A bipartisan AI bill has passed committee three times. It still isn't law. The agreement was never the hard part.

The consensus view of AI policy in Washington this year is that there is no consensus. The White House spent December trying to preempt state AI laws and stood up a task force to sue the states that don't comply. A coalition of state attorneys general subpoenaed OpenAI in June. A pair of frontier models went dark worldwide on a classified export order the same month. Brussels spent the first half of the year quietly gutting its own AI Act under the banner of a "digital omnibus." If you cover this beat, or just read the headlines, the fair summary is that American and European AI policy is a series of collisions — safety versus innovation, state versus federal, export control versus open science — with no room left for anyone to simply agree.

That summary is wrong, and the way it's wrong is instructive. It is not wrong because the collisions aren't real; they are, and I've written about several of them from this column. It's wrong because "AI policy" is not one fight. It's a label we've slapped over a dozen structurally different questions — who may release a model, who may use one to fire an employee, whose chips may leave the country, who gets to do the research in the first place — and treating them as a single undifferentiated subject is a category error with a specific, measurable cost. It is the reason a bill that four sitting senators from both parties agree on, that has cleared committee by voice vote more than once, and that has now been reintroduced in some form for three straight Congresses, still has not become law. Not because anyone is fighting it. Because nobody outside a small group of people is paying attention to it, and the loud fights are eating all the oxygen the quiet one needs to reach a floor vote.

The bill is the CREATE AI Act. It is, by a wide margin, the most boring acronym in AI policy — Creating Resources for Every American to Experiment with Artificial Intelligence — and that is precisely why it deserves fifteen minutes of your attention.

What the bill actually does

Strip away the branding and the CREATE AI Act does one specific thing: it would take the National Artificial Intelligence Research Resource, currently a pilot program run by the National Science Foundation, and make it permanent, statutory infrastructure. NAIRR has operated since January 2024, and its function is narrow and useful — it gives researchers who are not employed by a frontier lab or a hyperscaler access to compute, datasets, models, and training tools they otherwise could not afford. Reportedly, it has supported more than 600 research projects across all fifty states in that time, everything from language-model safety work to privacy-preserving synthetic data generation. It is, in effect, the plumbing that lets a university lab or a community college researcher do AI research at all, rather than watching the field consolidate entirely inside five companies with the balance sheets to buy their own compute.

The Senate version, S.4441, was reintroduced on April 29 by Sens. Martin Heinrich (D-N.M.), Todd Young (R-Ind.), Mike Rounds (R-S.D.), and Cory Booker (D-N.J.) — two Democrats, two Republicans, none of them positioning this as a signature culture-war bill. The House companion, H.R.2385, is led by Reps. Jay Obernolte (R-Calif.) and Don Beyer (D-Va.). The bill would require NAIRR's operating entity to set aside a "significant percentage" of its annual compute allocation specifically for research on AI privacy, ethics, safety, security, and trustworthiness — the unglamorous safety research that doesn't happen if every cycle goes to whoever can pay for it.

None of this is new. A version of the CREATE AI Act has now been introduced in three consecutive Congresses going back to 2023. The House Science, Space, and Technology Committee has passed it before, by voice vote, alongside a slate of other AI bills — the kind of unanimous, no-drama committee action that never makes a headline because it isn't a fight. And it still isn't law. Three years of a bipartisan foursome agreeing on the same idea, repeated across three Congresses, is not evidence that this bill is hard to agree on. It's evidence that agreement was never the obstacle.

The strongest case against calling this "free"

Before I make the case for why that gap matters, I want to make the skeptic's case properly, because it's better than the bill's boosters usually let on.

The NAIRR task force's own 2023 report put the price tag on a full-scale program at roughly $2.6 billion over six years — real federal money, not a rounding error, and not something Congress can wave into existence by changing a program's legal status. Calling four senators' agreement on a framework "consensus" elides the actual fight, which isn't over the framework, it's over the appropriation, and appropriators have shown zero comparable enthusiasm. The pilot itself is a case study in the gap between ambition and funding: in an early allocation round it received more than 150 proposals and could fund 35. That is not a program straining at the edges of demand because it's new. That is a program that has been resource-constrained since month one, and "codify it" does nothing, on its own, to change that math.

Worse, the timing is bad in a specific way. The National Science Foundation is, reportedly, trimming budgets across hundreds of its core science programs by 20 to 30 percent this fiscal year even as its topline funding fell only a few percent — the agency is reportedly redirecting money toward a new commercialization initiative, in apparent tension with a congressional directive that no NSF directorate take more than a 5 percent cut from 2024 levels. That is the funding environment NAIRR would be codified into. If Congress passes a statute that says NAIRR is permanent while the agency housing it is under this kind of internal pressure, the realistic version of "permanent" may be a permanent name attached to the same scarcity the pilot has run under since 2024 — 35 funded proposals out of 150, indefinitely, with better letterhead.

That is a fair objection, and I am not going to pretend it isn't. A statute is not a check. If the honest test of this bill is "will more researchers get more compute," passage alone doesn't clear that bar, and anyone selling you codification as a funding solution is selling you something.

The harder question: what codifying actually changes

But that framing — statute versus funding — is itself where the skeptic's case, as strong as it is, runs into the same category error I opened with. It treats "does this get funded" as the only question worth asking, when the procedural question is a separate one, and it's the one I spent most of a decade inside the government actually answering: who administers this, under what authority, and what happens to it when the people in charge change?

