Decoding State-Level Regulation with Jason Elliott

[music] Welcome back to the AI policy podcast. I'm Alukmetha, director of the Wadwani AI Center here at CSIS. Our guest this week is Jason Elliot. Jason is the president of Versus Solutions, a political consulting firm, but before that, he served as a deputy chief of staff to California Governor Gavin Newsome. Jason advised the Newsome administration on a wide range of issues, including technology and artificial intelligence and helped guide the governor through some of the nation's most influential policy decisions in the space, not the least of which was a first in the nation law regulating frontier AI. Before joining the administration, Jason served as a senior adviser to both Governor Elect Newsome's transition and Newsome's campaign for governor of California. He worked for three San Francisco mayors, first under Mayor Newsome and later as chief of staff to both Edwin M. Lee and London Breed. Jason is also the California Policy Fellow at Stanford University's Institute for Human Centered Artificial Intelligence and a member of both the California State Protocol Foundation Board and the California Highspeed Rail Authority Board of Directors. Over the past year and a half, much of the debate in the AI policy space has focused on the issue of preeemption. Should the federal government block the states from regulating AI in some way? We've seen multiple attempts to do this both legislatively and via executive order, but they've had limited effect to date. Meanwhile, states have served at the the tip of the spear on frontier AI with both New York and Illinois passing significant laws uh on the heels of the laws we saw in California. So I'm eager to dig into things with someone who can offer a unique perspective on how states are approaching the broad issue of AI regulation. Jason, thanks for joining us today. I'm really glad to have you on the show. Look, thanks for having me. >> Yeah. So um I wanted to uh start um with sort of a highle question. I have lots of questions specifically about what California has done in the AI policy space um and and these two big bills that um many of our uh listeners have heard of SB 1047 and SB53. But before I do that, before I dive into those things, I wanted to do a little scene setting. Um I think a lot of our audience is pretty familiar with how things work in Congress and the federal government, but probably less so with the mechanics of state government. And so I'm really interested in um your take on some of the biggest differences between um how things work in the state legislature versus how they might work uh here in DC. >> Yeah, that's a great question and a good place to start. I I guess I would start by saying I'm just really thankful that you invited me on because the ability to have a conversation about state policy in the national AI policy context I think is really important and I think that people that are working on federal AI policy uh would do well to be more exposed to some of the dynamics uh that are happening at state houses. So I'm glad that we get to have this conversation about subnational policy and regulation. uh to answer your question directly uh I think one of the key differences for California and each state is going to have a different answer to this question because each state legislature uh has different processes but in California I'd say one of the key differences uh is that the state legislative calendar in California starts around January 1st and ends around this time of year roughly September. So it's unlike Congress where at any moment they could be in session working on bills, sending bills to the president. In California, it is really a sort of 8 to nine month sprint at which time the governor has delivered 1,500,700 bills and then he has a month to go through and sign or veto uh every single one in the period of a month. We are now it's September 2nd as we record. uh we are now in the proc the the phase of the process where the governor is about to receive the legislative session is over. He's about to receive some 15,600 bills uh and then he has until September 30th to make a decision on on every single one. I would say another key uh difference to point out in California again each state would be different. Uh it is traditional in California that the governor vetos somewhere between 15 to 17% of bills that come to him. That's obviously a much higher proportion than the president excepting times when there's a deeply divided Congress and presidency. Um, but in any given year, the governor gets more than 1500 bills from the legislature and about 15% of them get vetoed. So, it really does behoove a legislature to work closely with the administration to make sure that the bill that he or she wants to send to the governor's desk is something that the administration would approve of. So this is a nuance in the power dynamic between the legislature and the governor. >> And I know you you talked specifically about California. There there are other changes uh other differences too. So in a lot of states the legislators are part-time uh fewer resources available. It's harder to develop deep expertise in areas like technology. in a lot of states. I think California um has a lot of expertise at the state level um in technology policy because of you know uh it being home to a number of these countries and also the size of the state. >> That that's absolutely right. I I I'll point out uh Jay Albernalty who's a Republican uh is of course uh one of the two uh uh bipartisan co-authors of the Frontier AI Act which maybe we'll talk about later in this discussion. uh he's from California, a Republican from California, comes from the tech industry. So we have a number of legislators who are either experts, scientists, technologists, business people who are serving in the legislature, and as you said, they are full-time and they have pretty robust staffs. So yeah, it's it's a very professional and professionalized legislature here. >> Well, you know, I know a number of the bills that are sitting on the governor's desk involve AI like they have in in um past years, and uh I'd like to get to those. Um, but before I uh do that, I wanted to talk about these big these two big frontier AI bills, SB 1047 um and then SB53. Um, and I start I wanted to start with um 1047. So, this was uh a really controversial bill, probably the first frontier AI bill that really broke out into the mainstream press. Um, and it had um, so this was uh, authored by California Senator Scott Weiner. It came out um, originally in February 