I don't understand the line as well. So its no to domestic surveillance, but all other countries are a fair game? How is this an ethical stand? What sort of mental gymnastics allow Anthropic to classify this as an ethical stance?

To me all of this reads like "we don't trust our models enough yet to not cause domestic havoc, all other is fine, and we don't trust our models enough yet to not vibe-kill people". Key word being "yet".

"vibe-kill" made me laugh then feel sick

The whole reason this is happening is because Anthropic looked into how Claude was used in the Maduro op and found it to violate the negotiated terms of service.

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

So... this would be fine with them?

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."

First one

Seems like an absurd distinction to me... Reminds me of "I was just following orders"...

I mean the distinction doesn't really matter

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

I agree the distinction doesn't matter, but im not so sure "just" having a human in the loop qualifies as an ethical stand. Just because your not pulling the trigger doesn't make you not culpible for the outcome.