Here is what a pilot is, structurally. It exists because an executive-branch entity — in this case, built off a 2020 statute and a subsequent executive order — decided to stand it up, and it persists for as long as an administration chooses to keep funding and staffing it through ordinary appropriations and program authority. Reportedly, the Trump administration revoked the Biden-era executive order that helped launch NAIRR's pilot phase, and the pilot kept running anyway — a genuinely encouraging sign, but also a bit of a fluke. It survived because its footing in the underlying 2020 statute and existing NSF grant authority happened to be strong enough, and because nobody with the power to defund it chose to. That is not durability. That is having dodged a bullet.

A statute is different in kind, not just in degree. Once NAIRR is authorized by law rather than sustained by program discretion, killing it requires an affirmative act — Congress repealing the authorization or zeroing its appropriation on the record — rather than the passive, low-visibility choice of an incoming administration simply declining to renew what the last one built. It changes who has standing to sue if the operating entity misbehaves or exceeds its mandate, because a statutory mandate creates a clearer legal basis for that kind of oversight than a program a Cabinet agency runs at its own discretion. It changes what the Government Accountability Office and inspectors general are empowered to audit against, because they audit against the law, not against a press release. And it locks in the structure the task force actually recommended — a single federal agency housing operations, a cross-agency steering committee setting direction, and an independent, non-governmental entity running it day to day — so that structure can't be quietly reshuffled by whoever holds the NSF director's office in 2029.

None of that guarantees the money follows. But it is not nothing, and dismissing it as nothing is the mirror-image mistake of pretending it's everything. The realistic version of this bill is: codification changes the legal durability and the oversight regime, and does approximately nothing to change the appropriations fight on its own. Both things are true. A columnist who tells you only one of them isn't being straight with you.

Statutory authorization does not buy you money. It buys you the fact that killing the program requires someone to vote for killing it, on the record, instead of just declining to write the check.

Why the quiet bill stays quiet

It's worth being specific about what makes 2026's other AI fights loud, because the contrast is the whole point. The state-preemption fight is loud because it's a jurisdictional fight over who gets to write binding rules on companies with trillion-dollar valuations, and it touches employment law, child safety, and free speech all at once. The export-control fight over frontier models — the kind that took two systems offline worldwide on a single classified directive in June — is loud because it sits at the intersection of national security and a handful of companies' commercial survival, decided by an authority that doesn't have to explain itself in public. The EU's rollback of its own AI Act is loud because it's an entire continent's regulatory body second-guessing itself in public, under industry pressure, with global businesses recalibrating compliance budgets in real time. Each of those is a fight over who gets to say no to something, backed by an enforcement mechanism with real teeth — a lawsuit, an export license revocation, a regulatory delay that reshuffles who's covered and when.

CREATE AI Act isn't that kind of question. It doesn't decide who may release a model, what a company must disclose, or whose chips may cross a border. It decides whether a graduate student in New Mexico or a community-college researcher in South Dakota gets a shot at the compute a frontier lab takes for granted — a question about access to inputs, not control over outputs. It has no natural adversary. Nobody's business model depends on NAIRR staying a pilot. That is exactly why it can attract four cosponsors who disagree about nearly everything else in tech policy, and exactly why it has no natural champion loud enough to force a floor vote when the calendar is dominated by fights that do have adversaries on both sides pushing for resolution.

• Loud in 2026, with an enforcer: the state-preemption task force (Justice Department, against state AGs); the Fable/Mythos export shutdown (an unpublished federal directive, against Anthropic); the EU Digital Omnibus rollback (Parliament, against its own prior AI Act); the state AG coalition's OpenAI subpoena (state attorneys general, against OpenAI).
• Quiet in 2026, with no natural adversary: the CREATE AI Act — reintroduced three Congresses running, cleared committee by voice vote, cosponsored across party lines, still not law.

That list is not a coincidence. Access-to-infrastructure questions are quiet because they're structurally the easy case: nobody is asked to accept a cost they're organized to resist. Enforcement questions are loud because somebody always is.

The better frame, and what it costs

So the case for passing the CREATE AI Act is straightforward, and I'll make it plainly: a bill that four bipartisan senators agree on, that has survived an administration change once already, that funds the unglamorous safety and privacy research nobody else has an incentive to fund, and whose only real defect is a funding gap Congress could close if it wanted to, should not still be sitting in committee limbo in its third Congress. Pass it. Then fund it properly, which is a separate fight worth having loudly, because a statute without money behind it is a plaque, not a program.

But the frame I'd push back on hardest is the one the bill's own boosters sometimes reach for without meaning to: the idea that this kind of quiet, genuine bipartisanship is evidence that AI policy, in general, is more tractable than the noisy headlines suggest — that if Washington can agree here, it can probably find its way to agreement on the harder stuff too, given time. That would be its own category error, just wearing a more optimistic face than the gridlock narrative it's replacing. The reason CREATE AI Act is easy is not that AI policy is secretly easy. It's that this particular question — who gets a seat at the research table — happens to be one of the few AI questions that doesn't force anyone to accept a real, organized cost. Model licensing forces labs to accept delay and disclosure they'd rather not give. Export controls force companies to accept revenue they'd rather not lose. State labor and consumer rules force employers to accept liability they'd rather not carry. Every one of those fights stays loud because it has to: someone with real power is being asked to give something up, and they are, entirely reasonably, resisting.

Infrastructure access asks nobody to give anything up. That is exactly why it can pass on a voice vote — and exactly why its ease tells you nothing about whether the harder fights are close to resolving. The honest reading of 2026 isn't "Washington can't agree on AI." It's "Washington agrees easily on the AI questions with no organized opponent, and disagrees hard, in public, on every question where somebody stands to lose something real." Those are different diagnoses, and they call for different prescriptions. Mistaking the second for evidence about the first is the harder version of the category error I started with — and it's the one I'd ask this bill's own supporters, myself included, to be careful not to make on their way to the floor vote they've earned.