>if Russian or Chinese weapons are automatically firing at the speed of a neural network while our weapons are waiting for meatware actuation then it's going to

>if Russian or Chinese weapons are automatically firing at the speed of a neural network while our weapons are waiting for meatware actuation then it's going to be a short battle. While I see your point on the ever present escalation risk in these things, are you sure you're not taking an overly restrictive view of "pull the trigger"? Take long existing weapons like just plain old guided missiles. In one respect it could be argued a computer system is making the kill choice there, after firing there will be a long period where the computer is seeking the target and the computer is what actually detonates the warhead that makes the kill. But in practice it's hard to argue that a human isn't still "pulling the trigger", making the decision that "this target should be killed" and then the level of intelligence the weapon employs to accomplish that order afterwards isn't so relevant. The real comfort factor seems come from there being inherent limits in terms of time/space (a missile only has so much run time and range) and specificity (a missile is going after a specific designated target). If we consider the most problematic weapons of modern history, I think the real uniting factor is when they get away from all those limits at once. Landmines can linger a long time and activate indiscriminately without further human intervention. Chemical and bioweapons also not well controlled in terms of time, space or specificity despite no AI. Project Pluto, the early Cold War R&D by the US military to create a hypersonic nuclear scramjet cruise "missile" (more of a autonomous bomber) that could be launched then linger flying over the ocean for months (due to nuclear reactor power) was abandoned precisely because it was considered far too destabilizing even for then. Despite still being a "missile", the longer time/space/area effect made a difference. Heck, even cluster munitions on conventional missiles, while not considered in the same category as landmines exactly, are still pretty hotly debated. By the same token in a future conflict with more advanced NN/ML weapons, isn't there still room for operator decisions? Ie., "here are cryptographically authenticated orders to kill targets matching X profile within Y geographic area for Z time period". It would seem like setting international restrictions on X, Y and Z there (requirements for safety interlocks and for any given kill authorization to be valid for no more then 20 minutes over no more then 400 km^2 to toss something out) could be a viable approach. So in the specific minutes and seconds and milliseconds of machine/machine conflict there'd be no disadvantage, but humans would still be need to be constantly "involved" at a higher level. Having killbots just roaming around or lying in wait (and potentially lasting after the conflict is totally over, ala landmines) would be banned, but sending robot tanks to attack a specific fortified position on one single day with active strategic direction would not be. Basically moving humanity "up the chain" a bit seems like something countries could practically agree to without feeling massively behind and could be monitored/verified. Edit: updated to add some comparison to non-AI weapon systems that have been agreed to be restricted with some success.