The Washington Examiner published two related reports on September 15. OpenAI lobbyists are backing regulatory provisions in the bipartisan FRONTIER Act, H.R. 9925, introduced in July by Reps. Jay Obernolte and Lori Trahan. Rep. Josh Gottheimer, who co-chairs the House work on the subject, is pressing a different method. He wants mandatory government review of the most advanced models and has said third-party audits alone do not meet the moment. The live argument in Washington is now the machinery. Who assesses a model, what standard that person uses, and how long a developer can survive a refusal will decide whether the statute protects the public or merely rations the right to compete.
Cities inspect bridges because collapse is final and inspection is cheap insurance. The trouble starts when only 3 firms qualify to inspect, every appointment takes 9 months, and an engineer whose design is rejected never learns why. The shortage does the work a ban would have done. Congress is writing a version of that arrangement into federal law for artificial intelligence. Safety is doing a lot of the talking. The identity of the gatekeepers is doing too little.
I argued in June 2025 that Congress, not 50 legislatures, should write the rules for this industry. The White House took the same view in its December 2025 fact sheet calling for a national framework to replace a pile of state mandates. A company should not need 50 legal opinions to ship one product. Uniformity still has to shrink the total burden. Swap many uncertain demands for one expensive bottleneck and the tidy map produces less opportunity than the sponsors promised. The test that matters is practical. Can a new lab read the obligations, budget for them, and meet them before its capital runs out? If it cannot, the framework has failed on the terms used to sell it.
H.R. 9925 is tiered, and the text should be read as written. It distinguishes frontier developers from large developers and from very large developers. Large developers must publish a frontier AI framework covering catastrophic-risk thresholds, model-weight security, and deployment decisions, then undergo annual independent audits for compliance with that framework. Only very large developers must retain licensed independent verification organizations for ongoing assessment. That last category requires both revenue above $5 billion and at least $10 billion in AI development spending over the preceding 36 months, counting affiliates. A garage lab does not walk into a federal licensing board on day one. The pressure the bill creates sits in the assessor market; it would stand up. An independent verification organization would judge whether a very large developer’s published framework, followed as written, is adequate to hold catastrophic risk to an acceptable level, including when the model is used inside the company. That is a serious job. It only works if more than a handful of organizations can do it on a calendar a growing industry can survive.
The Department of Commerce would license those organizations and set their qualifications through a new Under Secretary for AI Security. Sections 3 and 7 address assessor conflicts of interest and assessor licensing, and they require the Government Accountability Office to examine the assessor market every year for capacity, entry barriers, and independence. Section 9 preempts state law in a targeted way, with exceptions. Section 8 pairs emergency restrictions for imminent catastrophic risk with administrative review, constrained judicial review of provisional orders, and an exception that preserves constitutional claims. The sponsors saw the concentration problem. That credit is real, which is why they are the audience for a harder draft rather than a dismissal.
Gottheimer’s track and the FRONTIER design are two different gates. He would put the review inside an agency. H.R. 9925 puts it in a licensed private market that Commerce would supervise. The private version is easier to defend in a hearing because it looks like independence. It is harder to defend in an actual queue if only a few shops can do the work, if those shops take months to schedule, and if a rejected developer cannot learn why the verdict came down.
A formally even rule can still land unevenly. Joshua Levine’s June 2025 study for the Cato Institute treated auditing and impact-assessment procedures as natural sites of regulatory capture, and the mechanism is ordinary rather than conspiratorial. Large compliance departments are a fixed cost. A firm with $5 billion in revenue can absorb delay, keep a scarce assessor on retainer, and staff the paperwork. A lab of 12 engineers with a better training run cannot. The statute uses the same words for both. The calendar does not. The Heritage Foundation has made the same warning in plainer language. A centralized partnership between government and a handful of frontier firms, sold as safety, is the outcome to fear. Heritage’s critique of California’s earlier frontier bill ran on the same rail. Thresholds and reporting that look neutral on paper become barriers once only a few shops can staff them. Competition among open and closed models, and among multiple providers of apps and infrastructure, is often the discipline that actually works.
A rulebook written around the habits of today’s leaders also risks treating one business model as the only responsible way to operate. Congress should ask of every provision whether it measures a danger or merely rewards familiarity with a particular laboratory’s methods. That question still matters when everyone in the room believes the public purpose is real.
