Today Anthropic announced that new models for the EU will mark content generated by Claude.

To Americans, this comes out of left field. I find it fascinating that Dario was adamant to push back against the US government earlier this year, but gives into the EU without much fanfare. I’m sure activist groups on this side of the pond will pick up on this headline and push for similar regulations here.

Anthropic says that all text and graphics generated by Claude will include these watermarks, but doesn’t call out code one way or another. That will be a very slippery slope, because they could be laying the groundwork for accepting liability when the LLM’s code wreaks havoc down the line.

To me, liability is going to be the biggest roadblock to widespread agentic AI. Our world depends on having wringable necks to take responsibility when things go awry. When you remove humans from the loop, you also remove the easy placement of blame.

I doubt that Anthropic counsel and shareholders will allow the company to hold the bag, and this move to attach provenance to Claude’s output gets them one step closer to some very uncomfortable situations.