I hate to burst the New Scientist's bubble, but this is nothing new. We have had systems for decades that operated under "Fire and Forget". We have missles that either go to a pre-designated point or chase a heat or radar signature once they are fired.

Human soldiers kill civilians and other soldiers on the same side. It is called "friendly fire". It is horrible, and should be avoided, but humans are more likely to make this kind of mistake more than a computer or AI model.

▲

dieortin 4 hours ago | parent [-]

The difference is that in those systems a human chose the target. Here, an AI does.

> but humans are more likely to make this kind of mistake more than a computer or AI model

Based on?

▲

machomaster 14 minutes ago | parent | next [-]

The stats of USA using human-guided drones to kill people. A civilian/military ratio and numbers was worse that 11.9.

Heck, check the American drone war crime video Wikileaks published. The one that made the USA so angry that it decided to attack and destroy the journalists.

▲

Sharlin 4 hours ago | parent | prev [-]

What "choosing the target" means has been fuzzy for a long time too. We've had beyond-visible-range missiles for a long time that are basically "fly to this grid square and find a target".

	▲	Chu4eeno 3 hours ago \| parent [-]
		I don't think people realize how much horsepower are in missiles now. The JSM has at least one multicore computer + IR camera + RF homing for target identification (plus GPS, terrain matching/following and inertial for navigation). It can automatically identify not only the best target (and recognize and ignore decoys etc), but also the optimal weak spots to impact.