By Nathan McNaughton, MEng ’25 (EECS)
Image 1. The Air Force’s Variable Stability In-flight Simulator Test Aircraft (VISTA) is flying over Edwards Air Force Base in the Mojave Desert. (DARPA, 2024)
Your afterburners ignite as you bank hard left, chasing the enemy jet across the sky above the Mojave Desert. You line up the shot—miss. You try again—miss. Your opponent pulls enough Gs to flatten an elephant, and within seconds, you’ve been outmaneuvered for the fifth time in a row. You look over to see who has been tormenting you for the past hour and see an empty cockpit. This may sound like a science fiction piece, but this is reality at Edwards Air Force Base. The Air Force Test Pilot School’s VISTA (Variable Stability In-flight Simulator Test Aircraft) is the first ever F-16 to have autonomous dogfighting capabilities (Air & Space Forces
Magazine, 2024), leaving experienced aviators asking, “Would you let an AI be your wingman?”
With the rise of AI in aircraft like the VISTA and increased computing power, AI copilots will redefine what it means to be a military pilot by enhancing decision-making and getting pilots out of the cockpit. However, AI flight systems will never replace the experience and intuition of a human pilot; they will serve as a new framework for quicker decision-making.
For me, this debate of a human vs AI in the sky is not an abstract idea; it is the future of Air Force engineering that I will shape-and that will shape me. I commissioned as an Air Force Officer from UC Berkeley’s Air Force ROTC as the #1 cadet in my class, earning a slot at Euro-NATO Joint Jet Pilot Training to learn how to fly fighter jets alongside the most promising aviators from our NATO allies after completing my Master’s in Electrical Engineering and Computer Science (EECS).
Merging my two passions, flying and EECS, has been a dream of mine since I interned at Edwards Air Force Base, where I broke the sound barrier in a fighter jet (see Image 2 below) in the same airspace as the VISTA’s first autonomous test flight. After that experience, I wondered whether I would be a better engineer or a better pilot, and what better way of testing that question than coding an AI that I will eventually fly with… or against.
Image 2. Me right before I puked everywhere during my supersonic flight at Edwards Air Force Base.
The idea for the Air Force’s VISTA jet started at the Defense Advanced Research Projects Agency (DARPA) when they were developing software to prevent maneuvers that were out of an aircraft’s limits. Once the limit-protection software was successfully integrated, DARPA introduced an AI agent designed to function within those boundaries. And the question of “Could these agents dogfight?” was born.
As shown in Figure 1, traditional flight automation sits in the lower-left corner, where both problem complexity and pilot workload are low. An example of this is what we mentioned above, correcting and preventing an aircraft from maneuvering out of its limits. The current frontier lies in the upper-right corner, where both complexity and workload are high, and this is where we can truly use AI autonomously on the battlefield. The dogfight is the bridge from simple,
physics-based maneuver systems (the red curve) to AI capable of real-time action (the blue curve).
Figure 1. Dogfighting as the bridge from simple physics-based maneuver systems (red curve) to complex nonlinear systems (blue curve). Dogfighting is the key to true autonomy in combat (DARPA ACE program, 2020).
The top engineers and scientists trained the AI using a large-scale model and simulation to run various situations. Following the research, DARPA launched the AlphaDogfight Trials in 2019 and 2020, which had AI systems against human pilots in F-16 fighter jet simulators. (JHU Applied Physics Laboratory, 2020) The result was a shocking clean sweep of 5-0 for the computer! Interest naturally spiked after this simulation and propelled researchers to implement
this technology in real aerial combat. (DefenseNews, 2024).
The first step was more training, which had AI agents control virtual F-16s. After two years of development, by late 2022, the Air Force flew on the actual VISTA at Edwards AFB. The VISTA was put through the ringer with over 21 test flights from December 2022 through September 2023. The VISTA engaged in dogfights against the best F-16 pilots in the Air Force, flying at speeds up to 1,200 mph! (DefenseScoop, 2024) During these missions, safety pilots were ready to take over the controls but never had to deactivate the AI during any dogfight.
Now I know what you are thinking, “Who the heck won?!” And trust me, I was thinking the same exact thing. I searched high and low for this answer, searching dozens of websites, publications, and articles, only to find vague answers. This wild goose chase led me straight to the nest where I spoke to the former Commandant of Test Pilot School, Colonel James “Fangs” Valpiani.
Image 3. Still image of the VISTA from Inside the Air Force’s Elite Test Pilot School (Eckholm, 2025).
Colonel Valpiani has more than 1,400 hours in 30+ different aircraft, along with a PhD in Astronautical Engineering. He also currently serves as chief of combat autonomy at DARPA, making him the most qualified person in the world to talk about this subject.
