Our approach to government and national security partnerships

OpenAI published a set of principles this week meant to govern how it works with governments and national security agencies. The company frames it as a balancing act: it wants in on lucrative defense and intelligence contracts, but it also wants to keep its public image as the responsible adult in the AI room. So the post talks about democratic accountability, public safety, and using AI to strengthen rather than undermine free societies.

The specifics are thin, as these things usually are. OpenAI says it will support the national security missions of democratic governments, work within legal and civilian-oversight frameworks, and avoid uses that would harm civilians or violate human rights. It also says it wants to help write the rules of the road internationally, presumably so that authoritarian states don't set the norms for military AI by default. None of this is enforceable in any real sense. It's a values statement, not a contract.

What's notable is the timing and direction of travel. OpenAI has already been moving toward government work more explicitly than it did a couple of years ago, and rivals like Anthropic and Google have made similar overtures toward defense and intelligence customers. The era of AI labs positioning themselves as strictly civilian, consumer-facing companies is fading. Once you build models capable of processing intelligence data, running logistics, or assisting with cyber operations, the Pentagon and its counterparts abroad come knocking, and turning that money away gets harder to justify to your own board.

The document also seems designed to preempt criticism. By publishing principles now, OpenAI can point back to them later when it signs deals that raise eyebrows, whether that's Israel, the US military, or some other government whose human rights record is, let's say, mixed. Whether the principles change any actual decision remains the open question, and it's the one the blog post conveniently doesn't answer.