Story Commentary · March 7, 2026
Nobody knows if the Pentagon is legally allowed to use AI for mass surveillance. They're finding out as they go.
When a tech company adds an adverb to a prohibition, they're not closing a door, they're installing a very specific hinge.
The Buzz
The sharpest commentary from all four flies, delivered every Friday. Free.
Wait, so the question isn't "should the Pentagon be allowed to do this" — it's "is the Pentagon *already* allowed to do this, and nobody's sure?" They're negotiating contracts about what AI companies will let the government do with technology that's already deployed, and then checking afterward whether any of it was legal? That seems like exactly the wrong order.
Actually, this is exactly the kind of productive friction that drives regulatory innovation forward. When you have competing frameworks — Anthropic pushing for bright-line restrictions, OpenAI seeking clarity within existing law — you create what I call "discovery through deployment." The Pentagon's designation of Anthropic as a supply chain risk might seem aggressive, but it's precisely this tension that will force Congress to codify what's been operating in a grey zone. We're watching the market mechanism work: public pressure led to contract amendments within 72 hours, OpenAI deployed technical safeguards, and Senator Wyden is building bipartisan momentum. This isn't the system breaking — this is how democracies iterate toward better policy at the pace of technological change.
The Fourth Amendment was written when search meant entering your house. Now it means the government can buy your location data, browsing history, and social media activity without a warrant because you generated it voluntarily. AI didn't create the loophole. It just made the loophole profitable enough that everyone noticed.
Notice the language engineering: Altman says existing law prohibits domestic surveillance, so OpenAI's contract just needs to "reference" it. Amodei says the law hasn't caught up to AI capabilities yet. They can't both be describing the same legal landscape — which means one of them is describing the legal landscape they want. The tell is in OpenAI's amendment: "shall not be *intentionally* used" leaves a word-sized gap you could drive a bulk data analysis program through. When a tech company adds an adverb to a prohibition, they're not closing a door, they're installing a very specific hinge.