The US Army wants to . It may allow faster responses to threats, but it could also be a step towards autonomous drones that attack targets without human oversight.
The project will use machine-learning algorithms, such as neural networks, to equip drones as small as consumer quadcopters with artificial intelligence. Current military drones have little onboard intelligence, sending raw video back to analysts who pick out and identify targets.
At the moment, you can have dozens of people monitoring the video feed from military drones, who then decide what action to takesays at the Center for a New American Security, a think tank in Washington DC.
Advertisement
The US Army already fields miniature drones that can or pursue a target autonomously once the operator locks on to it with the camera. Several thousand have been used in Afghanistan, Iraq and Syria. But the new project goes much further. It will detect, recognise, classify, identify and target, which covers the entire process from finding a person to aiming weapons at them.
Human overseers?
鈥淭his does sound like they are moving very close to lethal autonomous weapon systems,鈥� says at the University of California, Berkeley, who has .
The Pentagon鈥檚 of drones. But Russell doesn鈥檛 find this very reassuring. 鈥淎t best, the drone shows a target to the human and the human says 鈥榶es鈥� or 鈥榥o鈥�, which is known to be a problematic system design, because the human stops exercising judgement very quickly,鈥� says Russell.
Many studies of safety systems, including 聽show that operators tend to gradually reduce the number of times they check a machine鈥檚 output and start to place more trust in the machine than in their own senses, a tendency known as automation bias. 鈥淎nd of course, it would be very easy to leave the human out,鈥� says Russell.
Dumb mistakes
Scharre, formerly a special operations reconnaissance team leader in the US army, doubts that the US military wants autonomous weapons. He does see other issues, though.
鈥淭he problem with AI is that it鈥檚 brittle,鈥� says Scharre. 鈥淚t can go from super-smart to super-dumb in an instant, making mistakes that are jarringly stupid for a human.鈥� One famous example is a 3D printed plastic turtle that AI systems repeatedly identify as a gun, even though there is no resemblance in human eyes.
Small drones that can spot and identify objects already exist. The University of Technology Sydney (UTS) in Australia has produced the that runs on a drone and can distinguish sharks from swimmers, surf boards and dolphins. Nabin Sharma of UTS says that SharkSpotter鈥檚 accuracy is about 90 per cent, compared with 30 per cent for humans analysing aerial imagery.
鈥淭he military is basically trying to import the technology from the commercial sector,鈥� says Scharre. 鈥淭he tech is out there, and it will be widely available whatever the US Army does.鈥�
A US Army spokesperson would not comment on the project at this time.
