A Russian drone equipped with an Nvidia chip killed three civilians during an attack carried out in July in the city of Zaporizhzhia, southeastern Ukraine. This was reported by drone experts, Ukrainian military commanders, and a forensic medical team that investigated the wreckage of the weapons and other attacks in the region.
This incident drew attention because, according to authorities and experts interviewed by The New York Times, the drone was controlled by an experimental artificial intelligence (AI) system without human involvement in selecting the final target.
The youngest victim was Titana Bubinets, a 19-year-old university student. She was near a gas station when a small Russian drone, resembling an airplane model, descended from the sky and attempted to approach the location. The aircraft crashed into a wall and exploded, scattering shrapnel that struck the young girl and other people.
Bubinets was taken to the hospital but died. The other two victims, Oleksiy Svyryn, 41, and Roman Karpiy, 48, later died from their injuries.
How Target Selection Occurred
According to experts and military personnel interviewed by Time, human operators programmed the drone to move toward a specific gas station. However, upon getting closer, the aircraft allegedly selected the precise target itself—likely propane tanks—based on the training it received to recognize and attack these objects.
According to Ukrainian authorities, the modules found in the drones were not encrypted. This allowed investigators to review the images of the area uploaded to the computer and check how the drone used visual references to follow its route.
Investigators also stated that they analyzed the code in the equipment, determining the types of objects the drone was trained to recognize and attack, including propane tanks.
Autonomous AI systems can be trained on thousands of images to recognize certain categories, such as rivers, military trucks, or people. When using drones with autonomous target selection, cameras search for these elements, and the system identifies those that match the training received. This technology potentially can provide greater accuracy than what a human pilot can achieve in certain circumstances.
However, critics of these weapons point out that the lack of human intervention may increase the risk of errors and violations of the laws of war. The system may fail to correctly distinguish between civilians and combatants or misinterpret an object on the battlefield.
First Documented Case
Kateryna Bondar, a senior researcher at the Wadhwani AI Center at the Center for Strategic and International Studies in Washington (USA), stated that the attack on July 6th is the first documented case where civilian deaths were caused by a Russian drone equipped with such an autonomous system. In Bondar's opinion, the presence of the Nvidia module is the 'best evidence' that Russia was experimenting with autonomous AI.
An advisor to Ukrainian President Volodymyr Zelensky previously published information on social media that Russia was testing AI systems capable of independently selecting targets. However, no prior reports had been made about the deaths of people or the details of the use of the Nvidia module.
Russia and Ukraine have already used AI in various drone models before the appearance of fully autonomous systems. In earlier models, AI was primarily used in the so-called 'last mile' of the attack. In this scenario, a person remotely selects the target, and the AI system takes control only at the final stage to help the drone reach it.
In fully autonomous systems, the selection of the final target is also no longer the responsibility of a human. Older models are usually controlled by operators via radio transmitters or fiber optic cables. An alternative option is pre-programming the coordinates the drone should follow.
With the expansion of drone use on the Ukrainian battlefield, pilots have become important targets. A drone controlled autonomously by AI, conversely, does not need a human pilot and can be more accurate than an ordnance programmed only to hit specific coordinates.
Local air defense authorities claim that Russian forces are targeting civilian objects, such as fuel pumps and propane tanks at gas stations, during tests of autonomous drones in Zaporizhzhia. Military conscription centers could also have been used as targets, although the reason for choosing these locations for testing is unclear.
Vadim Kushnikov, a drone expert from the Kharkiv Aviation Institute, noted that the drone responsible for the death of three civilians was a 'pre-programmed tool trained in virtual reality to track specific objects.'
Its functioning differs from traditional software. While ordinary programs follow step-by-step rules set by humans, AI systems develop their own rules when identifying patterns in huge volumes of data—a process known as machine learning. This allows the system to handle situations that were not specifically foreseen by its developers, but it also means that no one can completely predict its behavior.
Human rights organizations and the International Committee of the Red Cross strongly oppose the use of AI weapons without human involvement in the decision-making process. In their view, autonomous systems leave machines to interpret international law for determining legitimate targets during wartime. Nevertheless, some experts support a different approach, believing that future weapons must necessarily be equipped with AI systems. Their argument is that such weapons could be safer because they can make decisions before striking a target, even if imperfectly. Unguided munitions and conventional artillery shells, on the other hand, simply follow their trajectories without identifying what they hit.
The drone involved in the attack was part of a group of approximately six aircraft, according to the regional military governor, the city's air defense commander, and the air defense service on duty that day. None of the drones emitted radio signals. The aircraft also had other signs related to Russian tests of autonomous AI systems, including wave launches.
Some AI-equipped drones tested by Russia exploded upon hitting targets, while others did not detonate and remained intact, according to Ukrainian authorities. Ukrainian military personnel showed Times journalists the wreckage of two such drones. The newspaper photographed the Nvidia computer modules found inside the aircraft and sent the pictures to the company. Nvidia confirmed that the photos showed the Jetson Orin mini-computer. Modules of this type were found in various models of Russian weaponry. Ukraine is also conducting tests of AI systems capable of selecting targets. Mykhail Fedorov, the recently departed Minister of Defense, stated in an interview with Times at the end of July that Ukraine had been testing a fully autonomous AI system in the Russian-occupied Crimea in previous months. According to Fedorov, the devices were used against fuel storage facilities and military equipment, and he stated that there were no civilian casualties during these tests.
In the Zaporizhzhia attack, the drones lacked antennas for transmitting test data. However, Russian forces often use reconnaissance drones capable of recording attacks. The city is also considered by authorities to be a place where Russian spies providing damage assessments are located. The absence of an antenna itself does not prove that the drone could select its own targets; the aircraft could have been programmed to follow a set route using algorithms similar to those applied for decades in aircraft autopilot systems. In this case, the use of an Nvidia chip would be unnecessary.
Ukrainian air defense experts assert that Russia began conducting test flights with Molniya AI-controlled drones in May. Molniya is a mass-produced fixed-wing aircraft model. Russians also began testing autonomous target selection systems on V2U, another fixed-wing aircraft, several months earlier. Part of the V2U tests was aimed only at guidance in the 'last mile'. Zaporizhzhia is only 14.5 kilometers from the front line. Some residents interviewed on site stated that the possibility of aircraft hovering over the city, making killing decisions independently, causes fear. However, Ukrainian cities have been targets of attacks for years, and the lives of the population have already changed deeply due to the use of less complex drones. One response to this threat is interceptors, which are currently piloted by men and women considered local heroes. Another measure is the installation of anti-drone nets. In Zaporizhzhia and its surroundings, they cover about 386 kilometers of roads. Camouflage is also used as a protective mechanism. For example, plastic elements are installed around propane tanks at gas stations at unusual angles, attempting to confuse image recognition systems used by AI. Koronal Minaev, the city's air defense commander, commented on Russian forces: 'They are training the technology.' He added that AI is good, but not yet perfect. 'They are looking for errors and will correct them.'
}, {