Military AI does not need to fire a weapon on its own to influence who gets targeted. Systems can flag figures in a gunsight, connect battlefield information to an artillery unit or recommend a course of action. When a human is left with an approve button, the central question is whether that person still has a meaningful chance to judge the decision.
Human approval can hide machine influence
Fully autonomous weapons, which select and fire at targets without human input, have drawn serious concern from governments and civil society. After years of discussion, parties to the UN’s Convention on Certain Conventional Weapons agreed in May on recommendations that included limiting the duration, geographical scope and scale of operations. The recommendations were nonbinding, but they acknowledged that a person should be involved in the process leading to a killing.
At the same time, systems that assist a human decision have become more capable. A gunsight can identify a possible human target, while other tools can help analyze intelligence or match a battlefield target with an artillery unit. In these cases, a human may still make the final call, but machine-generated suggestions can shape what they see and how quickly they must respond.
That is a change from earlier tools that mainly alerted an operator to information, such as the radar retired Air Force lieutenant general Jack Shanahan used in the 1980s. Machine learning can offer greater apparent competence and breadth, potentially encouraging people to hand more decision-making to the system.
Speed and scale change the stakes
Some systems focus on a single weapon or observation. Others help coordinate many steps. The Ukrainian army uses GIS Arta to pair known Russian targets with artillery units the algorithm identifies as best placed to fire. A report by The Times said the software reduced the time between detecting a target and an artillery response: a process that previously took 20 minutes reportedly takes one.
Such speed can affect how much time people have to check information and weigh consequences. A tool may suggest where to look, which unit should respond or what route an attack team should take. Even if no single system makes the final decision, several recommendations can narrow the options a person sees.
Some militaries are working toward linking these tools into what the Pentagon calls a “kill web,” connecting weapons, commanders and soldiers. Rafael’s Fire Weaver, sold to the IDF and demonstrated to the US Department of Defense and the German military, finds enemy positions, directs a unit considered best placed to fire, and puts a crosshair on the target in the unit’s weapon sights. In a video of the software, the human chooses between “Approve” and “Abort.”
Responsibility needs more than a button
The Department of Defense lists “Responsible” first among its ethical principles for artificial intelligence. Shanahan, who oversaw development of those principles, says someone must ultimately be held accountable. Keeping a human hand on the trigger preserves a person’s formal role, but it does not settle what responsibility means when an algorithm has shaped the information and recommendation behind that action.
AI tools can glitch in unusual ways, and a person may not know when an answer on the screen is wrong. Their speed may also leave too little time to decide whether an action is legal. The challenge is especially difficult because military accountability already has to distinguish unavoidable tragedy from malign intent, misdirected fury or gross negligence.
Courtney Bowman of Palantir describes the shift as requiring a new ethical construct. That challenge grows when systems combine data and recommendations into a rapid sequence. In a demo of Palantir’s Artificial Intelligence Platform, the software flags a potentially threatening movement, suggests sending a drone to investigate, proposes plans to intercept the force and maps a route for the selected team.
Design can preserve room for judgment
One DARPA program, Urban Reconnaissance through Supervised Autonomy (URSA), was designed to help robots and drones act as forward observers during urban operations. Following advice on legal and ethical issues, the system was limited to labeling people “persons of interest,” rather than “threats.” Its designers also removed a proposed ability to discern a person’s intent.
Those choices reflect a concern that a machine’s label could push a soldier toward a conclusion the available evidence does not support. The program’s advisory group also raised a legal concern: according to Brian Williams, no court had positively asserted that a machine could legally designate a person a threat.
Keeping a human involved matters only if that person can question the system, understand the limits of its output and take time to decide. As military AI takes on more of the work of identifying and prioritizing targets, responsibility cannot rest on the presence of an approval button alone. It depends on whether people retain real authority over the judgment that follows.