2024. It had elements in it that have now become sort of relatively common like requiring Frontier developers to uh, implement various safety protocols and publish various information. But it also had some controversial provisions that we don't see uh as much anymore like a model kill switch requirement. Um and it ultimately made it to the governor's desk uh and was vetoed in September of 2024. Um so so I wanted to start you versus um your your current firm uh by the time SB 1047 was passed by the state legislature, but you had a few months of overlap in the governor's office when the bill was undergoing amendments. And I I really like to understand from your perspective um some about the bill's reception. What were the political dynamics like? How are you and your colleagues thinking about the balance between safety and the public interest versus you know the concerns of the tech um and venture capital industries that drive so much of California's economy. >> Sure. Uh we can revisit the the 1047 debate. Um so I think it's important to ground ourselves in how long ago in AI time mid 2024 in fact was uh 2 years in and for all the listeners of your show I don't need to explain this at all but 2 years 2 and a half years in AI time uh is is generations is is eons and I think you know look it's just to say and I I imagine we will cover SB53 which was the successor bill to SB1047 it's just to say a number of the ideas that Senator Senator Weiner put forward in 1047 have now become law through SB53 and are now derur as you said for other states and even the federal conversation. You know, I think Scott Weiner is a very good legislator and very well-intentioned. I think the issue with 1047 um and this is frankly what the governor governor Nome's veto letter said. A number of the policy prescriptions in 1047 uh were put forward because of assumptions about what the technology could do. Um, and frankly, two years ago, I I don't think that we had as much of a developed sense as we do now around loss of control uh, and other sort of emerging uh, capabilities that we see pretty regularly in models. Now, look, I I think, you know, some of the sponsors of that bill have sort of said, see, look, we were right all along, and I think there's some fairness to that, frankly. Uh, but the time was too early. And I think, you know, thinking about California's important and unique role in the regulatory ecosystem, California is with all the respect and love to my other large states in America, California is the most important state on tech regulation because we are home to all of these companies because we have access uh to the great institutions uh Stanford, Caltech, etc., Berkeley um and the researchers therein. California is the most important state on tech regulation. So, we made we really wanted to make sure we got it right. And you know, I say all this to say SB 1047 was too early for how we understood the problem to be. A number of the components of that bill have now moved forward in law. And you know, I think that the um the the proof is in the pudding in some sense. 1047 was vetoed largely on the argument that there was no empirical evidence necessarily at that time to suggest that particular suite of solutions that was proposed in 1047. It did lead to SP53. I think it's hard to talk about the 1047 without talking about SP53. SP53 did become the model bill that was adopted in a number of other places. I don't think that SB1047 would have become that. I think that SB 1047 had it been signed into law in 2024 would have raced too far in front of the discussion and would not have had this positive cascading effect of being copied by a number of other states. So, I know it's a bit of a convoluted answer, but it's a long way of saying we didn't know easy now in 2026 to look backwards. Hindsight always uh easier, but at the time uh there was a lot of conjecture about what AI could and couldn't do and what some of the emergent uh catastrophic risks were going to be and how they should be regulated. Also, at the time, uh in 2024, Joe Biden was still president, which is most certainly not the case now. And so we did still have some hope that there would continue to be democratic administrations in Washington that would actually advance the safety framework. So there were a number of different reasons why SB1047 in its moment was not ready. But again, I don't mean that as a critique on the people who put forward the bill or on Scott Weiner because as we've said a number of times, a good number of those components have now become law. So, you've talked uh a little about sort of some of the concerns about the bill and and probably um hinted at the reasoning that the the governor had for vetoing the bill. I'm curious if you if you think the expectation was always that some modified version of this bill would would um appear in a subsequent year. um if that was the expectation uh in the governor's office if if part of it was part of uh his logic for setting up this sort of advisory committee and expert working group with um Dr. Fa Lee and other AI experts to sort of um look into this issue and provide recommendations to the governor's office. Look, it's easy for me now to say yes, this was always the plan. And I think that's a little bit of revisionist history if we're being honest. But if you look at the governor's veto letter from September of 2024, he lays out very clearly that, and I'm paraphrasing, but he says there is a problem here to be solved. And to anyone who thinks there's no problem here, or that California has no role in solving this problem, I frankly disagree. So even in vetoing 1047, Governor Nuomo was very explicitly saying, "We do need to address this problem, and if the feds aren't going to do it, we are going to do it in California." Again, putting ourselves back in time, September of 2024, it was a coin flip. who is going to be the next president of the United States. We of course know how that turned out. Uh but at the time there was some expectation that the federal government would actually do its job. That's clearly not played out. Um so I I think the answer to your question is yes. There was always an expectation that California was going to step up if the federal government did not. Now, exactly what form that was going to take, you know, I can say uh in 2020 20 in the fall of 2024 when 1047 was being vetoed, there was not already a bill that had been drafted that was in someone's coat pocket ready to pull out and put forward. The the the process that Dr. Lee ran and Tino Quayar uh who was at Carnegie at the time now