In this conversation, Matt Perault and Collin McCune take stock of the current AI policy moment.
As AI policy moves beyond rhetoric and into a more consequential phase, Matt and Collin separate signal from noise. They unpack where momentum is building in Washington, how state activity continues to drive the policy environment, and what it all means for Little Tech.
Along the way, they dig into some of the most active debates including proposals focused on protecting kids, workforce disruption, data centers, benchmarking and licensing regimes, and the evolving balance between federal and state action.
Enjoy.
Topics covered:
00:00: The current AI policy moment
02:31: The White House National AI Framework: what’s new and what’s next
10:35: Kids, AI access, and the case against bans
16:35: Data centers, communities, and energy policy
18:42: Workforce disruption, retraining, and labor policy
24:18: Copyright, censorship, and other key debates
25:41: The Democrat perspective and response
31:13: Benchmarking, testing, and startup access
32:09: Licensing regimes and regulatory capture risks
37:02: What’s next in Congress?
42:46: States at the center of AI policymaking
47:08: Preemption, federalism, and the state-federal divide
54:13: Dormant Commerce Clause implications
57:45: Why Little Tech needs to stay engaged now
This transcript has been edited lightly for readability.
Collin McCune (00:00)
AI policy is moving faster than any other policy set that I’ve ever seen.
And now we have moved into more meat on the bone. We’re starting to have real conversations within committees about what could go in these packages. People want to protect the American public, which is ultimately the goal of these legislative proposals.
Now you’ve got 10 companies, the 10 biggest labs in the country or maybe in the world that are coming up with their own benchmarking ideas. They have to be blessed by a regulator. So if you have an enterprising young startup that is also trying to build, they can basically shape the rules such that no one else is going to be able to compete. That is how the regulatory capture game works.
We need people to be active on this and share ideas and concepts, especially from the Little Tech community. The time is now.
Matt Perault (00:54)
Collin, welcome back to the AI Policy Brief. It has been a while.
I think we lasted this in September and it’s hard to think back to exactly where we were in AI policy at that moment. So I think over the course of the summer states were moving toward the tail end of their legislative sessions. We had a couple of bills that I think had passed the legislature in California or awaiting signature, passed the legislature in New York awaiting signature. So we were really focused on state stuff. And then we had been through one round of trying to see if Congress could move forward with some kind of preemption and that had failed over the summer in July. There were a couple more ahead in the fall. A lot has happened since that moment in September. Where do you think we are now relative to where we were then?
Collin McCune (01:38)
Yeah, it’s funny. Everyone’s always complaining about the law not following as quickly as the tech does. And I feel like that’s probably still the case. The zeitgeist is absolutely like, we don’t have enough AI policy out there. I mean, I would say you probably agree with this from my perspective, AI policy is moving faster than any other policy set that I’ve ever seen.
Matt Perault (01:57)
It’s moving incredibly fast.
Collin McCune (01:59)
There’s an avalanche of bills at the state level, an avalanche of bills at the federal level.
Matt Perault (02:03)
Yeah, well, bills in both places. And then I think the thing that I have kind of underweighted, even in the last couple of weeks, is we’re seeing real legislation at the state level. We have check up bills in Oregon and Washington that have been enacted to law. And then like serious proposals that are really moving forward. So I agree, though, it’s not just the ideation, it’s the actual policy making that’s really moving forward at a rapid pace.
Collin McCune (2:10)
Yeah, I also think it’s worth, you know, just mentioning, we talk about this often, but maybe the rest of the people who are building this space don’t think about this. But we’re talking about thousands of bills that have been introduced at the state level. Hundreds of bills that have been introduced at the federal level. Now, not all of those will be enacted into law, but some of them will. I think that that’s generally positive, but there are obviously some downsides to that.
Matt Perault (02:49)
Positive that things are moving.
Collin McCune (02:51)
Yeah, I think it’s positive that lawmakers are thinking about this and it’s very front of mind. I mean, I think that there are certainly real issues with any technology that need to be taken seriously. So I think it’s positive. There’s also downsides to it because I don’t think that some of these policy sets are well thought through. We also don’t know where the technology is headed and where it’s going, which obviously we’ll get into today. Maybe just a little bit, you know, sort of back to your question. I would say things are progressing nicely.
