OpenAI Signed With the Pentagon While Anthropic Drew a Line — and Now the Industry Has to Choose a Side

When the Pentagon went looking for AI partners, it found two very different answers. OpenAI signed. Anthropic sent lawyers. The resulting gap — between a company that agreed to work with the Department of War and one that sued the US government over AI safety boundaries — has become the sharpest fault line in a conversation that has been building toward exactly this kind of confrontation. The Verge framed it as plainly as any headline this week: "Anthropic doesn't want its AI killing people unsupervised. The Pentagon isn't happy."

The specifics of what OpenAI actually agreed to remain genuinely unclear, and that ambiguity is doing serious work in the conversation. Coverage from Built In noted how different the OpenAI contract looks from Anthropic's terms — but without a public accounting of what either company actually permits, the comparison stays frustratingly abstract. What's filling that vacuum is a wave of analysis from institutions like Stanford HAI asking the question that nobody in Washington seems eager to answer: who actually decides how America uses AI in war? The accountability question isn't hypothetical anymore. Project Maven is already selecting bomb targets in Iran, and the governance infrastructure around that capability remains, as TNGlobal put it this week, a "governance gap."

The handoff of DoD's AI weapons programs from Dario Amodei to Sam Altman attracted less outrage than it deserved — a quiet reshuffling that would have been front-page news in a different news cycle. What's interesting is where the alarm is actually registering. It's not primarily on Reddit or X. Bluesky has been running consistently negative on military AI for weeks, while arXiv researchers are publishing in a noticeably more optimistic register — papers on AI enforcing the Biological Weapons Convention, analyses of AI accelerating defense acquisition, assessments of autonomous systems for Taiwan's defense posture. The gap between those two conversations is not a matter of disagreement about facts. It's a disagreement about who gets to define the terms: researchers embedded in defense institutions, or civilians watching from the outside.

The most caustic framing in the current conversation comes from Foreign Policy in Focus, which declared flatly that we've entered "a Golden Age for War Profiteers." That piece, and others like it, treat the OpenAI deal not as a policy question but as a moral one — and they're speaking to an audience that increasingly agrees. A Substack post arguing that "the information space around military AI is being weaponized against us" captured something real about the epistemic situation: the companies building these systems have become the primary sources of public knowledge about what those systems can and cannot do. A piece at opiniojuris.org this week named this directly, describing tech companies' "claims of epistemic authority on military AI" as a form of power that deserves its own scrutiny.

The AI ethics conversation and the military AI conversation used to run on parallel tracks. They don't anymore. Anthropic's decision to draw a public line — and to litigate it — has made the question of supervised versus unsupervised lethal AI into an ethics debate with real institutional stakes. The War on the Rocks argument that "warfighters, not engineers, decide what AI can be trusted" is a direct rebuke to that position, and it's getting serious traction in defense circles. Both arguments will intensify as the contracts get larger and the systems get more autonomous. The question isn't whether civilian oversight of military AI is possible — it's whether any company will pay the commercial price of demanding it.