Anthropic said on August 11 that Claude will mark the content it produces. Text carries a watermark embedded in the writing itself, so it travels with the text when someone copies and pastes it elsewhere. Files carry signed provenance metadata using C2PA, the open content provenance standard. Models released after August 2 include it, and the company says it plans to extend support to older ones.
##Why was this added?
Article 50 of the EU AI Act became enforceable on August 2. It requires providers of generative systems to mark synthetic audio, image, video and text in a machine-readable format that is detectable as artificially generated or manipulated. Anthropic has been direct that this is about meeting that obligation, and it is the third major model provider to describe how it intends to comply.
One distinction is worth getting right, because most coverage flattens it. The provider obligation is machine-readable marking, not a visible watermark. Nothing in Article 50(2) requires that a person can see a label. There is a separate duty on deployers, meaning the people publishing the output, to clearly label AI-generated text put out to inform the public on matters of public interest, and the Commission has been explicit that deployers cannot satisfy that duty by pointing at the invisible mark. Two obligations, two different parties, often reported as one thing.
So this is not really a Claude story. The requirement lands on anyone selling a generative system into Europe, and Anthropic is applying it worldwide rather than only in the EU, which is usually what happens once a company has built the plumbing anyway.
For imaging it sits closer to home than it first appears. AI editing is already a normal part of photography, and much of it does not generate a picture from nothing. It removes an object, extends a frame, replaces a sky. The legislation covers manipulated content and not only fully synthetic content, which is a considerably wider net than most photo tools are currently designed around. That reaches Adobe, camera and software vendors, generative startups and independent developers alike.
Which brings it back to us. MagicTrick modifies images substantially, and what we build after it will too. This is not a problem we have suddenly acquired. It is a design decision that has arrived early enough to treat provenance, disclosure and metadata as part of the product architecture rather than as something added under deadline pressure later. A system that carries its own history is easier to build than one retrofitted to explain itself after the fact.
The watermark is not the signal. The signal is that telling people what a machine made is turning from a feature into an expectation.
Sources
- [TechCrunch: Anthropic says it will watermark text generated by its AI models](https://techcrunch.com/2026/08/11/anthropic-says-it-will-watermark-text-generated-by-its-ai-models/)
- [European Commission: Transparency obligations under Article 50 of the AI Act](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)
- [EU AI Act: Article 50, transparency obligations for providers and deployers](https://artificialintelligenceact.eu/article/50/)