The Military's AI Alignment Problem
The US Department of Defense has been on an aggressive AI adoption trajectory since the establishment of the Joint Artificial Intelligence Center in 2018 and its successor, the Chief Digital and Artificial Intelligence Office in 2022. But as the Pentagon has signed contracts with a growing array of AI companies — from established defense contractors to Silicon Valley startups to frontier model labs — a new challenge has emerged: the proliferation of incompatible AI tools, different model versions, and inconsistent capability levels across commands and branches of the military.
The Pentagon's response is an effort to standardize its AI providers and get all of them operating on what officials are calling the same baseline — a shared set of model versions, APIs, safety policies, and performance benchmarks that would allow different military systems to interoperate and allow oversight officials to have a coherent picture of what AI capabilities are deployed where.
Claude in the Military Context
Among the AI systems the Pentagon is working to standardize access to is Anthropic's Claude — one of the frontier large language models that has been used by various defense-adjacent contractors and directly by DoD organizations for tasks including document analysis, intelligence processing, and decision support. The inclusion of Claude alongside systems from OpenAI, Google, and Microsoft reflects the Pentagon's desire to maintain optionality across multiple frontier AI providers rather than becoming dependent on any single vendor.
Anthropic has had a more complex relationship with military contracting than some of its competitors. The company's Acceptable Use Policy prohibits use of Claude for weapons development, attacks on critical infrastructure, and several other categories of harmful application. But the line between prohibited and permitted use cases in a military context is not always clear — document analysis, logistics optimization, and decision support are facially neutral applications that can in practice contribute to military operations including lethal ones.
The resignation of OpenAI's robotics head, Caitlin Kalinowski, over that company's Pentagon deal — citing concerns about lethal autonomy and insufficient deliberation — underscores the ethical tensions that frontier AI companies face as they engage with military customers. Anthropic has not publicly disclosed the specific terms under which Claude is available to DoD customers or what oversight mechanisms govern its military use.
AI for Military Targeting: The Core Issue
The most sensitive application of AI in military contexts is targeting — using automated systems to identify, prioritize, and in some cases help engage enemy targets. The US military's doctrine requires human authorization for lethal force decisions, but the role of AI in processing sensor data, identifying potential targets, and presenting recommendations to human decision-makers has expanded dramatically in recent years.
Project Maven, the Pentagon's AI-based imagery analysis program, processes drone surveillance footage to identify targets of interest for human review. The AI does not make the final lethal decision — a human does — but it shapes what information reaches the human and how it is framed, which raises questions about the meaningful quality of the human oversight that doctrine nominally requires.
The standardization effort is in part a response to oversight concerns: if different AI systems are giving different outputs for the same inputs, or if no one has a complete accounting of which AI systems are contributing to which decisions, meaningful human oversight of AI-assisted targeting becomes extremely difficult. Standardization creates auditability — the ability to know which model, at which version, under which safety settings, was involved in any given decision chain.
The Ethical Architecture Question
Critics in the AI ethics community argue that standardization addresses a governance problem but does not resolve the more fundamental ethical question: whether AI should be used in targeting decisions at all, and if so under what constraints. Getting all military AI providers on the same baseline means defining what that baseline is — what values, what safety policies, what prohibited uses are baked into the standard. That definitional process is happening largely out of public view.
The stakes of getting it right are high. AI systems that introduce bias into targeting decisions, that hallucinate threat assessments, or that are manipulated by adversarial inputs could contribute to catastrophic errors in combat situations. The Pentagon's standardization push is an acknowledgment of this risk — but the adequacy of the response will ultimately be judged by how the systems perform under operational conditions that cannot be fully anticipated in advance.
This article is based on reporting by MIT Technology Review. Read the original article.
