As much as the US military relies on drones to bolster its aerial arsenal, it has still relied on human operators to guide its aircraft — until now. Air Force Assistant Secretary Dr. Will Roper has revealed to Popular Mechanics that AI controlled a US military aircraft (and really, military system) for the first time, serving as the “co-pilot” aboard a U-2 Dragon Lady spy plane during a California flight on December 15th.
ARTUμ, a variant of the μZero AI used for games like chess, had ultimate control over radar during a simulated missile strike on Beale Air Force Base. It determined when to focus on hunting for missiles and when to focus on self-preservation. Effectively, that made it the mission leader — it wasn’t flying, but it determined where the human pilot should fly. There was no override, either.
The U-2 FedLab team spent a month training μZero’s algorithms to work the radar, teaching the AI to spot enemies and watch for danger while interacting with a pilot. The polished ARTUμ was cut off from other subsystems to minimize the risks involved with its decisions.
You won’t see AI take over combat duties for a while, even in tests. An AI-guided dogfighting drone isn’t scheduled to fly until July 2021, and it’s not hard for these systems to be fooled by human ingenuity. As it stands, the Defense Department has been developing AI ethics guidelines that stop short of allowing computers to launch attacks on their own.
Full control isn’t needed for the AI to prove useful, however. Roper saw the AI as helping pilots who might otherwise be overwhelmed by “complex” situations — the technology can assist or even take over some systems. The military could have an edge in battle simply by letting pilots focus on the most pressing issues during a mission.