They hypothesize reducing the misidentification rate will increase the usage rate, ultimately killing more people (but fewer wedding parties)
This is speculation at best, outright wrong at worst. Prove they're better than humans before employing them.
One soldier made this mistake. What if a faulty image fitting/target classification algorithm deployed in a fleet of autonomous drones was vulnerable to a similar misclassification and started flagging some innocuous objects in people's hands as weapons? Would the rules of engagement be permissive enough for the drones to begin killing those people on their own, would there be human operators in the loop confirming and authorizing the strike?
Any failure of the technology will be scoped away. If the drone blows up a city block targeting a kid with an ice cream cone, the cone will be classified as a potential weapon, therefore tragically the child was a militant.