And I would say, especially at the federal level, one of the principal things that you need for any piece of legislation to advance is you need to have White House buy-in in this day and age. Back, I think 20, 30 years ago, it wasn’t as important to have executive buy-in before the Congress was moving legislation. Nowadays, the executive sets a lot of sort of the overarching policy and then Congress fills in the gaps. And I think it’s really important.
Matt Perault (03:48)
So can you talk about the shift there? Cause I would say, like to most of the general public, including people who follow AI policy fairly closely, I think they would say the White House has been active on AI policy for a long period of time and vocal on it for a long period of time. But I do think there’s been a shift in the last couple of months from what you’re saying in terms of the real policy making component of that. Can you talk about what that shift is?
Collin McCune (04:11)
Yeah, look, and I would say at the front of the President deserves a lot of credit for leadership here, right? The President, the President’s team, Kratzios, David Sacks, Sriram, our former colleague, there are a number of people that deserve a lot of credit for advancing the ball, I think in a very intelligent way. Yeah. And, you know, it started really with the National AI Action Plan, which was kind of a six-month effort at the beginning of the administration to put forth a whole of government approach on how we’re going to attack AI. And I thought that was really, really important. But that was step one.
Step two was sort of the Genesis mission, which I think built in more on the infrastructure side of things, energy, data centers, et cetera. Research, right? It’s also become, I think, very apparent that the federal government needs to act from a legislative perspective. And those proposals early on did not have specifics on what should go in a legislative.
So now version three of this is the President and the White House along with I think collaboration with the executive agencies coming forward with proposals, real granular proposals on what should go in the legislative framework, almost like recommendations.
Matt Perault (05:25)
Yes, so that was flowing out of the December executive order. The December executive order said in 180 days, the White House will issue some guidance for Congress.
You understand in a way that I do not how a bill becomes a law. So just from the theory of how it works, President enforces the law, legislative branch writes the law. So I would have thought coming into this, this is on Congress. Congress’s job is to come up with what the legislative framework looks like. I feel like we heard in lots of meetings we did last spring, we consistently heard here, like, we want to see the White House weigh in here. So can you talk about that? Like, why was it important for Congress to have the White House not just say, we want to see AI legislation or, you know, Congress needs to play an assertive role here vis-a-vis the states. Why was it important that they actually come up with a framework for what legislation might look like?
Collin McCune (06:17)
Yeah, well, it’s still the Congress’s job. And the Congress has, I think, been very active on this. Back to our point, and probably the September pod that we did, obviously there’s been a lot of movement in the Congress. There has been the moratorium that went through the reconciliation bill. There’s been another conversation about some level of preemption, which we should talk about. So it’s been active. And I think there have been a number of different committee hearings. And there are certainly legislative proposals out there.
But it’s helpful on the front end to understand where the President of the United States stands on some of these things. So A, you don’t get all the way through the process and then get a veto once it gets to him. But also the executive and the executive agencies have to be able to implement the legislation. So legislation is like a framework and then the agencies take it and they put even more detail in it and they implement it that becomes really the enforceable law.
As you kind of just noted, it’s just really, really important to understand where the President is. A, for the reasons I just laid out, and B, also sort of the politics of it all, right? The president is the leader of one political party in this town. The president needs to be able to sort of guide that political party on a number, on a range of topics, especially something as important as AI. So it’s a really important signal to Republicans, hey, you’re heading in the right direction here, you’re not heading in the right direction here, and then go in and fill in the details. It will then come to me and we’ll sign it. That’s the general idea.
Matt Perault (07:47)
Let’s dig in a little bit on the proposal. I’d like to sort of get at three parts of it. One is just what we think it says, what it says and how we read it. And then I’d like to sort of see if we can think about how the minority receives it? How would Democrats look at this? What would they like in this and what would they see is insufficient? And then how has it been received? And we can talk about maybe how it’s been received publicly, but I think it’s more interesting for people to hear how you think it’s been received in the political process, like how it has been received on the Hill. So just like start maybe as a starting point, like what are the elements in this that we think are meaningful.
