How AI learns to fly a fighter jet

A grey F-16 Fighting Falcon climbed away from Eglin Air Force Base in Florida sometime in June 2026. A human pilot was strapped into the cockpit. But the human was not flying the plane. An AI agent was doing that.

This was not a drone. It was not a simulation. The VENOM F-16s are modified airborne test platforms equipped with specialised hardware, software, and instrumentation designed to enable artificial intelligence agents to pilot the aircraft while human pilots remain in the cockpit to monitor the AI agents. The pilot was there to watch, and to take back control if something went wrong.

DARPA and the U.S. Air Force have begun flight testing an F-16 controlled by an artificial intelligence agent, moving autonomous combat technology onto a standard fighter platform. The milestone is part of the Viper Experimentation and Next-generation Operations Model, or VENOM, programme.

If you've been following AI news, you might already know that AI can beat world champions at chess or Go, write code, and pass medical licensing exams. An AI flying an F-16 feels like a natural extension of that trend. But it's worth slowing down to understand what's actually happening here, because it's quite different from the language models most people are familiar with.

How you teach a machine to fly

Language models like ChatGPT learn by reading vast amounts of text and learning to predict what comes next. Flying a fighter jet is a completely different kind of problem. You can't learn to fly by reading books about flying. You have to try things, see what happens, and adjust.

The technique behind AI systems like this is called reinforcement learning. The basic idea is simple: the AI tries something, gets a score based on whether that action was good or bad, and gradually learns to favour the actions that lead to better scores. Repeat that process millions of times and the AI gets very good at whatever you're scoring it on.

For aerial combat, that might mean: did you keep the aircraft stable? Did you close the distance to a target? Did you avoid a simulated missile? Each action gets a reward or a penalty, and over time the system learns which combinations of inputs lead to the outcomes you want.

The tricky part is that real flying is unforgiving. You can't make thousands of mistakes in a real plane the way you can in a video game. So the Air Force began simulations in 2024 with one-versus-one engagements, expanded them to two-versus-two scenarios, and tested both within-visual-range and beyond-visual-range missions. Only after extensive virtual testing did the AI get to fly the real thing.

An individual scenario could be repeated 1,000 times to examine variations in decisions and aircraft behaviour. A human pilot can't safely do that. A simulated AI can do it overnight.

From a one-off experiment to standard jets

The VENOM programme didn't come from nowhere. The groundwork was laid by DARPA's ACE programme and its X-62A VISTA jet, which in September 2023 flew the first-ever within-visual-range dogfight between an AI-controlled F-16 and a human-flown one. That was proof that the concept worked at all.

But the X-62A was a unique, one-off experimental aircraft. While DARPA describes the X-62 as a "one-of-a-kind" aircraft, the agency says the VENOM project proves that standard fighters from the operational fleet can also deploy with AI enhancements.

That's the significance of VENOM. It's not a bespoke research machine. It's a kit. The VENOM Autonomy Kit connects with the aircraft's flight controls and mission systems without changing the F-16's core software. A pilot remains in the cockpit and can switch between human and AI control during testing. In principle, this approach could be rolled out across a whole fleet.

The VENOM autonomy kit interfaces with the F-16's flight controls and mission systems, with the pilot able to toggle between human and AI control of the aircraft "with the flip of a switch", according to DARPA.

That detail matters. The AI isn't taking over permanently. It's one mode among several, and a human can step in at any moment. This is a pattern you'll recognise if you've used autopilot on a commercial flight, or lane-keeping assist in a modern car. The AI handles the mechanics; the human retains authority.

What the AI is actually doing up there

When the AI agent is flying the F-16, it's processing sensor data from the aircraft many times per second and issuing control inputs: adjust the throttle, move the control surfaces, change heading. The decisions happen faster than a human can consciously react.

Earlier VENOM tests gave the AI system control over thrust as well as the flight controls, which is significant. Managing energy (speed and altitude) in air combat is one of the most demanding parts of the job. Human pilots spend years learning to do it intuitively.

