Remember December 11, 2025? An executive order landed with a title that promised a lot: “Ensuring a National Policy Framework for Artificial Intelligence.” The framework itself was not attached. It was a promissory note. Washington told us the homework was coming.
The homework arrived on March 23, 2026. Sort of. The White House released its National AI Policy Framework, establishing federal guidelines for AI development and deployment, mandating safety testing and federal oversight. What it did not release, and still has not released, is the part that actually matters: how the testing works.
A test with no answer key
I review AI tools for a living. My whole job is checking whether a product does what its marketing claims. The single most useful thing any vendor can hand me is a methodology, because a score without a method is just a vibe with a number attached.
So when a federal framework mandates safety testing and then declines to say what the tests are, who administers them, what counts as passing, or what happens on a failure, I have no way to evaluate it. Neither do you. Neither, I suspect, do most of the companies who will eventually be graded by it.
This is not a small gap. Safety testing for frontier models is a genuinely hard technical problem with no settled answer. Different evaluation suites disagree with each other. Results shift depending on prompt phrasing, temperature settings, and whether the model has been fine-tuned since the last check. Whoever writes the specifics is making enormously consequential calls about what “safe” means in practice. Right now those calls are happening somewhere we cannot see.
The pre-release review question
Reporting since the framework dropped suggests the White House is weighing pre-release reviews for high-risk models, and exploring an executive order that would set up a vetting process before models ship. That is a much bigger deal than the framework language alone implies.
Pre-release review changes the shape of the industry. It means a government body sits between a finished model and its users. Depending on how it is built, that could be a light-touch registration step or a genuine approval gate with the power to say no. Those are wildly different regimes for anyone building on top of these models, and the difference determines whether a small lab can realistically ship anything at all.
Consider what a review process needs to work: technical staff who understand frontier training runs, secure infrastructure to handle model weights, turnaround times that do not make a six-month release cycle into an eighteen-month one, and clear criteria so decisions are not arbitrary. Any one of those being weak turns the whole thing into theater or a bottleneck. Possibly both.
Even the coverage cannot agree
Here is my favorite symptom of how thin the public record is. Read the secondary coverage of this framework and outlets do not consistently agree on basic framing. Some describe it as a Biden-Harris roadmap. Others describe Trump’s federal AI policy framework aimed at fostering new development. Same document, same month, two different stories about whose policy it is and what it is for.
When trade press cannot settle on the authorship and intent of a major federal framework, that is not sloppy journalism. That is what happens when everyone is working from a summary instead of a text.
Why I care as a reviewer
My readers ask a practical question: can I use this tool for real work? Increasingly, part of that answer depends on whether the model underneath is going to be available in six months, whether its capabilities will be trimmed for compliance, and whether the vendor’s safety claims mean anything verifiable.
An undisclosed testing standard makes all three unanswerable. Vendors will start claiming compliance. I will have no way to check the claim, because the criteria are not public. That creates exactly the condition I spend most of my time fighting: marketing language that sounds like a technical assertion and cannot be falsified.
The most likely outcome is a new certification badge on landing pages, meaning something
What would actually help
The fix is boring and obvious. Publish the evaluation criteria. Publish the thresholds. Publish who conducts the tests and how disputes get resolved. Let independent researchers reproduce the results on open models so we can calibrate what the scores mean.
None of that requires exposing sensitive model internals or classified threat intelligence. Financial auditing standards are public. Drug trial protocols are public. Crash test methodology is public, which is precisely why a five-star rating means something to a car buyer.
A safety regime built on secret criteria asks for trust it has not earned yet. Mandated testing is a reasonable idea. Mandated testing that nobody outside the room can inspect is a brand, not a standard, and I am not going to review it as one until somebody shows me the rubric.
🕒 Published: