Anthropic has begun embedding invisible, machine-readable watermarks in the text its Claude models generate, a change the company says is designed to satisfy the European Union's new AI Act transparency rules — and one it is applying to Claude's output worldwide, not just for users inside the EU.
The requirement traces back to the EU AI Act's Transparency Code, which took effect Aug. 2 and requires companies to mark AI-generated or AI-edited content in ways other systems can detect. Regulators have said they expect a layered approach combining signed metadata, imperceptible watermarking and, in some cases, logging. Anthropic's version inserts a pattern directly into generated text that is invisible to a human reader but detectable by machine, alongside signed C2PA provenance metadata attached to generated image files such as SVG, PNG and JPG.
Applied at the model, not the app
Claude models released on or after Aug. 2 support the marking by default, and Anthropic says the coverage extends across every surface the model powers — its consumer app, its API, and its distribution through AWS, Google Cloud and Microsoft. "Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from," the company said in describing the rollout, as TechCrunch reported.
Anthropic is not alone in the commitment. Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have all agreed to follow the same EU transparency code, which carries penalties of up to €15 million or 3% of a company's global annual turnover, whichever is higher, for confirmed violations.
The watermark has real limits. It can be stripped by format conversion, re-saving a file, taking a screenshot, or other processing that alters the underlying text or image data. And a detected watermark only proves that content passed through Claude at some point — not that Claude authored all of it, since plenty of the tool's use is proofreading or lightly editing material a person wrote first. Anthropic has described the marking as one layer of a broader transparency effort rather than a complete answer to the question of what is or isn't AI-generated.