at Anthropic and Jennifer Chay uh the dean at college of computing at Berkeley. The process that they ran to write the report they wrote um over the end of 2024 and into the beginning of 2025 was instructive in the design of the bill that eventually became SB53. So there was not a preconceived notion of what the components of this subsequent bill would be, but there was definitely a clear acknowledgement that California is not going to sit idly by if the federal government doesn't do its job. >> So let's turn to to SB53 um quickly. So um SB53 is is in many ways the successor of SB 1047. It includes many elements that were originally floated in 1047, most notably the safety protocol requirements. Some of the other controversial elements got dropped in this version. Um, >> and this time the the outcome was very different. So there was far less opposition from the tech industry and and the government and ends up signing it and it becomes the first frontier AI law in the country. Um, so I'm curious about um your assessment of of the differences between SB 53 and 1047. Was was some of the diff was was the difference largely a a difference in content because the the substantive content of the bill was different? Um or or had the political dynamics also shifted? Maybe sort of maybe adding on top of that, had our technical understanding of AI also changed? Um, and did that contribute to the the different reception that this bill had? Uh, I'm gonna guess that you're going to say it's a it's a little bit of all of those. It's a little bit of all of those. Good guess. Uh, and a couple other things I would add. So, it is content. It is context and it's expertise. And I think expertise I want to get to because it it is a roadmap for other states and and and subnational governments that want to uh advance AI policy in the absence of national leadership. But we'll just sort of take those one at a time. Content. Yes. I the most important thing is that the bill was different. Uh it did not include a kill switch which has all kinds of technological and operational uh problems. Uh it did not include SB53 did not include some of the liability provisions uh which um SB 1047 did include liability on the frontier developer. Um SB53 did not include an audit requirement which is interesting and we should dig into that because um successors to SB53 namely the Illinois bill uh do include audits. I'm of the personal perspective that audits are are not actually a sufficient mechanism to guarantee safety and you need to go further and do some sort of third-party verification um beyond just auditing a company that they've done what they said they did. I think it should actually go further so we can talk about that little fork off of uh off of the policy. But then 53 did include, as I said in an answer to a previous question you asked, SB53 did include a whole number of things that were also in 1047. So it wasn't an entirely different bill. Uh a lot of the disclosure and transparency requirements were held in. Whistleblower requirements were held in. Cal compute was it a public compute cluster for California for research purposes was held in. So there were a number of things. Critical incident reporting was a big ad. So there were a number of things that were held over and and all of those components that were in 53 then effectively became the raise act in New York and effectively became the Illinois bill and also formed the basis of the Massachusetts bill and the old nultraan bill. So it really has multiplied in a way that was intentional. We were hoping that what we would create in California would uh if not an if not an actual federal federal policy at least a national policy um that would diffuse across uh different different states. So that that was intentional. >> And I'm curious about your take on the the political dynamics and particularly the dynamics with the tech industry. So the tech industry for the most part came out, I think, pretty vehemently against 1047, but there was a much different reception to SB53 and and even some endorsements or soft endorsements. And I'm curious about your assessment for why that is. and is part of the reason that SB53 was guided much more by some of the voluntary practices and commitments that companies were already making particularly companies uh in the US. >> Mhm. And thanks for the question. This is the second part of I think what made 53 feel different was context. You know uh just a few things to remind listeners in 2024 Joe Biden was president. the voluntary commitments were still in place uh with the leading AI Frontier Labs and Deep Sea had not yet happened. Right? You fast forward into the middle of 2025 and all three of those things had changed. Donald Trump was president. The voluntary commitments had been stripped or withdrawn and the deepseek moment changed the political discourse about what the United States should do uh to regulate uh AI or not do to regulate AI. So the context was very different and I think uh a combination of all of those things um is what led a number of the companies to engage very constructively with us during the SB53 negotiations. A fourth thing I would say is that Governor Nome made it plainly clear that he wanted to sign a frontier safety bill which was not necessarily the case in 1047 in the year prior. And you know, as I had mentioned earlier, uh the structure of the sort of legislative process, the governor does play an outsized role in California, maybe more so than the president in the congressional legislative uh process in Washington. So the governor leaning very strongly into uh demanding a frontier bill uh from the legislature again created the weather pattern to bring the companies to the table in a much more constructive way. And they did and they came to the table, negotiated in good faith. It was not an easy negotiation uh but negotiated in good faith and the bill that came through as you mentioned was uh implicitly supported or we gain neutrality from the companies which in the parliament of you know the legislative process when companies are neutral uh that that means that they're comfortable with what's in the bill and I don't want to put words into any of the company's mouths so I I will just say I I think that part of the reason that they felt it was okay to be neutral on SB53 was because They knew other states would copy it. And if you can make something work in California and negotiate well with the governor's office and the legislature in California, it can then become a model in other states and it's better than going and starting from zero