's AI law requires marking AI-generated material.
How it works, in plain words
When Claude writes, it constantly chooses between words that work equally well. The watermark uses those low-risk choices: instead of picking at random, it uses a cryptographic key plus the previous words to tilt the choice. The result is a pattern no reader can perceive, but that can be verified with the right key. Anthropic will publish a detection API.
Files and images take a different route: they get C2PA credentials in the metadata, not a watermark in the content.
What it does NOT do (and this is what matters most)
- It doesn't identify you. The mark contains no user or organization information, and it can't be traced to a specific person or chat.
- Not just anyone can detect it. Verifying it requires the key and Anthropic's API. Google can't "see" the mark on your blog, and neither can your client.
- It isn't foolproof. Anthropic itself lists its limits: it works poorly on short texts, on factual passages, on code that requires exact output, and it weakens with heavy editing.
- It doesn't tell how much you did. It indicates that Claude took part, not whether it wrote everything or only polished your draft.
What changes for an agency that produces content with AI
- Cost and quality: nothing. No extra tokens, no new price, no measurable degradation.
- Client content: honesty by design. If a client asks whether their blog was made with AI, the technical answer can now exist. The right move doesn't change: AI drafts, a human checks the facts and signs off. Seriously edited text also falls outside the reliable reach of the mark.
- Sensitive deliverables: review your contract. If you promised "100% human" content in any agreement, that wording was already risky and is now verifiable. Better to promise content that is verified and edited by humans.
- SEO: no panic. Google has said for years that it evaluates how useful content is, not how it was produced, and it has no access to the detection key. Weak content will still be the problem, with or without a mark.
What to do this week
- Nothing urgent: the change is transparent to your current workflow.
- If you use Claude through the API for your own products, note Anthropic's detection API for when it ships: it can help you audit your own pipeline.
- Review how you describe your content service in proposals and contracts: "AI-assisted and human-verified" is defensible; "no AI" no longer is.