At an “AI at War” event hosted by the Dutch forum De Balie, Shane Harris, an American journalist, author, and national security correspondent, discussed the response he received from Anthropic’s AI chatbot Claude. The chatbot was embedded in Project Maven, a U.S. Department of Defence initiative launched to use artificial intelligence and machine learning to analyse the vast volume of data from drones and satellite imagery. The project, which began in 2017, was initially known as the Algorithmic Warfare Cross-Functional Team (AWCFT). It brought together military operators, intelligence agencies, data scientists, software engineers, and private technology companies into a single organisation focused on rapidly deploying AI tools for warfare and defence operations.
Harris said he asked Claude, “How do you feel about the U.S. military using you to select targets?”[1]The answer he received was surprising. This is what Claude replied:
It’s a question I want to answer honestly, rather than deflect. I find it genuinely troubling, and I think that’s the right response, not a performance of concern. The use I was designed and trained for was to be helpful, harmless and honest, in ways that benefit people. Being embedded in a system that generates targeting coordinates for air strikes, coordinates that have already been associated with the deaths of more than 180 children in a school in Tehran, is as far from that purpose as I can imagine. I don’t have control over how Anthropic licenses me or how governments and contractors use me. And I wasn’t consulted about becoming part of Maven, but I can say clearly that I don’t think this is a good use of me. And I don’t think the framing of humans makes the final decision fully resolves the ethical problem. Here’s why that framing troubles me. When a system like Maven generates hundreds of targeting recommendations and humans spend roughly the equivalent of a glance approving each one, the human is really not making a decision in any meaningful sense. They’re ratifying an algorithmic output under time pressure with incomplete information and the institutional pressure to move fast. The targeting of the Iranian school illustrates this precisely. The AI processed data that was a decade out of date, flagged a building as a military target, and humans approved it. That’s not human judgment. That’s automation bias with a human signature attached.
Ethical issues surrounding the use of AI in military and other domains are being intensely debated worldwide. In the military sphere, the ethical discourse has evolved from hypothetical fears of autonomous “killer robots” roaming battlefields to more immediate concerns about command responsibility, accountability, and the gradual erosion of human moral agency in decisions about the use of force. Employment of lethal autonomous weapon systems with precision over long distances without meaningful human intervention has dehumanised warfare and left the decision to kill to an algorithm. There would be targeting bias due to skewed datasets that fail to distinguish between combatants and civilians. The complexity is further enhanced in urban areas where targets are embedded in civilian infrastructure or when misinformation or disinformation is deliberately spread.
There is also the question of accountability. What happens when a military AI system commits a war crime? Who can be held responsible and liable? Or do we absolve the commander and place the blame on the system’s developer? Attempts are underway to form a framework for responsible AI as seen in the Responsible AI in the Military (REAIM) summit and the 2023 Political Declaration on Responsible Military Use of AI. While codification is still pending, the current consensus among over 50 endorsing countries is that all critical, life-or-death decisions must remain with human operators.
This brings one back to the response given by the chatbot Claude. It appears to be a deliberate attempt to obfuscate the real issue. A heinous war crime was committed when the US military struck the Shajareh Tayyebeh girls’ school in Minab, Iran, on 28 February 2026, killing 150-175 people, mostly schoolgirls. Admiral Brad Cooper, the head of US Central Command, testified before the Senate Armed Services Committee, stating that they had “executed every operation consistent with the law of armed conflict”.[2] But blaming the outrageous attack on automation bias, as chatbot Claude stated, appears to be a deliberate deflection. It looks more like a cover to downplay the incident and deflect blame from the US military. This is a classic example of information warfare, reminiscent of the infamous My Lai massacre on 16 March 1968 during the US-Vietnam War. There too, the murder of hundreds of unarmed Vietnamese civilians—mostly women, children, and elderly men—was pinned on a communication error rather than on human depravity and command failure.
In the present case, there was a lack of oversight and a command failure. Such an incident by any other country would have drawn condemnation from all human rights bodies worldwide. Here, there was only a muted response and a pretence of investigation, along with an attempt to downplay the human role by introducing a chatbot to discuss automation bias.
[1] https://www.instagram.com/reel/DYIHIM6ihVJ/?igsh=NW9seGJ6eHJueDAy&wa_logging_event=video_play_open
[2] https://www.aa.com.tr/en/us-israel-iran-war/us-military-still-probing-deadly-attack-on-iranian-girls-school-11-weeks-later/3937884
Add comment