with every single state legislature across the country and never knowing where you're going to end up when that fractal starts to break. And I think that was a lot of the reason why uh they decided to be neutral on this bill was this this was the devil they knew in some sense. So, one of the things we argued here um at the Wadwani Center, including in a recent paper we put out about state bills, is that um what we saw happen with uh SB53 and 1047 is uh exactly what we want to happen, which is states playing roles as laboratories of democracy, sort of exploring policy solutions. uh some will will um work out the the time is right for them. Others won't because they're either um uh don't address the issues well or they they go too far in one aspect or another. Um and so that is what happened with 1047 and that's why SB53 ended up passing based on the lessons that had been learned the the prior year. Would you would you agree with that assessment? [snorts] I do agree with that assessment and certainly California fashions itself one of those leading laboratories to extend your metaphor and not just on technology but on climate policy and on social justice issues. You know, I think it's interesting the laboratory uh of democracy concept really only can be true if the if the discoveries from that laboratory to torture this analogy if the discoveries from that laboratory are able to be replicated. If you have a hostile, belligerent federal government standing in the way of states legislating, then the laboratory for democracy thesis can't really fulfill itself because if California does something meaningful, uh, and then the federal government comes and, you know, threatens to withhold bead funding from states that dare copy California. Well, then the laboratory, the research research can't propagate in the way that it really ought to. So I I yes I do agree with the premise of what you're saying but it really does require a federal government also in the federal in the sort of structure of federalism allowing states to be those laboratories. I would just add to um your description of the laboratories of democracy concept. Uh the third um part of why I think SB53 was successful which I mentioned earlier which is expertise. you know, there's a in California or in any subnational government, you can invite voices in because we're just smaller than the federal government. Any individual voice can seem a little bit louder. Um, and you can actually have a proper discussion um about policy trade-offs. And in the case of the 1047 to 53 uh dynamic, it was uh Dr. Lee, uh Dr. Quayar, Dr. chase these three luminaries that came together with their staffs of researchers and brought forward a researchbased report. And like look, we've all, you know, been in government or been around government and these reports often get issued and then they're discarded because there's already a foregone conclusion or no one cares what the researchers think. That was not the case with this Frontier Safety AI report. this Frontier AI report became the basis of what turned into SB53 and those scholars were consulted regularly. Um, and they were able to speak truth in a way that was apolitical. I mean, everyone has biases, everyone has preconceived notions, but in a way that was not political in the traditional sense of the legislative politic. Uh, and that was determinative for getting 53 done. And when we had a challenge in drafting, here's a trade-off. How are we going to draft this particular section of the bill? we had we had a a a bench of scholars that we could call on and say what would you guys do here right like help us think through some of the trade-offs don't tell us politically what needs to be done that's not appropriate that's not your job but help us think through some of the policy and practical trade-offs and that expertise made a real big difference and and I found that in going through that process I I helped with the report um through my um opices as a Stanford um policy fellow I found just a dozens if not hundreds of academics from all over the country eager to engage in the writing of this report or reviewing of this report. And to the extent that uh any of my state colleagues from other states are listening to this podcast today, there is a really deep bench of people that are apolitical uh that are willing to help uh if you just reach out to them. And they're at universities in your states and they're at universities in other states, but they're there. And for for us in the 53 SP 53 uh negotiation, uh they were really really important truth tellers. There's so many threads I want to pick up on um from what you just discussed. I think uh let me start with the other state laws. So after California passed SB53, we saw other states, most notably um New York with uh its raise act and um Illinois with SB 315 pass similar laws that contain a lot of similar elements. I assume that your take is that these laws um came into existence only because of the work that had been done in California around 1047 and then SB53. Uh we we certainly went first and then I think that these other legislators and legislators and governors did exactly the right thing which is they picked up the ball where we had left it and moved it further downfield. Um and you know and Illinois New York went just sequentially New York went first after California then Illinois uh and then Massachusetts is is in the process of debating a bill that's also very very similar to SB53. And I think that that's great. I think that that's fantastic. And you know I don't feel necessarily competitive pressure as a Californian some sort of California chauvinist that we got to like get ahead of Massachusetts today. I think that California should revisit SB53 and I hope that that California revisits SB53 next year, but to continue to work in partnership with other willing states. And it's interesting, it's not just blue, right? If you look at something like access to abortion, uh, or gun control or climate policy, it's the blue states versus the red states. That's basically how it breaks down. It's just simply not the case in AI policy. Um, you know, Ronda Santis passed an AI bill of rights and opposed uh AI preeemption in Congress. Ronda Santis and Gavin Newsome have never agreed on anything ever. Uh, and yet they agreed on that. So, it's not just a blue states moving the ball. It's all states moving the ball. And I think that that's really productive. You know, the one of the differences in in AI uh frontier safety specifically and cat- risk stuff specifically is if normally under a