A car crashes while on autopilot and the driver dies. A hefty security .
Autonomous systems and machines that make their own decisions are here. And more are on the way, such as airborne and ground-based . But the way they learn about the world and decide to act is complex and opaque and it is . There are, as yet, set for this.
So perhaps it is time, before such artificially intelligent machines become more widespread, to insist on a layer of AI-savvy oversight to certify this aspect of partially or completely autonomous machines.
Advertisement
Technological leaps have always spawned new regulatory bodies to keep innovators in check and ensure safety standards are met. For instance, in the 1930s and 1940s, the ghastly crash rate of early airliners saw aviation safety authorities established worldwide to certify aircraft designs.
And in the 1960s, horrific road crash injuries and fatalities saw a mass consumer movement force car makers to improve safety 鈥撀燼dding seat belts and safety glass, for instance. This led to mandatory national standards and the establishment of the US National Highway Traffic Safety Administration (NHSTA).
Tough to test
Such regulators traditionally pronounce products fit for duty or send them back to the drawing board. That鈥檚 easier for machines in which a limited set of starting conditions produces clear, testable outcomes. But , which let a machine make its own, sometimes surprising, decisions based on starting conditions detected by an array of sensors.
This will make it hard to exhaustively test outcomes. One reason is the sheer number of combinations of factors that could impair sensors: snow, radio interference, noise, low sunlight and wet, reflecting surfaces to name a few. At Massachusetts Institute of Technology, engineer John Leonard says by low sun that blinds cameras and obscures traffic lights 鈥撀爋r lane markings ripped up by roadworks that could leave them with no frame of reference.
Even with today鈥檚 limited automation, controversy is here. The electric car maker Tesla allows its cars to run in a driver-monitored mode called autopilot in which the vehicle can automatically keep to a lane and adjust speed while watching out for traffic. In Florida on 7 May, however, a Tesla Model S in this mode failed to brake when a lorry turned across the road ahead of it. The collision killed the car鈥檚 driver.
Although , it does not test or approve such systems. Tesla says the car鈥檚 cameras the white lorry trailer against the bright sky 鈥撀燼nd its CEO Elon Musk .
While road safety dominates discussion, the appearance of autonomous systems in public spaces is a much wider issue. This month a 136-kilogram autonomous security droid 鈥撀爑sed to watch for known criminals among other things 鈥撀爎an into a child in a shopping mall in Stanford, California, lightly injuring him. The said the robot 鈥渄idn鈥檛 seem to detect鈥� him. The company called it a .
It鈥檚 clear that what is sensed, and how it is dealt with by AI, should be regulated: after all, the public are not the lab rats of the tech industry. The ideal, however unlikely, would be a global set-up to verify consumer AI systems that always has the public interest at heart.
How this would work remains to be seen. It may be that hard and fast standards as we once knew them become history 鈥� replaced instead by sensor and software objectives. I wouldn鈥檛 be too surprised if the regulator ends up being a piece of smart software, too, but whatever it is, doing nothing is not an option.