Collin McCune (08:21)
Well, I should be asking you, really, shouldn’t I? Look, I mean, look, I think the White House and the President have been very careful to craft this around what they call the four C’s, right, which I think is actually a really smart framing. And the four C’s are are geared towards addressing the things that we that have arisen as problems from the general public and a number of the stakeholders over the last year and a half of real conversation.
Four C’s are children, communities, copyright, and censorship. And honestly, I think each of those issues has been around, I think, well before AI. But I think AI has the ability to sort of supercharge each of them in a way that we haven’t seen before. So I think it would be actually helpful to hear you kind of talk through a couple of them in terms of just from a policy set and then we can kind of get back into the politics.
Matt Perault (09:20)
So let’s start with what we’re seeing here. We can talk about things that we’re not seeing. But I think as a starting point, the cornerstone of our approach on policy, which I wish I could take credit for because I love it. But it really is you deserve it. It’s really you and what Marc and Ben and Jai established before I arrived at the firm, but is anchored around focusing on regulating use, not development.
And that means trying to ensure that when people are harmed in different ways, that when AI is used to harm people, that the people who are perpetrating the harm can be held to account. And there are elements in here that do that. I think that principle is very expansive because you can use it in a consumer protection context, a civil rights context, an antitrust context, criminal law context. So there’s a million things that could be in here related to that idea. But I think we see in here some elements of it.
There’s a focus on the use of AI for fraud, for instance. And I think when we see that, I get excited because so much of the framing of so many of the policy proposals we see are paperwork proposals. Correct. Like we need you to conduct an impact assessment. We need you to do an audit. And those things are the things that we think really burden startups and burden startups without, I think, a lot of benefits on the other side.
Collin McCune (10:42)
It probably doesn’t keep people safe…
Matt Perault (10:48)
Is an impact assessment likely to prevent fraud? Maybe in some circumstances, but I think the chances are pretty small. And when I think of when I’m moving through the world and I am a potential victim, what is the thing that can give me some comfort that either I won’t be harmed, that companies will be engaging in preventative activity on their own to ensure that they’re not incurring criminal liability or in the event that I am harmed, that there’s recourse.
And we primarily focus, how we think about AI policy, but also just how the law thinks about it, we think about that in terms of penalizing harmful use. Seeing those elements in here was really encouraging and exciting. And there aren’t a lot of development-oriented proposals. So I think that’s really critical. I can just tick through a couple of the other top line things that we see. We’ve spent a lot of time thinking about what the right approach is on kids, or what a right approach is on kids, it’s such a hard issue.
Collin McCune (11:40)
And it’s been around for a very long time, kids online safety.
Matt Perault (11:44)
It’s not specific to AI. And there are really challenging dynamics there. It’s important that we try to minimize harm to kids. It’s important that we give parents an active role in protecting kids and making decisions on behalf of their own minors. It’s also important that kids learn how to use AI tools. And so the idea that if you’re under 18, you’re not going to have access to AI tools, that’s not going to help kids prepare for an AI economy, it’s going to cut them off from important information. And there are Republican concerns about that information access and there are Democrat concerns about that information access. Kids have privacy rights too. And so striking this right balance I think is really challenging. And there are, I don’t think there are ultimate solutions necessarily in the framework, but there are ideas that move in that direction that I think are constructive starting point.
Collin McCune (12:36)
It’s very interesting. Actually, maybe we should just talk about this a little bit more. Some of the proposals that are out there that I think are really challenging, to your point, are banning children’s use under the age of 18. And I, you know, it’s tough, right? It’s a tough conversation because I think we’re now seeing things like Alpha School, for instance, who have kind of reinvented the way that children are learning in the classroom. And I think what we found is that it’s really important for kids to have playtime.
It’s really important for them to build social skills. And it’s really important actually for them to learn at their own pace. When you have a teacher and 30 kids in a classroom and everyone’s learning arithmetic at the exact same pace, it’s very likely that some of these kids are not going to be able to keep up. And if you now can shift that model to something like Alpha School, where you a personalized AI tutor in conjunction with a teacher in a classroom, some kids can learn, and the test scores are basically supporting this, they can learn for two and a half or three hours a day and the rest of the time they can be building social skills or playing or being a kid while other kids need to spend an hour in the morning and an hour in the afternoon, hour in the evening learning because that’s how they learn better. If you were to ban that, it seems like it would stunt the growth for all the reasons you’ve outlined but also certainly from an educational perspective. I don’t know, how do you think about that?