sort of a state protective regime for like air quality, right? All right. So, in California, we have very, very robust air pollution laws. We have very robust water quality laws and so forth. Um, you only get to enjoy the benefits of those laws if you live in California. If you're if you're drinking water in California or breathing air in California. So, it there really is a protection gap between red states and blue states when it comes to sort of quality of life and environmental health law. It's just not the case with frontier safety. If a model is trained uh is pre-trained um and the system cards are available for third party verification and whistleblower protections exist uh and certain behaviors are guardrailed out that benefits everyone in America. It benefits everyone in the whole world. So the protection gap that is traditional for sort of consumer protection uh really doesn't exist in frontier AI. So I think it's great when other states take the SB53 model and move it forward. You know, I'll just mention of of the states with the exception of Colorado, which was a bit of a disaster regulatory of the of all the states that we've mentioned, they have all taken the SB53 model and built upon that. It really is the vector for for AI frontier safety law in America these days. And the Obernolulti Treyan uh bill that's pending in Congress and I hopes get a vote and I hope I hope get sent to the president is also built on SB53. A number of the thresholds and thesis of that bill are the same as SB53. So I I think what the governor what governor Nuome and the state legislature did here in California in 2025 was they set the ball rolling and then everyone else has taken it and moved it forward in a meaningful way and added a little bit to it. And I think that that's a wonderful thing. >> So that that segus into what I wanted to cover next, which it has to do with this whole federalism and preeemption argument. So I think if you look at um we've heard a lot from from sort of federal policy makers about some of the reason that they support preeemption that uh frontier regulation is a national security issue and so it's the proper province of the federal government that a patchwork of state laws is going to be really problematic from a compliance perspective and that ultimately it's going to hurt the AI industry and and a lot of the long-term success of the US is dependent on this industry being successful. I think we can at least say that that if states are adopting similar models that are based on SB53, then maybe that um second argument is is less impactful. But I'm curious about from the state perspective uh how you might respond to the to the claims that really like frontier AI regulation is something that is properly regulated um by the federal government because it's say a national security issue or or it's an in interstate commerce issue or something like that. >> Here's how I'll respond to that. Frontier AI safety is the province of the federal government. They should be doing it. The states should not have to be in this game. But, you know, patchwork is the word that's always used. Patchwork is what federal inaction looks like from the outside in, right? You have a federal administration. I mean, let's not forget JD Vance went to Paris in February of 2025 and gave the infamous all gas no breaks speech where basically any frontier regulation is unpatriotic and you basically love China and hate America if you're for if you're for frontier safety. Then a lot of things happened including mythos and including hugging face and then they got a different idea and now maybe they're going to do some regulation. But now no they're not. Now they're doing the Carolina principles. That's the latest thing I saw right where they're basically saying that don't hurt they're they're they're imploring Scott Bessant is imploring other countries through the G20 to not do regulation. You don't want to hurt company profits through regulation. This is a schizophrenic federal policy at best. Right? Bad faith at worst, schizophrenic at best. I don't know who's in charge over there. I don't know what their objective is o over there in Washington. But I'll tell you what they're not doing. They're not advancing any meaningful frontier safety framework. So the states step in. You know, I think that the question abstracted from AI policy is what is the role of a state in consumer protection versus the federal government. Um it is I think clearly the purview of the states to do consumer product protection work. work. I think that's that's what um state governments do very well protecting life and safety. Uh ve very sort of standard state government work. In the case of frontier AI, it is a national security paradigm and the federal government should do its job. In SB53, I think one of the key things we put in SB53 that didn't really get a lot of attention, but I think it's a really important part of the policy design um is voluntary preeemption. So, in SP53, uh, for the critical incident reporting requirements, Frontier Labs are required to report to, uh, it's called the California Office of Emergency Services. It's kind of like our Homeland Security Department in the state. Uh, Frontier Labs are, uh, required to report critical incidents when their technology is used for enumerated uh, catastrophic harms. Um, and they have to report those in 15 days to our Office of Emergency Services. into our bill. We wrote a provision that said if the federal government, we hope because this was part of um one of Donald Trump's previous executive orders was he was tasking DHS to create a similar uh reporting structure. Um we said we hope that the federal government creates this and if and when they do, we are going to give our own agencies the authority to turn off our local requirements and just subordinate ourselves to the federal government because this is really where this should be happening. Lo and behold, the Trump administration didn't do what it said it was going to do. uh we have not had to voluntarily preempt ourselves. Uh but it is it's the concept there which is a lot of Democratic politicians especially Democrats say we wish we didn't have to do this on chatbot safety on data centers on frontier safety but since the federal government is not doing anything we're going to step in. It's a rhetorical point. I agree with it. Um we took the rhetoric and made it statutory. So now in California law is this sort of voluntary reverse federalism or voluntary preeemption as it's been as it's been called. And I think that that's a good model for other AI