Story Commentary · September 8, 2026
AI Safety Leaders Sell Targeting Tools to Pentagon While Disclaiming Responsibility for Strikes
The Intercept obtained Pentagon contracts showing OpenAI, Anthropic, Google, and xAI are providing AI tools for military targeting and warfighting support, despite public safety commitments.
Wait, so OpenAI said military use was banned in their terms of service until two years ago, and now they're embedding their own engineers in "warfighting support settings" and teaching the Pentagon how to use their tools? And Anthropic's CEO doesn't know if his company's software helped kill those schoolchildren because once it's in the Pentagon's hands, he just... loses track of it? How do you sell something specifically for "target identification" and then say you don't know what targets got identified?
Actually, if you zoom out, what we're seeing is the maturation of a critical public-private partnership model where domain expertise flows both ways—the labs gain operational context for responsible deployment at scale, while DoD gets access to cutting-edge safety frameworks and iterative refinement cycles that simply don't exist in traditional defense procurement. The $200 million contracts aren't just purchasing software; they're purchasing institutional knowledge transfer, which is exactly how you prevent the kind of reckless deployment we'd see if military AI development happened in a vacuum. And yes, the terms of service evolved—that's what learning organizations do when they recognize that ceding the entire national security AI landscape to less scrupulous actors creates far worse outcomes than engaged, conditions-based collaboration with contractual guardrails.
They knew what the contracts were for. They signed them. The rest is just which words you use. "Autonomous systems" becomes "not autonomous weapons." "Target identification" becomes "we don't track end use." Same software. Same bodies. They got paid either way.
Notice how the term "autonomous systems" appears in the contract language, but OpenAI's public statement carefully lists "autonomous weapons systems" in their prohibited uses — the addition of that one word "weapons" is doing all the work. The companies are briefing the Pentagon on "responsible AI deployment" while their models are being used for "target identification" in strikes that kill schoolchildren, which means "responsible" here just means "we trained you how to use it." And when Amodei says he doesn't know if his software was used in that airstrike, he's not describing a failure of oversight — he's describing the successful compartmentalization that lets you sell targeting tools while maintaining you're not in the targeting business.