“The United States is vigorously pushing ahead with the military use of artificial intelligence (AI): Defense Secretary Pete Hegseth has made it unequivocally clear that AI systems should be fully available to the armed forces for "all lawful" purposes. The Pentagon is currently at odds over this with Anthropic; while the company’s chatbot, Claude, is considered particularly well-suited for military applications, the firm refuses to authorize its use for fully autonomous weapons operations or mass surveillance. In Germany, the debate surrounding the use of AI in warfare is also characterized by reservations and fundamental questions: What roles should AI play in war, and what is ethically justifiable? What dangers does its deployment entail? How does AI alter the dynamics of the battlefield? And who bears responsibility when a system makes a mistake?
When it comes to military ethics, the US government has long pursued an outcome-oriented approach. In televised addresses, American presidents present the killing of terrorists as a success. At military events, soldiers are praised for innovations by celebrating the number of terrorists eliminated as a result. "In Europe, people would be embarrassed to applaud that," said Lieutenant General (Ret.) Ansgar Gerhard Rieks, who served as Deputy Chief of the German Air Force until the end of June 2023.
Rieks has been involved from the outset with a working group on technological responsibility within the Franco-German FCAS (Future Combat Air System) defense project and continues to shape its direction today. "In Europe, by contrast, we tend to focus on the measure itself: if the measure is justified or absolutely necessary, then we can implement it," he says. Together with Professor Wolfgang Koch from the Fraunhofer Institute for Communication, Information Processing and Ergonomics (FKIE), he is working on issues of technological responsibility in a military context—specifically, how AI can be developed and deployed responsibly within the armed forces.
Rieks currently sees the greatest benefit of AI for the military in the creation of situational pictures. Sensor and reconnaissance data as well as reports are aggregated in real time, threats are filtered out, and the overview is continuously updated. "A key advantage of artificial intelligence is that we are far more comprehensively informed." "We are able to process far more data." This reduces the level of uncertainty, says Rieks.
Situational awareness is far more precise today, and AI helps analyze information much faster. In the past, there was significantly more time for planning air operations: eight-hour cycles were common during the Cold War, with ground forces often operating on much longer timelines. What was considered fast back then seems almost absurd today.
Moreover, Rieks notes, AI often makes modern weapon systems operational in the first place. Flight control and assistance systems are what make "unstable"—yet highly maneuverable—combat aircraft like the Eurofighter controllable at all. "If you were to sit in a Eurofighter and fly it the way you flew its predecessor, the Phantom—meaning with a great deal of manual input and human control but without AI support—the aircraft would crash."
Only with AI support is it possible to handle modern technology in a way that ensures it performs effectively in the air, at sea, or on the battlefield. AI in the broader sense encompasses a wide spectrum: ranging from simple algorithms that map specific inputs to specific outputs based on fixed rules, to learning systems that incrementally improve their settings using data, and finally to complex systems that evolve through feedback and whose results cannot be precisely predicted.
The use of AI also enables new applications, according to Rieks. For instance, AI allows for relatively accurate space weather forecasting. AI also supports the armed forces in planning—such as in the rapid compilation of flight operation orders—and aids in training by allowing for better customization of exercises.
A direct advantage emerges in military operations: AI support makes attacks more precise, says Rieks. During the Second World War, entire cities and regions were bombed—often indiscriminately—to strike the armaments industry or, depending on the aim of creating terror. "The fact that today they can guide a weapon to strike precisely within a designated square meter is, of course, a development that is operationally advantageous, because they aren't using resources for something that holds no operational value."
Koch points out that technological progress is fundamentally ambivalent, including in the case of AI. As a "powerful tool," it can be used very responsibly, yet also "highly irresponsibly," he says. At the same time Koch emphasizes that AI and networked systems extend conscious perception far beyond the limits of human sensory capabilities: "Sensory organs" are spatially distributed in the form of sensors, and information can be aggregated and shared in real time—abilities humans do not naturally possess. "We are enhancing our perceptual capacity far beyond natural limits." This also expands decision-making capability—that is, the scope of what can actually be recognized, weighed, and willed. It also raises the question of whether this abundance of options and assessments might ultimately overwhelm people.
The question of responsibility for errors has reignited—most notably in the wake of the war in Iran. An incident with an as-yet-unclarified sequence of events is currently the subject of an internal US government investigation: on February 28—the first day of attacks in Iran—an elementary school in the city of Minab was bombed, resulting in the deaths of over 170 people, including many children. Preliminary findings reported by the *New York Times*, Reuters, and the AP suggest that outdated targeting data from the US military intelligence agency (DIA) may have contributed to a targeting error. The school complex had reportedly seen past military use and was located in the immediate vicinity of military positions.