Matt Perault (14:02)
I think bans are something that you do when the overwhelming majority of use cases are negative. And I think there are a range of different views about the value proposition behind AI. We obviously are incredible optimists who really think people’s lives will be fundamentally altered in ways that are net positive, net incredibly positive by AI. If you have our view, then clearly a ban means you’re taking enormous benefits off the table to try to address concerns and harms that we acknowledge exist, but the consequences in terms of the benefits you’re removing are not justified by the harms that you’re preventing.
And I think even a less optimistic view than ours, I think most people can recognize that AI is going to have really positive transformative impacts. And that even if you’re skeptical of some of those, we’re moving to a world where AI, knowing how to use AI, and being AI literate is gonna be an important part of how people do their jobs, run their lives, and interact with other people. And so the idea that you would not start to develop those tools until after you turn 18, I think is really problematic.
It was encouraging over the summer, I think summer, early fall, probably around the time we talked last, Governor Newsom vetoed a bill that arguably could be described as kid safety in California because he thought it was going to cut kids off from important ability to access the tools in ways that are meaningful and positive. And he also simultaneously signed another bill that he thought was going to be a narrower way to address potential harms without getting in the way of some of those benefits.
So I think the idea of bans is really probably not the right way to think about. But we worked really hard to make sure that the way that we think about policy is responsive to the concerns parents have, that lawmakers have…it’s something we hear all the time on the Hill.
Collin McCune (15:46)
It’s really challenging. There need to be real rules and restrictions. We have to keep kids safe, there’s no question about it.
The only other thing I would say on this, and if you have additional, we can do that, but the new political vector of attack for some of these really, really bad ideas is through kids’ bills. They’re using kids’ bills almost as a shell so that they can dump in a lot of ban-oriented things that look very innocuous but are actually very, very difficult and challenging, not only to implement, but I think also would be bad and negative through the lens of kids bills.
Matt Perault (16:30)
Yeah, it’s frustrating to see that in part because, again, you understand the political process, right? I sort of focus on, what’s the policy that we feel like we could get behind, like we’re looking for stuff we can get behind, and there have even some of the most there are sometimes bills that we see that are really equivalent to bans, really problematic. And then they’ll also include provisions like it should be a violation of criminal law to solicit a minor for unlawful sex, which is, I think already prohibited by criminal law, but like to the extent there’s a loophole, to the extent it’s not sufficiently covered, is definitely something that should be in place. To your point, those are the kinds of things that we can get behind. It’s harm-based, it’s targeted at protecting kids, it’s making sure that if, when you’re not using AI, you’d be held accountable criminally. Then when you are using AI, you should also be held liable criminally. We support those concepts but to include alongside them a broad ban of kids access is problematic and makes those kinds of legislative proposals something we couldn’t get behind.
All right, what other parts of the framework should we touch on? Okay, all right. So, in the community section, there’s one of the things that we’ve done some work on energy. The President developed the Rate Payer Protection Pledge, which most of the hyperscalers have signed onto. And the basic idea is like, they’re bringing their own power and water to data center projects. Like they’re not asking other people to pay for their costs.
We have thought that’s important for a Little Tech because if you’re a startup and there are sort of sweetheart deals that can go to large companies and you’re left out of those deals and you’re left holding the bag from a cost perspective or the hyperscalers are getting kind of preferential access to energy to compute, that’s gonna be problematic for Little Tech. And so I think the President’s steps there are really positive from our standpoint.
Collin McCune (18:23)
Incredibly positive. And look, I think that that should be an obvious thing. You can’t build a data center that also can explode energy prices and explode water prices for a community. That’s not realistic. And I think the companies rightfully acknowledge that. So I think that that’s less controversial. I think that that is exactly the way to solve this.