When I asked Colonel Valpiani the question that was eating me alive, “Who won?!” he chuckled and replied that it was classified information. Although I was unsatisfied with the answer, Colonel Valpiani said it did well enough to warrant further research.
Even though we do not know the results, I think it is safe to say that they were promising, given the interest in further research. However, after talking to Colonel Valpiani, I realized that the VISTA had a chink in its armor. During the flight test, the VISTA was being fed perfect state data from the enemy aircraft. This fact means that the VISTA knew the enemy’s precise position, velocity, and orientation, which is an issue because that would never happen in a real dogfight. I was shocked, and it made me question the potential of this technology.
As Col Valpiani mentioned, there is no public data from the VISTA program. Figure 2 shows what perfect-state data looks like in a research simulation (Chen et al., 2025). The blue line represents one AI-controlled aircraft, and the red line represents its opponent. Both agents know each other’s exact state in real time.
Figure 2. Example of simulated AI dogfighting using perfect state data. (Chen et al., 2025)
To dig deeper into the future of AI in flight systems, I spoke with the newest member of the Berkeley EECS faculty, Professor Victoria Coloneleman, who was the Chief Scientist for DARPA and the Air Force during the VISTA’s development. When I mentioned the issue of being fed perfect state data, the question of “Will this AI ever be ready for the battlefield?” was at the forefront of the conversation. Professor Coloneleman explained to me that there is no standard for an AI “being ready” because there has never been a technology like this before in the
military.
The interactions with Colonel Valpiani and Professor Coloneleman showed me that we are nowhere near an AI fighter pilot. From all the available evidence, AI is capable of making decisions on the fly and outmaneuvering real pilots in theory, but in the real world, without perfect state data of the enemy and human partnership, the AI is a target in the sky.
This whole time, I was searching for who won and if humans would be replaced. However, the Col Valpiani and Professor Coleman’s answers showed me that the goal of this program isn’t to replace anyone. The goal of AI in the cockpit is to expand what humans can do. In combat, a pilot’s brain has to process a flood of information in a split second, such as altitude, fuel, radar, threats, and friendly positions. AI in the cockpit can allow for human pilots to have more time to
use their judgment and creativity they developed from experience, while the machine handles the processing of information. AI will never replace the experience and intuition of human pilots; they will serve as a complement to make them more efficient.
I may never know who really won that dogfight over the desert, but I do know that AI can be my wingman anytime.
References
ACE Program’s AI Agents Transition from Simulation to Live Flight. ACE program’s AI agents transition from simulation to live flight. (2023, February 13).
https://www.darpa.mil/news/2023/ace-program-transition
Chen, C., Song, T., Mo, L., Lv, M., & Lin, D. (2025, March 20). Autonomous dogfight decision-making for air combat based on reinforcement learning with automatic opponent sampling. MDPI. https://www.mdpi.com/2226-4310/12/3/265
DARPA. (2020, August 26). AlphaDogfight Trials Foreshadow Future of Human-Machine Symbiosis. Alphadogfight trials foreshadow future of human-machine symbiosis. https://www.darpa.mil/news/2020/alphadogfight-trial
Everstine, B. (2020, August 21). Artificial intelligence easily beats human fighter pilot in DARPA trial. Air & Space Forces Magazine. https://www.airandspaceforces.com/artificial-intelligence-easily-beats-human-fighter-pilot-in-darpa-trial/?utm_source=chatgpt.com
Harper, J. (2024, April 17). Pentagon takes Ai dogfighting to Next Level in real-world flight tests against human F-16 Pilot. DefenseScoop.
Johns Hopkins University Applied Physics Laboratory. (2020). AI bests human fighter pilot in Alphadogfight trial at Johns Hopkins APL. Johns Hopkins University Applied Physics Laboratory.
https://www.jhuapl.edu/news/news-releases/200828-AI-bests-human-fighter-pilot-in-AlphaDogfight-trial-at-APL
Losey, S. (2024, April 19). US Air Force stages dogfights with AI-flown fighter jet. Defense News. https://www.defensenews.com/air/2024/04/19/us-air-force-stages-dogfights-with-ai-flown -fighter-jet/
Sam Eckholm. (2025, September 27). Inside The Air Force’s ELITE Test Pilot School. YouTube. https://www.youtube.com/watch?v=bweCUEA1URo&t=493s
Tirpak, J. (2024, April 22). First AI dogfights focus on safety, building for CCA applications . Air & Space Forces Magazine. https://www.airandspaceforces.com/usaf-first-ai-against-human-dogfights/