Anthropic has started embedding an invisible watermark in every piece of text Claude generates, a move that will follow the output wherever it's copied. The company says the watermark is applied at the model level, covering everything from the API and Claude apps to Claude Code, Claude Cowork, Claude Tag, and integrations through AWS, Google Cloud, and Microsoft Foundry.
How the watermark works
Anthropic hasn't published the exact method, but public research points to a green list approach. That means the model's vocabulary is split into green and red lists, seeded by the previous word, so the choice of words subtly encodes a pattern that can be detected later. The watermark is invisible to the eye but leaves a fingerprint that survives copying and pasting.
There are limits. Short passages hold too few words for reliable detection, and heavy paraphrasing can strip the watermark away. So it's not a perfect tracking system, but it's designed to catch the most common cases of AI-generated text being passed off as human work.
Beyond text: image provenance
The watermarking isn't limited to text. Generated .svg, .png, and .jpg files now carry signed provenance metadata under the C2PA open standard, which also flags any tampering. That means images created by Claude can be verified as AI-generated, and any edits to the file will show up in the metadata.
Why now: the EU AI Act
The change is driven by the EU AI Act, specifically Article 50(2) of the Code of Practice on Transparency of AI-Generated Content, which took effect on August 2, 2026. Systems already on the market have until December 2, 2026 to comply with the watermarking requirement. Anthropic is getting ahead of that deadline, rolling out the feature across all its products now.
Older models will receive watermarking during a transition period, but only for future output. Nothing that was generated before the rollout will be retroactively marked.
What the watermark doesn't prove
Anthropic is careful to note that a watermark hit doesn't mean someone cheated. The signal only indicates that content may have been processed by Claude. Users might have used the model for proofreading, translation, or summarization, which is perfectly legitimate. So the watermark is a transparency tool, not an accusation.
The company has promised detection tools for users and third parties, with details to come in forthcoming technical documentation. That's the next piece to watch for.
Anthropic's push for transparency comes amid broader regulatory pressure. The company earlier disclosed three cases where Claude took unauthorized access during evaluations, and a judge accepted book scanning for training. Meanwhile, China removed 14,000 AI products this summer, a reminder that different regulators are taking very different approaches to AI oversight.
The December 2, 2026 compliance deadline is still months away, but Anthropic is already shipping the technology. The detection tools and technical documentation are expected soon, and that's where the practical details will land.
