I think the difference between a targeting a specific piece of military hardware compared to training an AI model to target humans and infrastructure is quite different. This explains why drones that get misdirected will target oil infrastructure in friendly countries.
Agreed. Even some of the latest IR missiles (AIM-9X I believe) also include a visual seeking component to compliment the IR seeker, and try to identify aircraft types based on their outlines (presumably for orienting the missile for maximum damage).
You just can't make that distinction with people, especially not if just using IR or the likes. The guy with a rifle slung over his shoulder just happens to look like the guy with carrying a rake. Hand gun in hand happens to look the same as a power drill. Someone wearing a beanie looks suspiciously like a soldier with a helmet.
This all feels like a really bad idea.