The other thing here is, just from a messaging and a political point of view, it really does seem like data centers and the companies behind the data centers and I look, I think this would apply to us, have done a quite a poor job of actually messaging about the benefits of these things, right? Certainly there are some downsides, we’ve all seen the videos of low humming noises and things like that. Those need to be fixed and things need to be solved there. So I’m not suggesting that, but we’re talking about a real tax base increase, unlike these communities have ever seen. So now all of a sudden you have the best schools, you have the best sidewalks, you have the best public transportation.
That’s really hugely beneficial. I don’t think it’s being communicated effectively. That’s something that I think we as an industry collectively need to take on as a challenge.
Matt Perault (19:29)
Okay, so some other elements in the proposal are the workforce proposals, right? That’s been something that I’ve long been excited by generally, and wish that there’d be more legislative momentum in that area. Because I think Republicans and Democrats, well, they have like different constituencies in mind, when they think about someone who might be displaced by a worker who might be displaced by AI. And I should say, at the outset, like our view is that AI is going to have a significant positive effect on the labor market, not negative effect. That’s consistent with the history of the introduction of technologies. We’re always wary of disruptions and there are clearly disruptions, but overall typically there are gains for labor, not costs. But even if you have that positive view, there are workers who will be disrupted and those workers should be supported in their transitions. They should ideally be re-skilled into jobs that are jobs that will employ them in an AI economy and to the extent they’re unable to find jobs, they should be supported during the period of unemployment. And my sense is that Republicans believe that, Democrats believe that, there’s a lot of consensus behind it.
There are a whole bunch of different concepts from unemployment insurance to re-skilling programs to just things that don’t require any expenditures, just how we track disruptions in labor markets so that we can try to see them early on, see those disruptions. So the way that the framework is structured in that section, I think, is at least a good starting point. I don’t think it’s probably the end of the road, but a starting point for thinking about how we support workers.
Collin McCune (21:07)
Yeah, I agree. I think the big challenge with this is it’s been very difficult with everyone that I’ve talked to to pinpoint exactly where we’re going to see massive disruption, because the disruption hasn’t really happened yet. Yeah, there have been a couple, the infamous Block announcement where they slash 50 % of the workforce. And, I think that there’s a number of factors that went into that, obviously, the justification was AI, but I think there’s more to that story. But, the World Economic Forum came out with an article earlier this year that basically suggested there’s 1.3 million new jobs based on AI over the last two years, which is a pretty sizable explosion. Not to mention later on in that article, it’s basically like a three to one job gain versus loss in the AI market.
I think it’s really tough to quantify. We’ve talked about, and Marc talks about the lump of labor fallacy all the time, the idea that the pie is fixed and if you remove a slice of the pie, then there’s a collapse of the market, but that’s not necessarily the case because the pie we believe is going to grow. My comments aren’t to suggest that it might not be a real issue. It’s just really hard to pinpoint how we should attack this. That being said, there are a lot of really smart people that I think, inside and outside government, that care about this and are working on this.
You know, the one area that I think we should be focusing on is retraining because there are certainly going to be, you know, call centers seem to be one that comes up frequently. It seems to me like call centers are going to be an area in which AI could easily disrupt and probably disrupt in the next 12 months or so. Okay, so now we’re going to have a certain number of people that are going to be displaced from a job. How can we organize them into something new and hopefully in an area where I think AI is exploding and creating new jobs?
Matt Perault (23:03)
Yeah, I think regardless of the statistics, now we’re going to switch roles a little bit, the thing that you constantly urge me to do in my work is like, it’s not just what’s on the page, it’s not just what the data says, it’s what is it like when you’re sitting across from someone who says that they lost their job? When you’re hearing from the lawmaker who’s saying, I hear what you’re saying in terms of statistics, but I’m going to town halls and in the town halls, this is what I’m hearing. People are anxious. I think there’s the reality of the disruption we have to deal with.
But then there’s the perception of the disruption, which I think is also a big thing. It may be that you don’t lose your job because of AI, but you’re nervous because of losing your job because of AI. And I think if the policy solutions were extremely high cost, very disruptive, challenging for Little Tech to implement, challenging for the government to implement, then maybe we would want to have some caution and concerns about them. But I think there’s a lot of low hanging fruit.
