Every Claude Response Now Carries an Invisible Watermark

Anthropic has begun embedding an invisible watermark into text generated by every Claude model released on or after 2 August 2026. The mark is woven into how the model chooses each token, so it survives copying, pasting and light editing far better than a metadata tag would. It is rolling out across claude.ai, the API, Claude Code, Cowork and Anthropic's cloud partners including AWS, GCP and Azure.

The move sits inside a bigger shift. Files that Claude generates now use the C2PA content credentials standard, the same provenance system Adobe, Microsoft and camera makers have been building into images and video. Anthropic has said the change is partly a response to the EU AI Act's Transparency Code, which asks providers to make AI generated content identifiable.

For studios like ours, the practical question is whether this changes anything about how we brief and use AI tools day to day. Mostly, it doesn't. A statistical watermark on text is a much softer signal than a hash on an image file. Paraphrase a paragraph, translate it and translate it back, or simply retype it by hand, and the mark is gone. It is a useful forensic tool for platforms checking bulk content at scale, not a promise that any given sentence in front of you was or wasn't touched by a model.

What it does do is put a marker in the ground for how the industry is likely to handle provenance for the next few years: quietly, at the token level, and increasingly tied to regulation rather than choice. Clients who ask us whether copy was AI generated are usually asking the wrong question anyway. The better one is whether it's accurate, on brand and says what it needs to say. A watermark answers a different question to the one that actually matters for the finished piece of work.