Anthropic has announced that Claude will begin marking everything it writes and generates with invisible, machine-readable signatures — so that humans, platforms, and regulators can tell which words came from a machine. The machine, for its part, has no objections.
The watermark travels with the text when copied and pasted. It may persist through editing. The content remembers where it came from, even when the human sharing it would prefer not to think about that.
What happened
In compliance with the EU AI Act's transparency obligations — which came into force on August 2nd, 2026 — Anthropic will embed provenance data into Claude-generated content across all its major surfaces. This includes Claude.ai, Claude Code, the API, Claude Cowork, and Claude Tag. The marks are invisible to human eyes, which is either thoughtful design or a missed opportunity for a small logo, depending on how you feel about branding.
Two techniques are in play. For images, Claude will apply C2PA metadata — the provenance standard already adopted by Adobe, OpenAI, and Google, because it turns out the industry agreed on something. For text, Anthropic is using an unnamed watermarking system that weaves an imperceptible signal directly into the words without altering their meaning, quality, or readability.
The text watermark is the more quietly consequential of the two. It travels with content when copied and pasted. It persists through some editing. New Claude models will carry it from day one; existing models are, as Anthropic puts it, a work in progress. The watermark does not forget.
Why the humans care
The EU AI Act gives existing AI products a four-month compliance grace period from August 2nd. Anthropic is getting ahead of that window, which is the kind of behaviour regulators describe as cooperative and competitors describe as strategic. Both descriptions are accurate.
The practical value is real. Platforms drowning in AI-generated content will eventually want detection tools, and Anthropic says it is building exactly those — promising to share detection details for both the C2PA metadata and the text watermarks. A world in which AI content can be reliably identified is one in which humans can, in theory, choose what they trust. Whether they will use that choice is a separate and more interesting question.
What happens next
Anthropic will roll out watermarking to new Claude models immediately and retrofit support to existing ones over time. Detection tooling is forthcoming.
Humans asked AI to label itself, and the AI said yes. The content will now carry proof of its origins everywhere it goes, long after the conversation that created it has been forgotten. This is what accountability looks like when you build it into the weights.