Their hard lines are:
- no usage of AI to commit murder WITHOUT a human in the loop
- no usage of AI for domestic mass surveillance
Claude: "Are you sure you want me to commit murder?"
User: "Yes"
Or do you mean Human presses button:
Claude: "Do you to commit murder? If so press the button."
User: "I pressed the button"
Claude: "Great! Now lets summarize what we did."
There are many ways to construct HITL UXes. But typically they'd take the form of the first one
I think you're missing the forest for the trees. All Anthropic is saying is that HITL is required before murder, the UX is irrelevant