Featured
AI, Autonomous Warfare and the Severance from Responsibility
In autonomous warfare, AI may have found its ultimate application: Severing everyone from responsibility.
Last week, US Secretary of Defense Pete Hegseth announced the setup of a new taskforce to “study the future of warfare”. Members include Elon Musk, Palmer Luckey, the CEO of the defence tech company Anduril. The list of those invited to the group leaves little doubt what this is about. The “future” of warfare will be artificial intelligence.
While the public discourse has mostly focused on autonomous systems as means to execute mission (i.e. drones), another aspect has recently come to light: Target selection by AI.
This article will argue that such systems are inherently unethical due to AI not being able to be an ethical agent, and that with its involvement in autonomous warfare, AI finds its ultimate application: Where LLMs as black box decision-making machines are not a shortcoming, but a necessary feature to remove humans as far as possible from responsibility.
The AI-military complex
The title of the DoD working group is both misleading and a euphemism: Autonomous warfare is not a “future” to be evaluated but has already become an integral part of current conflicts. The choice of people appointed to the taskforce make it clear that not the “if”, but the “how” that is being “studied”.
In February 2026 US missiles attacked a school in Minab, Iran, killing 123 children. And while the US military has denied, deflected and given little concrete information on the incident, it brought to light the issues connected with AI systems used not to only guide missiles and drones to a target, but involving it in deciding about the target itself.
At the same time we also learned about the involvement of the “AI frontier labs” in the military. Defence contractor Palantir stated that the company “is not responsible for the underlying data nor identifying intelligence deficiencies and that there’s no evidence that its software was at fault in the Minab strike” (Bartenstein, Karra 2026), a statement that revealed more than it was likely intended to do.
Around the same time the US Department of Defense had its widely publicised “fallout” with “AI lab” Anthropic over the use of its services for military purposes. In brief, the DoD wanted to punish Anthropic by designating the company as a “supply-chain risk” for its alleged refusal to provide its AI services for certain applications.
However, we need to look carefully at how Anthropic choses its wording in the conflict: In a statement the company acknowledges that “fully autonomous weapons (those that take humans out of the loop entirely and automate selecting and engaging targets) may prove critical for our national defense”, justifying its refusal with the statement that current “frontier AI systems are simply not reliable enough to power fully autonomous weapons” (Anthropic 2026). A technical rather than an ethical concern.
When pressured, OpenAI later tried to have it both ways (OpenAI 2026), There are no known comments from SpaceX AI on that question, something that we have to assume factored into the decision to select Elon Musk.
The history of diffusion
For most of human history, soldiers have committed atrocious crimes during wartime: Variably justified with loyalty to a crown, nationalism or perceived threat. However, for millennia war took place far from the eye of the public, and without transnational agreements on the conduct of war.
Then the situation changed: The information age and the rise of modern democracy brought scrutiny and demands for transparency. The modern understanding as the military as part of civil society brought additional guardrails and raised the question of responsibility.
With this came new methods to diffuse responsibility for acts of violence ordered by governments: Technical progress in weaponry allowed for remotely operated deadly weapons, remotely guided missiles and loitering munition, Hence creating an environment “moral disengagement” within a large military operation that allowed to put a distance between acts and moral agents.
Moral disengagement “operates by obscuring, or minimizing the agentive role in the harm one causes. People will behave in ways they normally repudiate if a legitimate authority accepts responsibility for the effects of their conduct […] Selective activation and disengagement of personal control permits different types of conduct by persons with the same moral standards under different circumstances“ (Bandura 1999).
The question of individual responsibility however never went away until very recently, when automated-decision entered the chain of command under the disguise of “efficiency”.
“The concept of moral agency, central to traditional ethical frameworks, becomes problematic when applied to autonomous systems that lack consciousness, yet make consequential choices.” (Hamad 2025)
Automated decision-making as the ultimate diffusion
When people argue about the ethics of automated-decision making, the pro-AI arguments that are often made are efficiency and even fairness, claiming that humans are biased, too, and therefore machines may even provide a fairer approach, not being influenced by human moods and feelings.
Those arguments fall short on several levels: While humans are certainly show bias, algorithms are biased on multiple levels: First the bias that went into implementing the system and training, secondly at the stage of application, and through the lack of ability of taking unforeseen parameters into account.
The pro-AI argument also ignores one of the central aspects of human decision-making: That human decision-making involves a semantic understanding (the ability of a system or person to grasp the meaning, context, and intent of language).
“…ethical illegitimacy of AWS stems from their incapacity to satisfy the foundational requirements of jus in bello. Just War Theory demands not only adherence to proportionality and distinction but the presence of moral agents capable of making contextually informed value judgements — a capacity algorithms fundamentally lack due to their inability to comprehend meaning” (Guo 2025)
Modern LLM attempt to mimic semantic context by “reasoning”, however not only is the motivation to implement this inherently dishonest (as an attempt to further anthropomorphize), it has also shown that those attempts “face a complete accuracy collapse beyond certain complexities” and have been correctly described as an “illusion” (Shojaee et al. 2025).