So I agree. It’s great to see the White House doing this. I thought in the action plan the previous guidance from the White House that was more focused on executive branch activity. There was a ton of stuff in there that was really positive. I think if you had just cut it out of the action plan and you put it on a blank document and you showed it to Democrats, they would have thought, these are a bunch of Democrat ideas. They were really like little “p” progressive concepts for how we support workers who are disrupted. so this is not a sexy topic, but it’s a topic where there’s a lot of bipartisan consensus.
Also, importantly, a lot of room to really help people.
Collin McCune (24:34)
Yeah, the only thing I would note on that is just we’re also, we’re going to have a conversation, I think, on federal versus states at some point here. I think it’s important to discuss that. But this is also an area where I think states are stepping up and they’re also starting to have real conversations. So the collaboration between federal governments and state governments to try and address this problem with the private sector, is ultimately going to be the all hands on deck approach that we’re going to have to take.
Matt Perault (25:00)
So I’m going to tick through a couple things quickly because I want to get to a different perspective on this. I want to get to the Democrat perspective on this. Importantly, there’s stuff in here on intellectual property rights and signaling clearly the White House’s view that we view as the correct legal determination that most of AI training is fair use and protected by copyright law.
Collin McCune (25:24)
You have written about that a lot, obviously JaI has as well.
Matt Perault (25:27)
And then some stuff on censorship that I think actually has more bipartisan appeal than you’d think on its face. They’re often Republicans who are pushing the censorship-related concepts out of their concern of a bias against conservative ideas. I think there are principles in here that also Democrats might use their ideas here related to jawboning for instance, and that’s something that is of concern to Democrats as well. So I actually think this is kind of grounded in a First Amendment approach to speech issues, which I think is really positive. Then there is, well, maybe we should put a placeholder on the preemption stuff and talk about that in more detail in a few minutes.
So we’ve been through what’s in the document and I think sort of talked about the White House’s perspective. If you’re a Democrat on the Hill and you are reading this, you’re receiving this, are there things in here that you like? Are there principles in here that you think are, you know, things that are offered to Democrats as a way to hopefully bring them to the negotiating table. And what are you not seeing here that would be important at the negotiating table to see that gets you to feel more comfortable with the concept of some sort of congressional action in AI?
Collin McCune (26:44)
Yeah, it’s a very important question. Three different categories in which I would evaluate this, and this is what I’ve heard from Democratic members of Congress.
One is we agree with the overall framing with some of these categories, but there’s not enough substance here for us to evaluate. Which, you know, I think is, you know, somewhat fair criticism, but that’s also Congress’s job. Like we talked about, Congress needs to fill in some of the gaps that have been in this general framework.
Matt Perault (27:15)
Is it a fair criticism? I’ve heard that and it’s hard for me to, I mean, it’s not a legislative proposal, but it wasn’t intended to be a model bill.
Collin McCune (27:25)
It wasn’t. It wasn’t at all. I think it’s some of the gamesmanship that you would see from any sort of cross party negotiation. But look, I think that there is enough meat on the bone to get started. And that was exactly what it was intended to be. But if you’re asking them, OK, is this something that you agree with? You know, the answer is reflexively going to be I need more detail. And I think that that’s effectively what we’ve seen.
The second category here is they want to see something like an SB 53 safety framework in here, a quote-unquote safety framework for all the reasons that we talked about at the top, you know there these disclosure frameworks that look seem fairly innocuous, these safety reports that must be put out, these transparency standards, etc. They all sound really good. But in practice are they actually keeping people safe? You know they in fact create sort of a regulatory barrier because the people with the most lawyers are able to comply, which is why we’ve been critical of this from a Little Tech perspective for a long time. You can talk about this eloquently.
And loo, the third category, which I think is probably the most important political perspective, is that anything Donald Trump does they’re gonna dislike.
Matt Perault (28:37)
So I hear that as being fairly bleak in that it wasn’t going to be a longer document by design. And I think in a helpful way, it’s short because, as you’re saying, this isn’t the president’s job, this is Congress’s job. So you don’t want the president to be dictating what the actual text is. You want it to be conceptual.
