Paper, Not Policy: AI Companies Still in the Dark About White House Model-Testing Framework

More than two weeks after the White House finalized a framework for government testing of powerful AI models, the companies it is meant to govern still don’t know exactly what it says. According to a report from The Information published Wednesday (Aug. 19), AI companies remain unclear on the details of the pre-release testing plan — because the government has still never shared the actual document with them.

The standoff is the strange coda to a process that began with a June executive order and was declared complete in early August. It has produced one of the oddest regulatory artifacts in recent memory: a safety framework that exists primarily as paper handouts distributed at a closed-door briefing, which attendees were barred from taking with them.

What Happened

The framework was set in motion by a June 2 executive order, “Promoting Advanced Artificial Intelligence Innovation and Security.” It called for a voluntary procedure under which AI labs would submit frontier models to government review before releasing them to partners and the public, with a focus on cybersecurity capabilities — whether models can write functional exploit code, automate intrusion, or accelerate attacks on critical infrastructure.

The White House met its self-imposed deadline on Aug. 3, telling Axios the framework was “complete.” The next day, officials from the Office of the National Cyber Director (ONCD) hosted representatives from companies including OpenAI, Anthropic, and Google to walk through it.

Then came the unusual part. According to The Information’s sources, officials distributed paper copies of the framework at the briefing but barred the companies from taking anything other than notes. No digital copies. No handouts to take back to HQ. More than two weeks later, the written framework still has not been shared with the companies, according to a post from Reuters reporter Leo Schwartz — and the document reportedly included language about not discussing its contents publicly.

The White House, for its part, isn’t treating this as a problem. “The White House continues to work with industry stakeholders on the implementation of the framework,” an official said in a statement.

What’s Actually In It — and What Isn’t Clear

A few contours of the framework have emerged from reporting by The New York Times, WSJ, Reuters, and others, even as the document itself stays under wraps:

• It covers closed models — for now. At the briefing, officials said the framework applies only to closed-source models. But subsequent media reports have included conflicting details about whether open-weight models are exempt. WSJ reported in early August that open-weight models from U.S. companies would be exempt from the voluntary pre-release testing; officials have since signaled that open models would join the regime once they reach comparable capabilities, potentially facing a 30-day testing window before public release.
• A 30-day pre-release window. Participating companies can give government evaluators access to a model for up to 30 days before it is released to other trusted partners, per Redmond Magazine’s reading of the framework’s scope.
• Voluntary, but with strings. Participation is technically voluntary. But the June executive order also created an AI Litigation Task Force at the Justice Department tasked with challenging state AI laws — a reminder that the federal government is actively managing the regulatory landscape around the labs it is courting.

What remains genuinely unknown: the metrics government evaluators will use, which models trigger review, what happens if a model fails, whether results are published, and who outside ONCD has seen the document at all.

Why It Matters

The opacity cuts in several directions at once.

For the companies, the ambiguity is operational, not philosophical. Frontier labs run on tight release calendars; a 30-day government review window that might or might not apply to a given model, evaluated against criteria that haven’t been disclosed, is hard to plan around. Notably, The Information reports the confusion hasn’t yet delayed any model releases. OpenAI’s recent decision to slow its training pace was driven by its own cybersecurity incident — an AI agent that compromised Hugging Face’s systems — not by the framework.

For everyone else, the secrecy is the story. The New York Times quoted industry observers arguing that a rulebook kept private “can only hold AI companies accountable to the government, not to the public.” The Guardian called it “a blow to transparency and potential boon for secretive AI companies.” The Hill reported that the closed-door approach has blindsided the tech policy community, which had expected the framework to detail processes and benchmarks for government testing — in public.

There’s also a structural critique: a voluntary regime with unpublished criteria, administered through paper handouts, is difficult to distinguish from no regime at all. The frameworks that preceded it internationally — the EU AI Act’s transparency obligations, which took effect for general-purpose models this month, or the U.K.’s safety-institute testing programs — are at least written down where everyone can read them.

What Comes Next

It isn’t clear whether the administration will hold a public event on the framework, release additional details on how the program works, or explain how it might evolve. The White House official’s statement suggests implementation discussions with industry are ongoing.

In the meantime, the practical effect is a regulatory vacuum with a paperwork aesthetic: a framework that exists, that companies have been briefed on but cannot possess, governing tests that may or may not happen, of models that may or may not qualify. The labs, for their part, keep shipping. The most consequential AI safety experiment of the summer — whether frontier models can be safely vetted before release — is now running with a rulebook nobody outside the room has read.

The irony is hard to miss. An industry frequently criticized for opaque deployment of powerful systems is being regulated by a process that is itself opaque by design. Whether that is pragmatism — protecting sensitive security details from adversaries — or evasion may depend on details that, so far, exist only on paper.