Drones... may well make mistakes - I'm not sure that AI Vision is anyway 100% foolproof, but besides - Drone operators are human, and i'm under the impression they most certainly are fallible.
They're certainly fallible - but they can err on the side of civilian life rather than the drone's. That's a big improvement even if the errors are more common.
The fact that the drone doesn't require safety effectively means it shouldn't be armed to comply with typical use of force guidelines. Drones essentially are no different than the crime cameras in the UK, only they're mobile.
Reading the article, it's mostly talking about remote control, not AI