Without wishing to pass judgment on the specific case at hand, General Rieks advocates for a more nuanced view of responsibility—moving beyond the traditional approach of simply holding the operator accountable. "That is why, when addressing the issue of responsibility within a systemic framework, I always suggest analyzing who bears what responsibility: sometimes it is not the pilot in the aircraft, but perhaps the person inputting the data. Sometimes it is the person operating the sensor that generated the data. And sometimes, it might even be the person responsible for the data link." It has always been the case that when decisions are based on information processed by machines or other people, a dependency arises regarding the quality of that information and the procedures used to generate it.
Responsibility is therefore "always shared," says Rieks. It begins with the researcher who develops the AI process, extends to the engineer who integrates it into the system and the person who trains it with data, and finally reaches the operator who uses it in the field. While the operator remains "ultimately responsible," in a digitized world they can "no longer be viewed as the sole bearer of responsibility." In this context, Koch emphasizes: "Data collection and data maintenance must be transparent and verifiable." Anyone wishing to authorize an AI system must therefore meticulously document the data used to train it.
He describes the combat situation as a particular challenge: because reality changes rapidly in such environments, an AI system that has been trained once may quickly become unsuitable. In such cases, it is necessary to recognize the drop in performance and carry out retraining—including "record-keeping," meaning a precise log of what was changed and what data was used for the retraining. Furthermore, sensors might supply erroneous data, or an adversary could engage in jamming and deception. Consequently, checks for data validity and robustness against manipulation are required.
Vanessa Vohs, an expert in international law at the Bundeswehr University Munich, identifies additional sources of error in AI deployment beyond the issue of outdated data. One such factor is the human tendency to "diminish critical thinking and trust the machine more than our own intuition"—a phenomenon known as automation bias.
"Excessive reliance on the machine" can lead to violations of international humanitarian law—for instance, if military and civilian targets are confused, or if disproportionate force is used when less severe measures would have been possible. Furthermore, she identifies a "black box problem" regarding the use of AI: often, it is impossible to trace why a system arrives at a particular result. "In such cases, the data basis is unclear to me, and I run the risk of drawing the wrong conclusions from the outcome." That is why extensive education and training are required—including exercises that simulate deliberately incorrect output, says Vohs. "Soldiers must be thoroughly trained to maintain their critical thinking skills."
She points out that international humanitarian law applies "regardless of the technology involved." Although there have been attempts at the United Nations for over a decade to regulate the military use of AI, these have been unsuccessful. Consequently, it will now be necessary to observe how practices evolve and what regulations emerge from them. In principle, international law distinguishes between state responsibility and individual criminal responsibility. States are bound by international humanitarian law; there is also a state obligation to review new weapon systems for compliance with international humanitarian law prior to deployment.
Individual criminal responsibility, by contrast, applies to natural persons; it is typically linked to a high threshold of intent—that is, a high degree of deliberate purpose. Regarding the use of AI, Vohs notes that problems arise when a serious violation—potentially amounting to a war crime—occurs objectively, yet "no one intended it"—neither the operator, the state, nor the manufacturer; in such cases, the requisite intent is often lacking. Consequently, there is discussion about lowering the threshold for intent.
AI manufacturers, Vohs emphasizes, are "naturally not subjects of international law in the first instance." Therefore, they are not subject to the same direct obligations under international law as states. Their responsibility lies primarily in adhering to self-imposed standards and complying with national law—for instance, regarding product liability, criminal law, and export regulations. Furthermore, if an AI is used for both civilian and military purposes (dual-use), additional compliance obligations under EU rules, such as the AI Act, come into play. At the level of international law, states cannot evade responsibility simply by pointing to a "malfunctioning product."
AI is increasingly assuming the role of specialists, says Wolfgang Koch of the Fraunhofer Institute. It provides analyses and "reports," while the human acts more as a supervisor who contextualizes these results, weighs them against one another, and bears responsibility for them. To do this, the human must maintain a sufficient understanding of—and a "feel" for—the systems' capabilities and limitations. General Rieks, too, does not anticipate fully automated warfare. From a military perspective, this would be inefficient, as the objective could be missed in many situations without human contextualization and decision-making.” [1]
1. Was darf KI im Krieg? Frankfurter Allgemeine Zeitung; Frankfurt. 07 Apr 2026: 12. Von Magdalena Tröndle
Komentarų nėra:
Rašyti komentarą