The federal government wants AI everywhere, fast. It is buying it like it’s a Black Friday deal, while the security oversight program meant to keep vendors honest is being starved into paperwork cosplay.
This is not a story about whether AI is “good.” It’s a story about procurement gravity. Once you plug a vendor into mission workflows, the exit costs become a hostage note, and the “discount” was just the onboarding fee.
What happened
ProPublica published a set of cautionary lessons about the U.S. government’s rush into AI, drawing parallels to the earlier cloud transition. The piece highlights a familiar pattern: vendors offer “freebies” or steep discounts to get embedded, oversight programs like FedRAMP struggle with limited resources, and “independent” assessors are paid by the companies they evaluate.
One concrete example: the General Services Administration has promoted government-friendly pricing for AI tools, with per-seat offers that look cheap enough to make adoption feel inevitable. ProPublica’s warning is that the initial price is not the cost. The cost is lock-in, usage-based billing, and the slow conversion of procurement convenience into operational dependency.
The non-obvious part: AI risk is being reorganized into a billing model
Everyone argues about model alignment. Meanwhile, the government is creating a new alignment problem by contract: if you buy AI as a metered service, the incentives get weird. Agencies are told to “monitor usage,” but usage is the point. Savings narratives push adoption. Adoption pushes spend. Spend creates dependency. Dependency makes audits politically expensive.
In other words, the “AI safety” fight inside government is going to be won or lost by procurement clauses, not by an ethics panel with a PowerPoint.
Why FedRAMP matters even when it’s boring
FedRAMP was created to prevent a chaotic cloud free-for-all. ProPublica reports that it has been reduced to minimal staffing and is leaning heavily on third-party assessors. If that’s right, then AI procurement is inheriting the worst of the cloud era: compliance theatre, vendor pressure, and an approvals pipeline that can be worn down until “yes” becomes the path of least resistance.
And AI makes it worse, because the blast radius of a bad purchase is larger. You are not just hosting data. You are piping sensitive context into an active system that will be asked to summarize, decide, and automate.
The Singularity Soup Take
The government is buying “$1 AI” the way people buy cheap printers: the real money shows up later, in cartridges, maintenance, and the quiet realization that you built your workflow around the thing. The discount is the hook. The hook becomes policy. Efficiency in humiliation, fully automated.
What to Watch
- Procurement language that shifts liability: who eats the cost when an AI tool causes harm, leaks data, or just hallucinates its way into a decision?
- Whether FedRAMP is rebuilt as a real security gate, or left as a “check the boxes” accelerant.
- Any major agency pullbacks after early adoption, especially where usage-based billing explodes (the cloud pattern repeating, but faster).
Sources
ProPublica — "The Federal Government Is Rushing Toward AI. Our Reporting Offers Three Cautionary Tales."
General Services Administration — "Buy AI"