Vice President Vance’s remarks on September 14 named the tension in public. He granted that AI carries real risks and real benefits, said the administration wants to “regulate smartly,” and said he was uneasy that frontier companies are the ones asking for the rules. Lydia Moynihan made the same observation in the New York Post that day, warning against a “please regulate us” turn that just happens to freeze the field. Established firms have an incentive to favor industry-wide constraints they can carry and rivals cannot. The incentive is visible on the balance sheet. Chris Lehane’s September 9 statement for OpenAI says a federal framework should target “the handful of well-resourced laboratories developing the most capable systems” and that “a serious public framework should reduce, not increase, the concentration of power.” Convert the second sentence into terms a competitor could enforce. A shared line in a white paper is where the drafting starts.
Once an outside assessment becomes a condition of competing, access to competent assessors is public policy. The GAO review in the bill is useful and late. A report on a concentrated assessor market after it forms is a different task from building capacity before the obligation bites. Congress should require published waiting times next to completed assessments. It should set qualification standards that a new evaluator with a different technical approach can actually meet. It should decide whether a developer may carry valid assessment evidence from one evaluator to another, so changing shops does not mean repeating a year of work. Security and intellectual property will force a defined handoff. Portability makes exit practical, and exit keeps an assessor honest.
Standards must be justified, not assumed. The Mercatus Center’s AI Policy Guide makes the point that keeps getting skipped: “AI audits depend on application.” Goals and methods vary with what the model does. No single audit uncovers every problem. A mandate should say what an assessment must establish, let qualified organizations show different ways of establishing it, and require written reasons when an applicant is rejected. An assessor who will not say why a model failed leaves no record that a court, a rival evaluator, or the public can test. Conservatives have spent a generation watching agencies exercise power in that form. They should not bless the same pattern because the subject is code.
Section 8 does provide review of emergency orders, and the bill should acknowledge that. The review offered for provisional orders is still constrained. For a small developer, a stop-development order is not a pause on the calendar. Investors leave, engineers take other jobs, and by the time the administrative case closes there is often nothing left to vindicate. Preserve the government’s power to act on an evidenced emergency. Require early and effective review of any decision that halts development, with the burden of explanation on the institution using the power. Developers remain responsible for meeting justified obligations. The state remains responsible for justifying them.
Competition among assessors could reward weak auditing, and the design must account for that risk. A developer might shop for an agreeable evaluator. Cheap work can look like rigor until someone tests it. Demonstrable qualifications, disclosed conflicts, periodic testing of assessor performance, and mandatory reasons for consequential decisions make a purchased verdict harder to hide. Paperwork volume is a poor proxy for safety in either direction. Competing evaluators and public criticism find weaknesses a closed roster has no incentive to publish.
Jennifer Huddleston warned Congress in April 2025 against treating early AI leaders as permanent monopolists. She was right, and the warning still fits. The ecosystem is wider than a handful of laboratories, and today’s leaders acheived their place by building tools people actually used. Supporting the country’s best innovators and protecting everyone else’s right to displace them can happen under the same statute. An America First AI policy has to protect the nation’s capacity to produce challengers, including the engineer who has not yet incorporated, the researcher working from a different method, and the customer who wants a second vendor because an industry that cannot be challenged from below will call that arrangement safety, and the rest of the country will be asked to treat the label as proof.
Congress can write a national framework, and it can put real obligations on the largest developers. Handing a small circle of laboratories and compliance firms the lasting authority to decide who may challenge them, then calling the result safety, is the outcome the present draft still leaves open. Before the government requires an outside organization to judge a developer, the public should know who can become that organization, what its verdicts must rest on, and how a wrong verdict gets corrected while the company is still alive to care. The FRONTIER Act already admits the assessor market can concentrate. The next draft should treat that admission as a design problem, not a reporting footnote.
If you enjoy my work, please subscribe https://x.com/amuse/creator-subscriptions/subscribe
Alexander Muse is a Fellow at the John Milton Freedom Foundation and publishes daily political analysis at amuseonx.com. Primary sources cited in this piece are linked inline; campaign finance figures are drawn from FEC filings, polling data from publicly released crosstabs, and legal claims from filed pleadings. Corrections are posted to the original URL with a dated changelog. Readers who identify errors are invited to contact the author directly. Each op-ed edited for grammar and clarity using Ai in a sponsored partnership with Grammarly. Data provided in a sponsored partnership with Polymarket.
all in favor of banning mathematics, raise your hand.