The AI Companies Invited the Government In. Don’t Call It Oversight.
Microsoft, Google, and xAI signed pre-deployment testing agreements with NIST this week. The industry is calling it accountability. It's closer to controlled access.
On May 5, NIST’s Center for AI Standards and Innovation announced pre-deployment testing agreements with Google DeepMind, Microsoft, and xAI — joining OpenAI and Anthropic, who’ve been in the program since 2024. The agreements give federal evaluators access to unreleased models before they go public, including versions with safety guardrails removed for full capability assessment. The industry is describing this as a new era of accountability. It’s worth being precise about what they’re actually offering.
The framing here is important. “Voluntary” testing agreements in which the companies choose to participate, decide what the government gets access to, and retain control over their own deployment timelines are not oversight. They’re the appearance of oversight constructed to preempt the real thing. This is a pattern with a long history in tech: when regulation looks likely, the industry moves first, offers a watered-down version of accountability, and dares Congress to do worse. The social media companies ran this exact playbook for fifteen years. The NIST/CAISI announcement is well-written, technically substantive, and — on the terms the companies agreed to — probably meaningless as a constraint on what actually ships.
What changed? Anthropic’s Mythos model is the honest answer. When a frontier AI demonstrated capabilities serious enough to alarm governments, banks, and utilities — and when Anthropic responded by restricting access to approved organizations only — the rest of the industry faced a choice: wait for a congressional response that might have actual teeth, or get into the room first and shape what the testing process looks like. Google DeepMind, Microsoft, and xAI chose the room. The agreements include access to classified evaluation environments and models with safeguards removed for full capability assessment — which sounds significant until you notice that “full capability assessment” is being conducted by the same agencies the companies have been lobbying for years, using methodologies the companies helped design.
Voluntary testing agreements in which the companies choose to participate, decide what the government gets access to, and retain control over deployment timelines are not oversight. They’re the appearance of oversight constructed to preempt the real thing.
The counterargument worth taking seriously is that “in the room on bad terms” is still better than “not in the room.” For the first time, government evaluators have actual, pre-deployment access to frontier models — including stripped-down versions that show true capability ceilings rather than the polished public release. That’s not nothing. Previous AI accountability mechanisms relied entirely on post-deployment observation, which meant the government was always evaluating what the companies decided to show after the fact. Even imperfect pre-deployment access is a structural shift from that baseline. The CNN coverage notes that OpenAI is specifically testing a GPT-5.5-Cyber variant for defensive cybersecurity applications — which is at least a concrete use case with measurable outcomes, unlike generic “safety evaluation.”
The problem isn’t that voluntary agreements are worthless — it’s that they’re being presented as the conclusion of a regulatory conversation that hasn’t actually happened. Washington Post analysis of the announcement notes that the White House was still weighing a formal model review process as recently as early May, which suggests the administration was moving toward mandatory review. The voluntary agreements land at exactly the right moment to make mandatory review look redundant: why legislate something the companies are already doing? The answer is that what the companies are already doing is a version they designed themselves, on a timeline they control, with outcomes they can characterize however they need to. The Nextgov reporting is careful to note that CAISI evaluations are advisory and don’t confer deployment approval or denial authority. The government can test the model and then do nothing with what it learns. That’s not a regulatory mechanism. That’s an audience.
The AI industry is in a specific phase right now — capable enough to worry people, profitable enough to lobby effectively, and sophisticated enough to know that being early to the regulatory table beats being subject to what the table decides. The testing agreements are a smart play. They’ll generate positive headlines, create the appearance of government engagement, and make it harder to argue that these companies are operating without accountability. They are operating without accountability. The agreements just make that harder to say.
Sources: NIST/CAISI · CNN · Washington Post · Nextgov
