- The US military is moving forward with its development of AI pilots in a recent groundbreaking test.
- Air Force officials said the AI learns from flight and mission data and has made “dramatic” progress.
- Shield AI's CEO said AI is an affordable, mass solution to the problems facing the Air Force and the future battlefield.
Last month, two US Air Force fighter jets engaged in a fierce dogfight over California. Air-to-air combat training is not uncommon, but what made this incident strange was that one of the planes had no human pilot at the controls.
With the Secretary of the Air Force in the back seat and piloted by an artificial intelligence agent, the fighter jet engaged in a dogfight with a manned F-16, demonstrating the rapid progress being made by the Air Force and Defense Advanced Research Projects Agency's AI fighter pilot program.
An Air Force commander and CEO of Shield AI, the company that developed the AI, told Business Insider that the moment marks a groundbreaking step in the program's development.
The Air Force also appreciated this. Col. James Valpiani, head of the Air Force Test Pilot School, told BI: [Frank] Kendall's flight marks a “significant milestone” for the program and speaks to growing “confidence in the state of combat autonomy.”
Kendall flew for about an hour in the X-62A Configurable Flight Simulation Test Aircraft (VISTA), a highly modified Air Force F-16, at Edwards Air Force Base on May 2. “We witnessed several AI agents autonomously piloting the X-62 against manned F-16s in dogfights,” Valpiani said.
Throughout testing of this emerging technology, the Air Force has been impressed with the speed at which the AI can learn and modify new information based on large amounts of data and flight experience.
“This iterative process allows the team to rapidly mature their AI agents, often resulting in dramatic performance improvements with each test flight,” Valpiani says.
Across the Pentagon, defense officials and leadership are fast-tracking affordable, mass-producible AI capabilities, described as a key element of great power competition with China and a kind of arms race to keep the U.S. at the top of its game for any future conflict.
Kendall himself called last month's flight a “transformative moment” that marked the transition from a time when autonomous air-to-air combat was merely a “distant dream” to a reality.
Ryan Tseng, CEO and co-founder of Shield AI, told BI that the US should take a Manhattan Project-like approach to acquire the resources, talent and technology to deploy AI at scale, and that autonomous vehicle pilots are just the first step.
“What we're looking at is [from Kendall’s flight] “We've made impressive progress in the performance, reliability and safety of our AI Pilots,” Tseng said, “and we're making significant efforts and investments on our side to apply them to missions relevant to the security challenges we face around the world.”
He added that the AI Pilot would meet the Department of Defense's call for intelligent, affordable, mass AI, address concerns about the number of pilots and the training required to fly, and remain effective on the challenging modern battlefields where GPS and communications may be degraded or cut off.
Tseng co-founded Shield AI to “provide the Department of Defense and its allies with the best in AI and autonomous technologies to protect military and civilian populations,” according to the company's website.
Work on the AI to pilot the X-62A, in collaboration with the Air Force and the Defense Advanced Research Projects Agency (DARPA), has involved trial and error, massive amounts of data, simulated dogfights (Shield AI has had a 99% win rate in dogfights), and real-world applications.
But amid all this, the AI pilot, named Hivemind, is maturing faster than expected. “I think things are moving a lot faster than people probably realize,” Tseng said.
Lt. Col. Ryan Heffron, DARPA's Air Combat Evolution program manager, expressed a similar sentiment in April when the Air Force announced that the AI would complete a groundbreaking real-world dogfight with a manned aircraft in September 2023. The event marked the first time such a test had been conducted.
Officials have remained quiet about whether the aerial battle was won by the AI or the human pilot, but they have been eager to share progress reports.
“Things are moving as quickly as we expected, if not faster,” Heffron said at the time.
A lot of this can be attributed to how quickly AI agents can adjust based on new data and performance.
Training an AI to fly starts with basic maneuvers and fundamentals, like a human pilot, then new missions, skills and abilities are added to the AI's repertoire. If the AI exhibits “unexpected behavior,” Valpiani said, many changes can be made to the program within hours, or even overnight.
“These changes are often related to the 'simulation-reality gap', a common challenge in machine learning research, where AI agents perform differently in simulators and in real-world applications due to inevitable discrepancies between simulators and real-world environments,” he explained.
When such errors occur, the team would replay the dogfight scenario in a simulator, identify what went wrong, update the scenario, and retrain the AI agent—an iterative process that means the AI system is constantly learning and relearning.
But there are big questions about the role of AI in the US military, and recent events, such as accusations that the Israel Defense Forces used an AI program to target Hamas operatives in the Gaza Strip, indiscriminately endangering civilians, only heighten fears about future wars fought with AI.
The Pentagon has made clear that AI will not make life-or-death decisions, and those decisions will be left to humans, but there are concerns about how that distinction will change as AI continues to develop rapidly.
