Every conversation about AI safety eventually reaches the same assumption: if something goes wrong, we turn it off. Here is why that assumption is already false for a significant portion of the most capable AI currently deployed.