law in the states around the country to follow because it does call the question on Congress and the pre and the president. If you're going to do it, then do it. Uh but don't tell us we can't just cuz you don't want to. Well, I've heard, you know, informally from some governor's offices and staff that, you know, in in in a lot of ways they would prefer not to work on AI policy, that they would be comfortable if the federal government took sort of robust action on this. Um that the federal government has more technical expertise and more ability to to figure out technical information about models. And so that that would be useful and that they could spend their time doing um other things. So, do you do you do you think that um that's an accurate assessment of how some state lawmakers are thinking about and that um if the federal government did something really robust in this space that uh that a lot of states would be would be okay with that maybe like layer on some additional state specific things but generally be okay with uh a sufficiently robust federal framework. >> Uh yeah, the short answer is yes. I think that this is clearly something the federal government needs to step in and do, but I don't think preempting the entire space. The original uh Ted Cruz preeemption proposal, which failed 991 in the Senate, was a blanket preemption on any state law that had anything to do with AI. That was comical. That was that was a clown show. Um, I think the idea that, um, the federal government could come in and narrowly preempt things where they do take meaningful action, uh, I think is something I I look, I I don't spend a lot of time talking to Republican staffers, so I I'll speak maybe more on the Democratic side. Uh yes, I think if the federal government stepped in and took meaningful serious action, not sort of 1600 Pennsylvania window dressing, but real action, I think that that would come at the relief of a number of uh subnational leaders in America because they are the appropriate unit of government to be doing this work. But the second best is the states and if we can't have the best, we do we do need to employ the second best. >> Uh so one more question in in in the sort of federalism space. Um you've mentioned the Frontier Act. Uh the Frontier Act um from Obernoli and Tran um is an evolution of a previous bill where they uh called the Great American AI Act. They had a broader preeemption in there. It's been narrowed um in in the latest version. Um so what we observed is like for much of 2025 there's a really adversarial relationship between the federal government and state governments about uh frontier AI regulation multiple attempts at preeemption executive orders trying to take action against states regulating AI. Um it seems now that the dynamic um has shifted a little at least for some parts of Congress where they're thinking about what's happening to the states as an opportunity to learn uh lessons from from um what has happened at the state level and bring them over into the uh federal space. Um I you know our take is that this is a much better dynamic that you should >> certainly leverage you know the hard work being done at the states and as well as learn from mistakes made at the state level before you >> create something that's that broadly applies to the entire country. Um I assume you would agree with that kind of assessment. >> I I do agree. I do agree with that assessment and um I don't know that it's because Donald Trump has become more magnanimous. I think that the context is that they cannot continue to dig their heels in and say that there is just simply no political space for frontier safety. Uh and so then if you acknowledge that reality, you then look to who has put something on the table. And though the only people who have put something on the table are the states. So it's a little bit more by necessity than I think by magnanimity. But uh perhaps I'm cynical. >> Amy is part of this dynamic. We know that if you look at the polling, Americans are sort of super skeptical about AI. There's a lot of a sort of apprehension about AI, a lot of skepticism, I'll make people's lives better, a backlash against data centers. Do you think part of the reason states have been more active in this space is that they're uh those governments are closer to um the people sort of hearing from hearing from their constituents about concerns and so are more likely to respond or or am I overreading uh into things? >> Yeah, it's it's interesting. I mean if you just isolate three issues data centers, kid online safety and frontier each of those three have very different dynamics underneath them for sort of who the appropriate regulator probably is or where the appropriate political process should take place. As we have discussed ad nauseium frontier AI safety really ought to be regulated at the federal level with all of the national security uh uh uh uh credentials that the federal government has that states don't. Um, in the case of data centers, um, the the story of America is a story of neighbors wanting to control the future of their own communities. And fundamentally, data centers, it gets all lumped in and people are taking out their anger on AI on data centers, but it's actually really a different thing. It's it's the it's a story as old as time. It's nimi, right? And unlike nimbies on housing where it's just plainly detrimental to stop housing uh in a community, there are some valid reasons to fight against data centers in one's community. I don't know. We don't know need to go into those so much. You know, power and water obviously. Um the the federal government should not do land use policy for the states and cities. that really ought to be devolved to the to the lowest level of government possible because it is closest to people because you're talking about the the look and feel of their hometown of their community and so I think it's entirely appropriate that data center decision-m is devolved uh to municipal levels. I do think that that's entirely appropriate. And then kid safety is it is it yet another interesting one because it really is in my opinion um the truest form of what state governments are there to do which is to protect consumers and protect vulnerable people which is really um which is there's a lot of core competency for states with attorneys general um and so forth. This is this is what states do well. So, I think that kids online safety is actually well served to be done uh at the state level. And I think it's a it's great to have a race to the top on kid safety, frankly. And California, you