I tested some scenarios myself and honestly the results were pretty good (and unsetlling at the same time)
Also food for thought: we say human is in the loop for every AI decision. but in a military raid where you need to act in milliseconds, how do you even judge if what the AI told you was right? You just… trust it?
I don't think we have a good answer to this. And that's the scariest part.
https://nanonets.com/blog/anthropic-pentagon-ai-control-problem/
3 Comments
If you’ve seen how developers are treating Claude code, then you’d realize human in the loop quickly devolves into spamming the accept button.
I believ we’re building these systems faster than we’re building the frameworks to govern them. And the human in the loop promise is becoming a fiction because the tempo of modern operations makes meaningful human judgment physically impossible.
Eg. if AI is synthesizing intel from 50 sources and surfacing a go/no-go recommendation in real time, and you have seconds to act, what does “oversight” even mean anymore? You’re not evaluating the decision. You’re just the hand that pulls the trigger on a decision the AI already made.
and this is only going to get worse. As these systems get faster and more autonomous, the window for human judgment shrinks further. We’ll keep calling it “human in the loop” because it sounds safer, but the loop will get so tight it’s basically a point.
So, do we need a hard international framework that defines *minimum* human deliberation time before AI-assisted lethal decisions? And if yes – who enforces it when every major military is racing to be faster than the other?
Because right now, nobody’s slowing down, lol
bruh. The “I tested it myself” section genuinely made me chortle. Did you really just query the consumer product with hypothetical military scenarios and screenshot the results? “here’s what the evidence suggests” BRUH WHAT EVIDENCE. There is 0 useful information in this article. I swear, LLM’s are to blog posts what cheap recording equipment is to podcasts. Not everyone should do them. Unless this is just SEO spam for whatever the fuck nanonets is.