The order that bans AI licensing also tells agencies to pick the trusted partners
More than two dozen groups asked the White House to publish its frontier model review framework, completed on 1 August and previewed only with selected companies.
On 9 September, more than two dozen advocacy groups from both political wings asked the White House to publish the voluntary framework it uses to review frontier AI models before release. The administration finished the framework by its own 1 August deadline, previewed it with a selected group of technology companies, and has not published it. Five weeks on, it is still unpublished. Smaller and open-source developers were largely left out of those conversations.
The same day, The Information's Amir Efrati described the practical result.
🤔 There is a de facto “whitelist” of companies the White House permits to use SOTA closed source AI models from Anthropic & OpenAI …
But nobody seems to know how to get on it t.co/lNoPogfMkB
A whitelist nobody knows how to join is a strange thing to find in a country whose executive order on the subject explicitly forbids licensing. And you can read both halves of that in one document.
Executive Order 14409, signed on 2 June, directs agencies to design a voluntary framework under which developers give the federal government access to "covered frontier models" for up to 30 days before release, and — this is subsection (b)(iii) — "collaborate with the Federal Government to select trusted partners that will have early access to covered frontier models". Two subsections later: "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement."
Both sentences are in the order.
The list is authorised in one and disclaimed in the next. So which is it?
CNBC reported in July that the administration had begun deciding which companies and entities get access to the newest models, a decision that used to belong to Anthropic and OpenAI. A White House official told CNBC it "doesn't provide approvals for AI releases from private companies" and that "decisions on timing and scope of releases rest entirely with the companies". In the same period, per CNBC, the administration blocked two frontier models over national security concerns and restored access after weeks of negotiation, and OpenAI said it would limit new models to "trusted partners" to comply with government requests.
Our read: treat this as industrial policy, because the economics behave like it. A pre-deployment review is a compliance cost. A compliance cost you can read is something a small team can probably build for.
One you cannot read is a lottery. And lotteries favour whoever already has counsel in the room. The coalition — Americans for Prosperity next to Public Citizen, which tells you how unusual the alignment is — makes exactly that argument: secrecy entrenches the largest labs while open-source developers work blind.
The strongest case for the other side has not been made publicly. So we will make it. Publishing what a security review checks tells adversaries what it does not check. Evaluation criteria are withheld for that reason across government, and a leaked list of the specific cyber capabilities being probed would be genuinely useful to the people you least want holding it. But parameters are not test content. You can publish who qualifies, on what timeline, with what appeal, without publishing the exam.
We would bet the framework becomes public through Protect Democracy's freedom-of-information suit rather than a White House release, and not before 2027. A voluntary publication this year would probably mean we misjudged the politics.
If you are buying frontier capability for a company, there is a duller question in here that your procurement team should be asking. Is our access to this model contingent on anything we are not allowed to see, and what happens to the roadmap if the answer changes?