Why DARPA put AI at the controls of a fighter jet

In December, a special fighter jet made multiple flights out of Edwards Air Force Base in California. The orange, white, and blue aircraft, which is based on an F-16, seats two people. A fighter jet taking to the skies with a human or two on board is not remarkable, but what is indeed remarkable about those December flights is that for periods of time, artificial intelligence flew the jet. 

As the exploits of generative AI like ChatGPT grip the public consciousness, artificial intelligence has also quietly slipped into the military cockpit—at least in these December tests.  

The excursions were part of a DARPA program called ACE, which stands for Air Combat Evolution. The AI algorithms came from different sources, including a company called Shield AI as well as the Johns Hopkins Applied Physics Laboratory. Broadly speaking, the tests represent the Pentagon exploring just how effective AI can be at carrying out tasks in planes typically done by people, such as dogfighting. 

“In total, ACE algorithms were flown on several flights with each sortie lasting approximately an hour and a half,” Lt. Col. Ryan Hefron, the DARPA program manager for ACE, notes to PopSci via email. “In addition to each performer team controlling the aircraft during dogfighting scenarios, portions of each sortie were dedicated to system checkout.”

The flights didn’t come out of nowhere. In August of 2020, DARPA put artificial intelligence algorithms through their paces in an event called the AlphaDogfight Trials. That competition didn’t involve any actual aircraft flying through the skies, but it did conclude with an AI agent defeating a human flying a digital F-16. The late 2022 flights show that software agents that can make decisions and dogfight have been given a chance to actually fly a real fighter jet. “This is the first time that AI has controlled a fighter jet performing within visual range (WVR) maneuvering,” Hefron notes.

[Related: I flew in an F-16 with the Air Force and oh boy did it go poorly]

So how did it go? “We didn’t run into any major issues but did encounter some differences compared to simulation-based results, which is to be expected when transitioning from virtual to live,” Hefron said in a DARPA press release. 

Andrew Metrick, a fellow in the defense program at the Center for New American Security, says that he is “often quite skeptical of the applications of AI in the military domain,” with that skepticism focused on just how much practical use these systems will have. But in this case—an artificial intelligence algorithm in the cockpit—he says he’s more of a believer. “This is one of those areas where I think there’s actually a lot of promise for AI systems,” he says. 

The December flights represent “a pretty big step,” he adds. “Getting these things integrated into a piece of flying hardware is non-trivial. It’s one thing to do it in a synthetic environment—it’s another thing to do it on real hardware.” 

Not all of the flights were part of the DARPA program. All told, the Department of Defense says that a dozen sorties took place, with some of them run by DARPA and others run by a program out of the Air Force Research Laboratory (AFRL). The DOD notes that the DARPA tests were focused more on close aerial combat, while the other tests from AFRL involved situations in which the AI was competing against “a simulated adversary” in a “beyond-vision-range” scenario. In other words, the two programs were exploring how the AI did in different types of aerial contests or situations. 

Breaking Defense reported earlier this year that the flights kicked off December 9. The jet flown by the AI is based on an F-16D, and is called VISTA; it has space for two people. “The front seat pilot conducted the test points,” Hefron explains via email, “while the backseater acted as a safety pilot who maintained broader situational awareness to ensure the safety of the aircraft and crew.”

One of the algorithms that flew the jet came from a company called Shield AI. In the AlphaDogfight trials of 2020, the leading AI agent was made by Heron Systems, which Shield AI acquired in 2021. Shield’s CEO, Ryan Tseng, is bullish on the promise of AI to outshine humans in the cockpit. “I do not believe that there’s an air combat mission where AI pilots should not be decisively better than their human counterparts, for much of the mission profile,” he says. That said, he notes that “I believe the best teams will be a combination of AI and people.” 

One such future for teaming between a person and AI could involve AI-powered fighter-jet-like drones such as the Ghost Bat working with a crewed aircraft like an F-35, for example. 

It’s still early days for the technology. Metrick, of the Center for New American Security, wonders how the AI agent would be able to handle a situation in which the jet does not respond as expected, like if the aircraft stalls or experiences some other type of glitch. “Can the AI recover from that?” he wonders. A human may be able to handle “an edge case” like that more easily than software.