The U.S. military's current approach, the "Third Offset" which is still in development, is based on the theory that AI is good at defined tasks (e.g., identifying enemy planes) but not at handling the chaos of war (e.g., everything involved in flying a plane on a mission); that is, it's not good at general intelligence. AI can park your car, but the other cars aren't intelligent adversaries who are trying to stop you from parking, destroy your car, and kill you.

The general design is to combine AIs and humans in what they call "centaurs", utilizing the strengths of each in a team. At least, that is what is said publicly.

But autonomy already is deployed. For example, in ships, the last line of defense against incoming missiles are basically large, autonomous machine guns; a human could never act quickly enough against supersonic incoming missiles. Computer system attack and defense ("cyber") is expected to be run by AI, simply because humans can't keep up with an attacking AI. Autonomy also is necessary because you can't depend on having secure, effective communication with your drones; that would be a huge vulnerability.

But autonomy is much trickier to define than it first appears: If you shoot a 'dumb' artillery shell at a target, you lose control of the shell the moment after you pull the trigger. If a group of children subsequently enters the target area, that 'autonomous' shell is going to kill them and you can't stop it. (An AI weapon might be safer in that kind of circumstance; it can change course.) Is it different to shoot a dumb shell to kill everything in the target area than to send an AI-operated drone to the same target area with the same instructions?