Opinion | The Ukraine ‘Kill Zone’ Is AI’s Testing Ground

Lviv, Ukraine

A fixed-wing drone circles below the clouds, and its camera streams back an aerial view of a field. I’m standing in that field, and on a laptop screen I see the two vehicles beside me, both marked with a green X that indicates a target. The drone can carry more than 26 pounds of explosives, and it approaches far too fast to outrun. The pilot intervenes seconds before it could crash into us, sending the device swerving away with a flick of his finger.

This wasn’t a near-death experience but a glimpse into how artificial intelligence is changing warfare. Norda Dynamics, a Ukrainian-Estonian defense company, was showing me its terminal-guidance system, Underdog, which lets drones autonomously complete the last mile or so of a strike mission. CEO Nazar Bigun says Underdog is widely used in combat in Ukraine, and once a pilot lets this system take over, the drone successfully hits its target more than 90% of the time.

Underdog is one example of how Ukraine is pioneering wartime AI. Experts here are tapping AI both to analyze data and automate weapons. Ukraine collects an incomprehensible amount of battlefield information. A single brigade can collect the equivalent of 10 months of video every day, and nationwide the Ukrainian-developed battlefield-management system Delta takes in petabytes (quadrillions of bytes) daily.

The country’s most sophisticated units use their own closed AI systems to turn this data into intelligence to act on. With financial support from the British government, Ukraine in March launched A1, a center within the Defense Ministry focused on integrating AI into the military. “Who understands the battlefield data and can make decisions faster, Russia or Ukraine—this side will have the biggest benefits,” says Danylo Tsvok, A1’s CEO.

The center is rolling out Spectr, a system in which agentic AI can quickly analyze data and perform other tasks that typically require a staff or several operational officers. A1 has also developed tools to let tactical-level units work with some sensitive data in the battlefield on a laptop without an internet connection. In its next stages Spectr will be able to help plan operations, providing commanders a menu of options for how to carry out a mission as well as each suggestion’s probability of success. A future development stage involves AI-enabled prediction, such as forecasting enemy actions.

Such analysis isn’t without risk. There’s a long history of governments thinking they understand problems because they have extensively quantified them. That can lead to catastrophic overconfidence and error if the data set used is inaccurate or fails to account for some essential facet. But across the front line in Ukraine, I observed soldiers and commanders approaching AI with healthy skepticism. If it saves lives or brings victory closer, they embrace it; in combat you find out quickly what works and what fails.

Meanwhile Mr. Tsvok and his team are thinking about how to encourage, not curb, critical thinking. Spectr explains its reasoning and notifies the user when it lacks data to complete a task. A human commander decides whether to accept Spectr’s recommendations and is responsible for the outcomes.

It’s good when AI can reduce the danger to infantrymen. But not when AI replaces a commander in the decision-making process, especially when that choice involves human lives, says Yaroslav Azhnyuk of The Fourth Law and Odd Systems, two Kyiv-based companies that develop autonomous drones and interceptors. Autonomy in weaponry inherently means “replacing humans . . . in the loop or process or action,” he says. But “the mission is still decided by the human,” who “is deterministically setting the boundaries.”

Autonomy in weapons is a decision-making spectrum, not a binary choice. Using terminal-guidance systems, a drone locks on to and strikes a human-identified target. In that case, is launching that autonomous drone much different from shooting artillery? It could be more ethical if it’s more precise and lowers the risk of collateral damage. Such last-mile technology is less autonomous than a drone that recognizes a target on its own and proceeds to carry out the strike. A fully autonomous mission, which is on the short-term horizon, could involve a drone that takes off, flies to the combat area, identifies targets and attacks all on its own, but within the framework of rules set by a commander.

Kyiv already relies heavily on autonomy in air defense. Commercial planes aren’t flying over Ukraine, so there’s minimal risk to civilians. The air-defense system knows where Ukrainian drones, missiles and fighter jets are. That means a commander could draw responsible boundaries and issue orders, such as: We know that none of our pilots or drones are in this 10-mile radius, so take down anything in that airspace.

Extend that concept to the front, where an order could soon look like this: None of our troops are within this area, and our reconnaissance suggests neither are civilians. So go ahead, drone, destroy the highest-value target you can identify, factoring in cost and your payload.

Both AI analysis and autonomous weapons will change the battlefield and tactics. In Ukraine the front already is no longer a line but a “kill zone” where anything or anyone moving is vulnerable to attack. Autonomous drones will likely expand the kill zone, while AI analysis will accelerate strikes within it. AI analysis can also speed up military innovation, while autonomous systems offer an asymmetric advantage that can offset manpower disadvantages or compound manpower advantages.

While some Western tech elites fret about agents going rogue, the Ukrainian AI pioneers I interviewed worry about authoritarian adversaries winning the AI arms race. “We can’t avoid using AI,” Mr. Tsvok says. “The highest risk in Ukraine now is, don’t take the risk.”

Ms. Melchior is a London-based member of the Journal’s editorial board.