By Javier Rollon · March 26, 2026

Last April, something happened at Edwards Air Force Base that stopped me mid-coffee. DARPA announced that an AI-controlled F-16 had just dogfought a human pilot in another F-16. Not in a simulator. In the actual sky, at actual speed, with actual G-forces. Two pilots sat in the AI jet monitoring the algorithms, hands off the controls, while the machine made split-second combat decisions around them.

I've been building flight simulator aircraft for over fifteen years now. I've modeled cockpits switch by switch, tuned flight dynamics parameter by parameter, spent months making sure a virtual CRJ-200 behaves the way real pilots tell me it should. And honestly? That DARPA announcement made me rethink what I do for a living.

The X-62A VISTA — When AI Learned to Fight

Here's what actually happened. DARPA's Air Combat Evolution program — they call it ACE — took a modified F-16 known as the X-62A VISTA and loaded it with machine learning algorithms developed by companies like Shield AI and EpiSci. Between December 2022 and September 2023, they flew 21 test flights. The AI controlled the aircraft through offensive and defensive combat sets, including the first-ever real-world AI versus human dogfight.

Did the AI win? DARPA won't say — "national security reasons." But they did say the AI agents "performed well" across multiple scenarios and never violated safety norms for flight. Think about that. A machine learning system, trained initially in simulation, flew a real fighter jet through combat maneuvers and stayed within human safety standards the entire time.

What gets me is the speed of iteration. The ACE team was sometimes updating flight control software within a single day based on test results. In my world, tuning a flight model takes weeks of testing. These AI teams are operating on a completely different cycle.

What This Means for Aircraft Development

Here's where it gets personal. When I model the Boeing 747 or the Space Shuttle for X-Plane, I'm essentially hand-coding physics that took engineers decades to understand. Blade-element theory for propellers, ground effect coefficients, engine spool-up dynamics. It's painstaking work, and I love every minute of it. But AI can now learn these relationships from data.

Boeing and Airbus are already using machine learning for structural optimization — figuring out where to place reinforcement material in a wing, for example. Traditionally that's finite element analysis by human engineers. AI doesn't replace the physics. But it finds solutions in the design space that humans would take years to discover.

The military side is moving even faster. The Air Force's Collaborative Combat Aircraft program plans to spend billions on autonomous drones that fly alongside manned fighters. One human pilot commanding multiple AI wingmen. Secretary Kendall put it bluntly: "If a human being is in the loop, you will lose. You can have human supervision... but if you try to intervene, you're going to lose."

That's a sobering statement from the person in charge of buying fighter jets.

AI in Military Simulation — Already Here

The simulation community was actually ahead of the curve on this. Games like DCS World have had AI opponents for years. They're decent but predictable — you learn their patterns after a few fights and exploit the same weaknesses every time. But the new generation of AI, trained through reinforcement learning rather than scripted behaviors, is something else entirely.

The same principles apply to naval simulation too. Silent Hunter III, the legendary submarine sim, used convoy AI that adapted to player behavior back in 2005. Escort destroyers would change search patterns based on your attack profile. Twenty years later, the Silent Hunter III community is still pushing boundaries with mods that make the AI even more unpredictable. If you're interested in how AI opponents can feel genuinely threatening in a sim, that community is worth checking out.

The Certification Problem Nobody Talks About

There's a catch nobody's solved yet. As DARPA's own team admitted — there is currently no pathway to achieve flight certification for an aircraft controlled by machine learning. That's huge. We can prove an AI can fly a fighter jet through a dogfight. We cannot prove, in the regulatory sense, that it will always do so safely.

Traditional avionics get certified through deterministic testing. You know exactly what the system will do given any input because you programmed every response. Machine learning doesn't work that way. The AI's reasoning is opaque — a neural network with millions of parameters doesn't explain why it made a particular decision. It just... decides.

For military applications, that's a risk commanders might accept in exchange for combat advantage. For commercial aviation? No chance. Not yet, anyway. We're probably a decade away from AI having meaningful control over passenger aircraft beyond what autopilot already does.

What It Means for Flight Sim Development

From where I sit at JRollon Planes, AI is going to change flight simulation in three ways.

First, procedural flight modeling. Instead of me spending six months hand-tuning the Jetstream 32's propeller behavior, AI could learn it from real-world flight data in a fraction of the time. I'm not there yet. But I can see the path.

Second, adaptive AI opponents. The current state of AI enemies in combat sims is embarrassing compared to what DARPA's doing. We'll see reinforcement learning agents that adapt to your flying style, force you to change tactics mid-fight. That's when sim combat gets genuinely interesting.

Third — and this is the one that keeps me up at night — AI-generated aircraft. Not just the flight model, but the 3D model, the textures, the systems simulation. AI tools that can take a specification document and produce a flyable aircraft add-on. We're not there yet, but the gap is closing. Whether that's exciting or terrifying depends on your perspective. Probably both.

The one thing AI won't replace? The obsessive attention to detail that makes a great simulator aircraft feel real. A machine can optimize parameters. It can't tell you that the throttle friction on a real SF-260 feels like dragging your thumb through warm butter. That kind of knowledge only comes from sitting in the actual cockpit. And that's what keeps human developers relevant — for now.

Javier Rollon is the developer behind JRollon Planes, creating aircraft add-ons for X-Plane since 2010. Follow his work on Twitter.