“Reasoning-models” are a hocus-pocus of making models “look” like they go through take human-like decision process. Even Anthropic had to admit as much: “But our research shows that we can’t always rely on what they tell us about their reasoning”(Anthropic 2025).
Get Wolfgang Hauptfleisch’s stories in your inbox
Join Medium for free to get updates from this writer.
Despite this, managers from Anthropic have been reportedly lobbying religious scholars in privately hosted meetings to acknowledge the potential that LLMs can have “moral standing” (Dias 2026) and both the CEOs of Anthropic and OpenAI have repeatedly implied to assign LLMs “quasi consciousness”. It is this anthropomorphisation of AI systems creates the illusion that, while fallible, AI can make “moral” decisions.
Severance
The black box of AI systems provides the severance of ethical considerations from decision-making the military has wished to establish for centuries. In this context, the opaqueness of automated decision-making is not a shortcoming, but the required feature.
The “perfect” AI decision-making system is functionally a black box with a certain randomisation and the outward appearance of reasoning. The users of the system might even “believe” in it “moral standing”, in the way the audience of a soothsayer wants to believe. This self-delusion might be the final step on achieving total severance from responsibility.
Ethical implications for society
The establishment of automated-decision making systems as described has implications for the wider society.
Because of their military nature, we do not know what those systems are trained on. Considering the enormous appetite of modern LLMs for data to serve their purpose for “target selection” and “execution”, we have to assume that “our date” has become part of the training corpus. It is not unreasonable to assume that data from social media, network graphs and blog posts have whatever small influence on a system that makes decisions over life and death.
In addition, those who work at the AI companies involved need to ask themselves to what extent they can diffuse their own responsibility.
It is questionable if the companies who develop the current “frontier models” have the ethical standing or even the seriousness that would allow them to take on such responsibility.
Proliferation
There is the additional danger that, if autonomous systems in the military context become widely accepted as “ethical”, this attitude will proliferate into other areas of society. Police and private security will certainly share arguments with the military.
It is often assumed that automated-decision making will slowly proliferate from trivial applications to more series onces that make decisions over people that could do real harm. It is however also possible that once autonomous decision-making becomes accepted for military purposes, the argument will be turned around: If accepted to make life and death decisions, why deny such systems to make decisions in other areas.
Conclusion
It is a central aspect of current ethical frameworks to recommend that to prevent bias those affected by automated-systems should be included in the development. This creates an obvious absurdity for autonomous systems in the military and a demand that can not be fulfilled.
The increasing anthropomorphisation of AI is more than a layer of user experience, it is a way to “humanise” an inherently unethical system.
With the application of AI in autonomous warfare, the AI industry has found its ultimate product: A service that makes society dependent on them while it isolates them and the users from any responsibility and shields them from ethical considerations.
The challenge, and responsibility, for AI ethics is to point out the absurdities created by automated decision-making.
For more, subscribe to the “misaligned bits” newsletter and follow Misaligned on Medium, LinkedIn, Threads or Mastodon.
Guo, J. (2025). The ethical legitimacy of autonomous Weapons systems: reconfiguring war accountability in the age of artificial Intelligence. Ethics & Global Politics, 18(3), 27–39. https://doi.org/10.1080/16544951.2025.2540131
Elisabeth Dias (2026). Religious Scholars Met With Anthropic. What They Heard Stunned Them. New York Times
Bartenstein B, Karra K. (2026). Inside the “US Kill-Chain” that Destroyed An Iranian School. Bloomberg
Badreldin Hamad (2025). The Impact of Digital Technology on International Humanitarian Law: Ethical and Legal Implications of Autonomous Weapons Systems. (2025). European Journal of Law and Political Science, 4(4), 1–14. https://doi.org/10.24018/ejpolitics.2025.4.4.182
Bandura, A. (1999). Moral disengagement in the perpetration of inhumanities. Personality and Social Psychology Review: An Official Journal of the Society for Personality and Social Psychology, Inc, 3(3), 193–209. doi:10.1207/s15327957pspr0303_3
Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S., Farajtabar, M. (2025). The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. arXiv preprint arXiv:2506.06941.
James O’Donnell (2026). A defense official reveals how AI chatbots could be used for targeting decisions. MIT Technology Review
Anthropic (2026). Statement from Dario Amodei on our discussions with the Department of War
Nick Robins-Early (2026). Pete Hegseth appoints Elon Musk to US taskforce on future of warfare. The Guardian
Anthropic (2025). Reasoning models don’t always say what they think
OpenAI (2026). Our Agreement With the Department of War
