Anthropic will watermark every Claude output globally — but it won't prove a human didn't write it

Starting August 2026, every piece of text and every image generated by a new Claude model will carry a label. Anthropic committed to the EU AI Act’s Code of Practice and is taking the requirement worldwide — across the API, the Claude app, and tools like Claude Code. The move makes Anthropic the first major AI lab to apply a mandatory, universal watermarking system to its models.

Here’s how it works. Text gets an invisible watermark embedded at the model level. It survives copy-paste and, in Anthropic’s phrasing, “may persist through some editing.” Generated images and files get signed provenance metadata using the C2PA standard, which can flag later tampering. The company plans to release verification tools for third parties, though there’s no launch date yet. Cloud partners like AWS and Google Cloud will support the text watermarks, but signed metadata may not carry through those platforms.

Anthropic is refreshingly blunt about what the system can’t do. A watermark doesn’t mean Claude authored the ideas — someone might have used it to polish or translate their own writing. And a missing watermark doesn’t clear human authorship either. Heavily edited text, short passages, format conversions, or screenshots can all strip the signal. The company acknowledges the detection will be more reliable than black-box tools like Pangram, but it’s no silver bullet.

The timing matters because AI detection has become a minefield. Google DeepMind already open-sourced its SynthID watermark for Gemini models, though it struggles with edited text. OpenAI, meanwhile, has been sitting on a text detector it claims is 99.9 percent accurate for roughly two years — and refuses to ship it, citing fears of stigmatization, circumvention via translation, and potential blowback to its own business. Anthropic’s decision to move forward anyway puts it in a different camp. The real question is whether these watermarks hold up under the editing workflows that make Claude useful in the first place. If a few rounds of revision scrub the signal clean, the whole exercise becomes a compliance checkbox rather than a genuine transparency tool.