I read the leaked draft at 2 a.m., and the room felt colder than it should. You can almost hear lobbyists recalibrating their scripts. What looked like a hard line has thinned into polite language and a wink.
A fluorescent copy of the draft sat open on an aide’s laptop — the order split itself into two tracks
Axios reported the latest leaks, and the picture is now less dramatic than the first whispers suggested. Instead of a single gatekeeper agency forcing a full review of every new model, the order appears to carve the work into a cybersecurity track and a frontier models track. The first arm strengthens federal defenses around networks and critical infrastructure; the second offers a voluntary check-in for makers of the most advanced AI systems.
The West Wing conversation started with a slide titled “covered frontier models” — then stalled
At one meeting an aide flagged the phrase and someone joked that the government wanted a peek at the biggest brains in the room. The draft now gives developers a 90-day window to ask for a vetting session. If companies accept, there’s a review; if they don’t, the draft leaves them to their own devices.
Will companies have to share AI models with the government?
Short answer: not according to the latest leaks. The order, as reported by Axios, appears to make that sharing voluntary. That will matter a lot to OpenAI, Google DeepMind, Anthropic, Meta (Llama), Microsoft and other heavy hitters whose business models hinge on model secrecy and speed to market. If compliance is optional, only reputational or market pressure—not the law—will move many firms.
An engineer at a startup texted: “Why would I invite a regulator in?” — the market calculus is simple
Business logic cuts both ways. A voluntary check gives firms a placard to show customers and partners that they cooperated. But it also hands competitors a blueprint for avoiding oversight: no one wants to be the first to expose proprietary weights, training data or safety testing. The result looks less like a dam and more like a gated cul-de-sac.
What does ‘covered frontier models’ actually mean?
The phrase isn’t technically precise. Leaks suggest it targets models whose capabilities exceed a threshold tied to social risk, scale, or compute. That could sweep in multi-modal systems like GPT-4-style models, Anthropic’s Claude, Google’s Gemini, and Llama-class projects. But without a single, transparent metric, companies will argue about definitions and claim ambiguity when they decline to participate.
A cybersecurity adviser left a sticky note on the draft: “harden, don’t police” — priorities have shifted
That sticky-note spirit shows in the draft’s split personality. The cybersecurity side feels like defensive engineering: tougher authentication, incident reporting, and federal hardening programs aimed at infrastructure. The other part—public-facing engagement with frontier model creators—reads like an invitation, not a summons.
I’ve seen Washington do about-faces before: an initially muscular announcement that becomes a compromise pieced together by agencies, lawmakers, and industry. The current language gives White House officials political cover while leaving blunt instruments—mandates, export controls, or licensing—on the shelf.
A lobbyist laughed and said, “Voluntary is the new mandatory” — gamesmanship will start now
Expect a parade of voluntary codes, safety councils, and third-party auditors. Companies will tout compliance badges and certifications. Think of this as a theater of assurance: a velvet glove over an iron fist is promised, but the glove might stay in the prop box. That metaphor is deliberate—public comfort handed out in neat, branded packets.
Axios’s sourcing points to infighting inside the Administration. Vice President J.D. Vance’s comment at last year’s AI Action Summit in France signaled skepticism toward strict rules; the March policy document hinted at a light-touch approach. Now, drafts suggest the White House is trying to square those instincts with rising public and congressional alarm about misuse, job disruption, and national-security risks.
A handful of engineers at OpenAI and Google scanned the memo — they’ll watch incentives, not rhetoric
For developers the question is simple: does cooperation buy anything? A voluntary review could offer a fast-path to government contracts, liability protections, or preferential procurement—if the order ties benefits to participation. Without incentives, participation will likely be patchy. And patchy oversight is a poor substitute for consistent rules.
Is the framework enforceable if it’s voluntary?
The short technical truth: enforcement requires mandate. Voluntary programs can be nudged by funding, procurement, or public naming, but they cannot force code disclosure or model weights. That means tools like export controls, liability regimes, and criminal statutes remain the real levers if the government wants strict compliance.
I’m not suggesting the drafts are final. Leaks are drafty things designed to shift negotiating positions. But you should feel the momentum: an Administration that once seemed ready to shrug at AI is now quietly designing a menu of options—some protective, some polite—and asking firms whether they’ll take the entrée.
Two metaphors, no more: think of this moment as a fragile bridge and a slowly closing curtain.
So what happens next? Will companies accept a voluntary check-in, or will they treat it as a PR checkbox while keeping the keys to the kingdom hidden?