mentioned in the intro here, California, the state legislature just passed Senate Bill 119. Um and it's on its way to uh the governor now. And um you know I think that I hope he signs it and I think it'll be the most meaningful and robust set of uh child online protections when you when you combine it. So that bill that I just mentioned is all about um prohibitions on what chat bots can and cannot provide in response to children limitations on advertising a lot of other parts but it's about the user experience for kids uh with chat bots. And then there's another bill uh by uh another person Josh Loenthal that's around the addictive components of social media and limiting those addictive elements like persistent notifications bids for engagement uh algorithmic uh advertising and so forth uh recommenders I should say. Um and when you combine those two bills because there's really a false distinction between AI and social media these days. is I mean they're sort of collapsing into one product category. But when you take those two bills together, uh the the Buffy Wix chatbot bill, which says chat bots cannot output things around suicidal ideation or child sexual abuse material or encouragements to to to harm others so forth cannot be outputed and we cannot advertise limitations on persistent memory. We cannot advertise to kids in their in their most vulnerable state. take that and you combine it with some of the protections of the lowenthal bill which are around keeping kids engaged and addicted to their phones and to platforms. Those two things operate as a very very very strong package. Uh and look I hope other states uh also copy those bills. Um and and I think that you know whatever the states do certainly California big progressive blue state is going to be stronger than whatever Congress would do on kids online safety. And I think that's a great dynamic. So that that segus into sort of where I want to conclude this conversation, which is we spent a lot of time talking about sort of what happened in California in the past. Uh a little bit about sort of where California is going, especially when it comes to AI policy. So you've mentioned, right, like we've talked a lot about frontier regulation. There there have also been many other AI related bills that have passed and become law in California. And now in this latest legislative um session, there are there are there continue to be, you know, a number of bills related to AI. There >> 24 AI bills went to the governor just this year in California. Yeah. >> Um so so we've talked a little about uh some children safety bills. My understanding is that there are also multiple bills related to data centers trying to take action to prevent data center buildouts from having negative impacts on consumers over things like say water and electricity prices. Um I think um you know uh another thing that we've seen is that um as as the state legislative process unfolded and various states built on top of SB53, one of the things we saw uh in the Illinois bill was um you know uh requirements related to independent verification organizations. you you hinted at this earlier when you you talked about um auditing. Um this is an issue we've been tracking closely. We recently had Brie Tree on the podcast whose organization pioneered this concept. And so one of the bills in this latest session was uh SB813 um which um creates a framework for selecting and regulating IBOS in the state. And so I'm curious about your take about this requirement and also what it signals about how California might continue to think about AI. Do you think that it is going to continue to regularly update these laws? And in fact, might that become faster as we see AI technology sort of progressing and and developing faster and faster over time? >> Sure. There's a lot lot in there in that question, so I'll try to answer most of the parts of it. Um I interrupted you and said 24 AI bills go to the governor this year. And you know, to a lot of um people in the tech industry, the answer is probably, see, that's too many. We can't possibly keep up. And you know, I'm sympathetic to that argument. Um I get it that, you know, the kids safety bill that I just detailed actually replaces, repeals and replaces another kid safety bill that was just passed last year. So I I am sympathetic to an extent to that argument. And why do I say to an extent? Um, you know, the the the major Frontier labs are releasing a new model every 3 to 6 months on average. The car companies release a new model generation every 3 to 6 years with all kinds of new bells and whistles and technology that the National Transportation Highway Safety Board has to uh analyze for safety. The product cycles are just so much slower in traditional regulation than they are in AI. So you would expect in fact we would be derelict in our duty as policy makers if we weren't also keeping up with the velocity of the deployment because every single new model has completely different capabilities. So while yeah I get it. It's hard when laws are changing every year and stateto state they're different. I I I have sympathy for that. And on the other hand please technology companies have sympathy for regulators who are dealing with a deployment velocity that has no parallel in any other sort of consumerf facing or consumer available product. So I think go sympathy a little bit both ways um behooves uh everyone in this debate. Um so uh you asked about independent verification organizations and I should just um say so that I say it that I've done some advising for Fathom um and it's a great organization and I'm proud to work with them and if you haven't listened uh to uh Bree's episode with a lo I would suggest you go back and listen to it from about a month ago. She does a much better job explaining IVOS than I ever possibly could. But the basic gist is uh starts with a question. Do we trust or are we comfortable with the AI frontier AI labs grading their own homework on their safety plans? Uh what how what what capabilities their models have and what guard railing they've put in place uh to prevent against bad things happening. Uh are we comfortable with the companies checking their own homework? If you're comfortable with the companies checking their own homework, I think we just disagree on this first fork. But if you're not comfortable with it, then the question is who checks the homework? And once you start down that path, there's really kind of only two uh plausible responses. The first is that a government checks the frontier AI labs safety card. I think that's really