The stereotypical danger isn't that they'll make fewer mistakes than a human, it's that the mistakes they do make will be much bigger--continually offloading decisionmaking and decision enforcement to AI without a coherent theory of friendliness is like using Martingale betting: You don't lose often, but it only takes one to wipe you out.
True. I would imagine that the way to use AI in the context of managing information overload on the battlefield is to have AIs prioritise information and propose options and have humans review the options and take decisions. By having the AI deal with incoming information and humans making the decisions you get the best of both worlds, at least with current technolody. Taking humans out of the loop entirely won't be a good idea for a long time yet.
Keeping humans in the loop doesn't do as much as we think, because humans have a natural inclination to trust computers. In a high-stress situation, such as a battlefield, a human will defer to a computer even if the computer reports information that, under normal circumstances, would be obviously false. In 1988, for example, the USS Vincennes shot down a civilian airliner because its AI targeting system had misidentified it as an enemy aircraft. Giving individual humans access to an off-switch is a pretty suboptimal way to prevent AIpocalypse.
