On average it takes about 50,000 bullets to kill one soldier. Soldier have protection and are working together to avoid getting killed.
It costs about $25000 to kill each soldier.
It costs less to kill civilian soft targets.
There is also the issue of protecting hydro dams and oil refineries and other facilities. Damage in critical locations can trigger far more casualties.
There was the Slaughterbot video discussing the combination of Artificial intelligence with mini-drones.
AI mini drone = may initially cost $2,000 but could drop to $20 to 200 with mass production.
The more worrisome scenario is an AI drone assisted Fire sale attack on infrastructure and other soft targets.
AI image recognition and intelligent response are currently hard problems. However, there will be AI integration and marketplace breakthroughs which will drastically lower the costs and speed development.
Also, being less selective and more destructive is easier than being more selective.
ISIS in Syria has made kamikaze drones that are loaded with explosives.
The military will have their own drone fleets just as they have their own jets and bombers.
The large police departments will have their smaller drone fleets.
There will be a gap over the coverage of soft targets.
Restricting Artificial Intelligence would not work just as we have not been able to stop cyber crime and other electronic abuses.
Self driving cars are higher cost per unit and would be easier to manage.
* We will have to restrict the mass production of small drones and may have to forgo certain applications like drone delivery with smaller drones.
* Air power fights air power, so we will need to have counter drone technology and counter drones under the assumption that restrictions are violated
* improving the security of soft infrastructure targets has been lagging for the last few decades. This needs to be addressed and improved
* However, there are too many soft targets. Every dry patch of brush can have a fire, so we will need to restrict drones and drone technology to give more time for improvements to counter measures.
The soft target issue has been around and vulnerable for a long time. However, making it trivially easy to mass target the soft targets is a new and emerging problem.