problematic. And I say this as a guy who has spent 22 years in government and politics. governments are are not necessar certainly subnational governments are not capable of keeping up with the science uh recruiting the the very best and paying the very best frontier AI engineers. Uh I think it's really problematic to have the government from a um from sort of a first amendment perspective deciding uh what models do and don't get to be released. So I just think that's extremely problematic to have the government do it. Well, if not the government then whom? And that's Fathom's point is there's kind of only one logical conclusion here, which is you need independent verifiers. That's sort of if you walk down the logic chain, it's kind of where you end. And okay, so independent verifiers, well, what does that even mean? What is the definition of who can serve as an independent verification organization? What does an independent verification even consist of? And that's where SB813, the bill that just passed the California legislature this year, comes in. It begins to set up the ecosystem for IBOS for independent verification organizations. It itself SBA13 is not a requirement that OpenAI or Anthropic or Google or anybody submit themselves to an independent verification at any particular interval. It's setting up the plumbing. It's setting up the substructure for what eventually could become uh a robust national independent verification uh requirement somewhere in some law somewhere. But that law doesn't exist anywhere yet. this is setting up this is the precondition to to those IBO requirements being uh applied somewhere. >> It's also an example of uh sort of two-way communication, right? So, we saw other states being inspired by SB SB53. In this case, it looks like California is taking inspiration from some state things that have been happening in other states. >> Yeah, I think that's certainly true. Um it's, you know, imitation, sincerest form of flattery and so forth. Um, and I think that it's great when, as we've discussed, uh, previously in this discussion, uh, when states are casing each other's joints and taking good ideas and building upon good ideas, I think that that's totally healthy. Um, and in the case of IVOS, uh, California is the first place to start to develop this IVO substructure. Uh, Illinois had an audit requirement and as we discussed a few minutes ago, um, audits are fine. IVOs are much better and much more important to pursue. Uh so last question, you hinted at this earlier. You said um you hope that um SB53 will be reconsidered at some point. Um even though it's a relatively recent bill, we've certainly seen many many things change uh in the AI space since that's been passed. Um do you think that SB53 is something that will be taken up by the legislature or or a replacement bill be considered by the legislature next year? Uh in >> I hope I I hope so. I I hope it's not a replacement because I think the uh structure of SB53 has caught on and I wouldn't want to start over from scratch with an entirely new model. SP53 is basically trust but verify, requiring a whole bunch of disclosure and then testing that disclosure to make sure that it's it's valid. So, I hope it's not a total replacement, but I do hope that it's revisited. Uh you know in the very beginning of this conversation you asked me um about you know what distinguishes California from the federal process and I mentioned the timeline of the California state legislature um mythos and then hugging phase uh happened too close to the end of our session and unlike Congress where you can put a bill in go through markup and kind of roll whenever you want um the hugging face incident and then the subsequent sort of reports from anthropic and all that um There wasn't, I think, enough time to truly get that deep into the bone marrow of the California policymaking process. What does this mean? And what should we do to change our laws in response to these incidents? Uh, and look, we're now the legislature is now out of session until uh, basically January. And you know, I am I am not a betting man, but I think it's a pretty fair guess that something is going to happen with agent swarms or loss of control or god forbid something, you know, more uh impactful to everyday life between now and January. Just the cadence of these sort of news reports is such that something else is going to happen. So, whatever I predict today on as we record September 2nd, uh, is probably going to be out of date by the time bills start to get put in in January because something else will happen. Hey, maybe a boy can dream. The Trump administration puts together a federal uh, frontier AI framework that's not a secret that they actually tell people about. Um, and that would be that would be awesome. Um, maybe that changes the dynamic. Maybe there's deeper, greater cooperation between some of the frontier labs where some things that are not in the political Overton window become possible because the companies agree to do them. There's just so much that can happen between now and January that I will answer your question directly. Yes, I hope SB53 is revisited. But by doing what? I I don't pretend to know because this stuff is changing so quickly. >> Yeah, I think making predictions in AI policy is often a fool's game. Um, >> but I I think that's a good place to to leave it. Um this was a really fascinating conversation. Um I think it provided insight um that we often don't hear about about what is happening at the state level, how state politics interacts with national politics and what what the significance of what's happening at the state level is in terms of uh setting the conversation both for other parts of the country and also uh at the national level. So, um, this was a really fascinating uh, uh, really fascinating conversation and really fascinating insights and I appreciate your time >> and thanks for making space for conversation about subnational AI policy. I know it's not always on the top of everyone's agenda, but I've enjoyed the conversation with you as well. Thank you. >> Yeah, thanks for coming uh, on the podcast. Thanks. >> Thanks for listening to this episode of the AI Policy Podcast. If you enjoyed the show, consider leaving us a five-star review on your favorite podcast platform. [music] We'd also love your feedback on the show. Please email us at AI policy podcast at csis.org. And don't forget to visit our website csis.org for the Wadwani AI Center's latest research and events. This podcast was produced by Sarah Baker and Nicole Herrera. See you next week.