Surface Story
The White House is exploring a federal pre-release vetting process for AI models, requiring government review before public deployment. The proposal would establish standards for what counts as 'safe enough' AI before it reaches consumers and businesses.
Underneath the Story
A vetting regime controlled by the executive branch does not constrain AI power - it consolidates it, handing whoever sits in the White House the authority to greenlight or kill any AI system before the public ever touches it. The question of who sets the safety standard is inseparable from the question of who benefits from the answer, and the companies already embedded in Washington are positioned to write those standards in their own image. This proposal arrives precisely as American AI labs are racing to deploy frontier models, which means a federal gate does not slow the race - it just determines who controls the finish line.
Whoever defines AI safety gets to decide which AI wins.
Every regulatory regime produces two things: rules and gatekeepers. The rules are debatable. The gatekeepers are the point.
The White House proposal to vet AI models before public release is being framed as a safety measure. That framing is doing a lot of work. Safety from what, exactly, measured by whom, using which methodology, appealed through which process? None of those questions have answers yet. But the entity that answers them will hold more structural power over the AI industry than any board, any investor, or arguably any CEO.
That entity would be the federal government, specifically the executive branch, specifically whatever administration happens to occupy the White House when the standards get written.
The history of industry regulation in the United States follows a consistent arc. A sector grows large enough to attract political attention. Regulatory frameworks get proposed. The largest incumbents, the ones with Washington offices, lobbyists, and former officials on their payroll, engage early and shape the standards. By the time the rules are final, they function less as constraints on the industry than as barriers to everyone trying to enter it.
AI is moving through this arc in compressed time.
OpenAI, Google DeepMind, Anthropic, and Microsoft are already present in Washington. They have submitted testimony, published safety frameworks, and hired policy staff. They are not waiting to see what the standards will be. They are participating in writing them. A pre-release vetting process that draws on existing safety literature will draw heavily on safety literature these companies produced.
That is not a conspiracy. It is just how regulatory capture works. It does not require bad faith. It requires incumbency.
This proposal surfaces at a specific moment. Several frontier AI labs are preparing to release models of substantial capability. The regulatory window, the period between 'AI is powerful' and 'AI is everywhere,' is closing. Once capable AI is widely deployed and integrated into critical infrastructure, pre-release vetting becomes logistically absurd. You cannot gate what is already inside the walls.
So the proposal arrives now, framed as precaution. But the mechanism being built, a federal checkpoint on AI deployment, would be extraordinarily durable. Checkpoints do not dissolve when the emergency passes. They expand.
Consider what the checkpoint actually controls: which models reach users, which companies can deploy at scale, which open-source projects can publish weights, which foreign models can enter the American market. A vetting regime is simultaneously a safety tool and a trade policy instrument and a domestic industrial policy mechanism. Those three functions will not stay separated.
Pre-release vetting as currently imagined maps neatly onto the large-lab model of AI development, where a single identifiable entity produces a model and releases it. It maps very badly onto open-source AI development, where models are released as weights, modified by thousands of independent actors, and deployed without any central point of control.
Meta's Llama models, Mistral's releases, the entire Hugging Face ecosystem - none of these fit a review-then-release framework without restructuring how open-source AI works. That restructuring would, by coincidence, advantage the closed, proprietary systems built by well-capitalized labs with legal departments capable of navigating federal compliance.
The safety argument does not require this outcome. But the structural incentives point directly toward it.
None of this means AI safety concerns are fabricated. Frontier models trained on vast data, capable of generating persuasive content, writing functional code, and advising on sensitive decisions, do carry genuine risks worth taking seriously. The argument for some form of evaluation before deployment is not irrational.
But safety is always implemented through a specific institutional architecture. That architecture distributes power before it distributes protection. The question worth asking about any proposed AI vetting regime is not whether safety matters. It is who holds the stamp, who appeals to whom when the stamp gets denied, and which models were already through the gate before the gate was built.
Whoever defines AI safety gets to decide which AI wins. Everything else is implementation detail.
Editorial Note
underneath.news analyzes structural patterns, power dynamics, and the conditions that shape contemporary events. This is original analytical commentary, not reporting. We do not summarize, paraphrase, or replace coverage from any specific publication.