The flights featured increasingly complex air combat scenarios, including offensive high-aspect nose-to-nose engagements that involved the dogfighting jets passing within 610 metres of each other at speeds of 1,040 knots. Shield AI later told FlightGlobal that its Hivemind autonomy software showed the ability to "improvise" and develop novel combat tactics.

That word "improvise" is worth noting. The AI wasn't following a scripted set of moves. It had learned, through millions of simulated repetitions, a general strategy for handling situations. When faced with a scenario it hadn't seen before, it applied that strategy in a novel way. This is what makes modern reinforcement learning systems different from the rule-based autopilots of earlier decades.

The human is still in the loop, on purpose

It would be easy to read this story as "AI replaces pilot". That's not quite what's happening, and the distinction matters.

"It's important to understand the 'human-on-the-loop' aspect of this type of testing, meaning that a pilot will be involved in the autonomy in real time and maintain the ability to start and stop specific algorithms," said Lt. Col. Joe Gagnon, 85th TES commander. "There will never be a time where the VENOM aircraft will solely fly by itself without a human component."

Before flight, engineers used software-in-the-loop and hardware-in-the-loop testing to verify that autonomous commands could not exceed structural limits or impose unsafe physiological loads on the pilot. In other words, the system has hard constraints built in. The AI cannot tell the jet to do something physically impossible, or something that would harm the person inside it.

This "human on the loop" model is increasingly common in high-stakes AI deployments. The AI does the fast, repetitive, data-intensive work. The human provides oversight, makes high-level decisions, and can override at any point. It's different from the simpler "human in the loop" model (where a human approves every single action) and very different from fully autonomous operation (where there's no human involvement at all).

Where this gets philosophically interesting is in thinking about what happens next. The near-term goal isn't to replace pilots entirely. The goal is to pave the way for human pilots to oversee teams of autonomous, uncrewed aircraft, DARPA said. One human pilot, in one plane, directing several AI-controlled wingmen. That's the operational vision.

If this one was useful, there's plenty more on the site. Pieces on how AI works, plus coverage of AI news, the downsides included. All free to read, no account needed.

What this tells us about AI more broadly

The VENOM programme is a military story, but it illustrates something important about how AI systems work in the real world, beyond chatbots and image generators.

First, simulation is essential. DARPA said the approach creates infrastructure for faster and more scalable development of combat AI across the joint force. The reason AI can learn to fly at all is that you can run millions of virtual training flights cheaply and quickly. The same principle applies everywhere: AI systems for driving cars, managing power grids, or scheduling logistics are trained in simulation before they touch the real world.

Second, the gap between "it works in simulation" and "it works in reality" is where most AI deployments either succeed or fail. "AI has tremendous potential to help humans manage this complexity in beyond-visual-range combat, but many hard questions remain concerning the performance and trustworthiness of combat AI in the extreme fog and friction of modern war," a DARPA programme manager noted. That caution applies well beyond the military. The real world is messier, noisier, and more unpredictable than any training environment.

Third, and perhaps most important, the question of how much autonomy to give an AI system is never purely technical. It's a question about accountability, trust, and what we want humans to remain responsible for. Letting an AI fly a plane is one thing. Letting an AI decide to fire a weapon is another, and that boundary is something DARPA is being deliberately careful about.

VENOM aircraft will support DARPA's Artificial Intelligence Reinforcements programme, which will evaluate multiple AI agents during live-flight tests. Future experiments are expected to include multi-aircraft operations and support development related to Collaborative Combat Aircraft programmes.

So the story isn't really "AI replaces pilot". It's closer to "AI expands what a single pilot can oversee and control". One human, multiple machines, all working together. Whether that's reassuring or unsettling probably depends on who you are and where you're standing. But it's where things are heading, and it's worth understanding how the technology behind it actually works.