If you ship a feature built on Claude, its output now carries a mark you did not add. That changes what you should tell your users. It does not give you a detection tool.
What changed
Since 2 August, Anthropic has been watermarking Claude output globally, outside the EU too. Text gets an invisible statistical pattern that survives copy and paste. Generated .png, .jpg and .svg files get signed provenance metadata, which is closer to EXIF than to a visible badge. Models launched on or after that date support it at launch, and older ones are being retrofitted.
Why it happened
The driver is Article 50 of the EU AI Act, which came into force the same day. Penalties go up to 15 million euro or 3 percent of global revenue.
Someone else made a product decision for you
If your app generates copy for customers, that copy is machine-attributable by default. You should probably tell them.
A missing watermark proves nothing
Anthropic says plainly that the metadata can be stripped by format conversion, editing or a screenshot. So if you don't find a watermark, you've learned nothing about where the content came from. Read the actual terms before someone in your company builds a detection feature on top of this.
Building something like this?
I'm Ahmed Mamdouh, a senior full-stack and AI engineer. I reply within one working day.
The systems I am proud of are built from boring parts
The systems I'm proudest of use Postgres, a queue and a cache with a written invalidation rule. The real engineering was in the failure handling.
MCP went stateless, and that matters
The July MCP revision went stateless with OAuth 2.0 and OpenID Connect, so MCP servers can now run behind load balancers and serverless.
What actually breaks when AI agents go to production
Production agents fail on state, partial failure and knowing when to stop. Those are distributed systems problems, and prompts won't fix them.