AI Pulse · Frontier Labs Watch

Anthropic just turned on watermarking for everything Claude produces. Text gets an invisible marker that survives copy and paste. Images and files get signed C2PA metadata. It’s live worldwide, not a pilot or opt-in. They’re calling it a compliance move for the EU AI Act, but it’s also the first time a major lab has committed to always-on provenance at this scale.

The watermark doesn’t break if you edit the text or move it between apps. That matters because earlier attempts at text watermarking fell apart the moment someone rephrased a sentence. C2PA on images is more established, it’s the same standard Adobe and others use, but Anthropic is now applying it to everything Claude generates by default. No toggle to turn it off.

This sets a new baseline. OpenAI and Google have talked about provenance tools, but neither has shipped a mandatory, global system. Now they’ll be measured against this. If you’re using AI to draft contracts, reports, or marketing copy, you’ll start seeing these markers in your workflow. Some tools will surface them, others won’t, but the metadata is there. That could matter for audits, compliance checks, or just proving where a piece of content came from.

If you’re building AI into your operations, this is the kind of shift that changes how you think about output integrity. It’s not about blocking AI use, it’s about knowing what came from where. The Omni Command Centre tracks these kinds of provenance signals across models, so you’re not guessing which drafts carry metadata and which don’t. You just see it, log it, and move on. Anthropic moved first. The rest will follow, and your stack needs to be ready when they do.