The second category are things we would strongly oppose. Cause we think it’s going to make it really hard for a Little Tech to compete. If we, if we get all the things in your second category, then we will have very concentrated AI markets where the startups will be at a significant disadvantage.
And then the third is, I don’t know, impossible to deal with in terms of really trying to get a framework to the table.
And so I guess part of what I’m wondering is like, what are the things, if we can envision a final package, what are the kinds of things that you could add to this that would get Democrats to yes? I think there’s probably a lot of stuff in the regulate harmful use category that appeals to Democrats. Like enforcement of civil rights law. That’s a bipartisan concept because Democrats think about equity issues, but enforcement of civil rights law includes First Amendment violations, which Republicans care about. So making clear that when AI is used to violate a person’s civil rights, I think that’s appealing to Democrats. Democrats have concerns about enforcement agencies not having the resources they need, whether that’s the Justice Department, if you’re talking about civil rights specifically, it’s the civil rights division, the Federal Trade Commission is going to be enforcing consumer protection law.
So I wonder if that’s something that maybe gets Democrats a little bit more excited about some of the ideas here? Is there anything else that you think could be on the table for Democrats that is compelling?
Collin McCune (30:25)
I think all of those things are germane. I think the other piece, my mind goes to the Kids Online Safety Act. It’s been a major bipartisan bill for a very long time. Obviously originally geared towards more social media, but is, I think, also applicable in the age of AI. I think there are some problems with it, but both Republicans and Democrats have acknowledged the fact that this needs to move forward in some way, shape or form. So if there is a strong form of KOSA that moves forward and gets attached to this bill, that’s also something that would get them excited about voting for this.
Collin McCune (31:08)
Yeah, and the AI Safety Institute, has now been sort of has been changed into CAISI, which is effectively a very similar mission, but also now with an innovation backbone to it, codifying that and making sure that it’s permanent, I think is something that they’ve been advocating for for quite a bit of time. And I think that that is something that is relatively non-controversial at this point and something that could get done.
Matt Perault (31:37)
Everyone seems to like CAISI, which I think is interesting because not everyone liked the Safety Institute. Everyone seems to like CAISI, everyone likes it but it seems like there’s not a lot of content now on what it is. And so we heard some Democrats talk about ideas of CAISI being used for cyber security capabilities, testing and benchmarking, which I think is an interesting idea. We’ve done some thinking internally on how you could make benchmarking resources available to startups and then have some kind of regulatory mechanism where startups and larger companies are testing their AI models against the benchmarks that CAISI offers for instance, or I think ideally it wouldn’t be one benchmark. We’d have concerns about regulatory capture if CAISI or NIST is publishing just one benchmark that startups have to publish a test against. But if there’s a plethora of benchmarks and startups are able in a reasonable cost way to test their models against, then consumers can make determinations about what models they think are safe and make decisions about what they use. And the government could use it for procurement, CAISI, NIST could write reports about the benchmarking industry, they could write reports about what benchmarks are a little better, a little bit worse. I think there’s something there that might be workable.
Collin McCune (33:00)
Now, we should also talk about the ideas that probably would not end up in a bill that is likely to pass, but are still democratic ideas that they would want in there. Licensing regimes are one thing. Creating an entirely new agency that effectively centralizes all AI policy. And before a model is released, the public must go to this agency. And the agency must bless it. And then it can be released. We should talk about, let’s talk about the downsides of that.
Matt Perault (33:32)
We should also talk about, when people hear us talk about this, they think…”I’ve never heard any Democrats propose a licensing regime…” And that was proposed on the Hill a couple of years ago. You had companies coming out in favor of it. You had an administration in support of it. But that aside, I think it’s really important to talk in specific ways about what we hear as licensing, even if other people in the ecosystem hear it really differently.
There are a lot of policy proposals that when we hear people talk about them with us, we hear licensing. One of them is a permutation of this benchmark concept. So I think if CAISI or NIST puts out a benchmark and if you build at the frontier, you are required to test against this benchmark. And if you don’t hit a certain score, you cannot release your model into the world. We hear that as licensing. Because that is permission to build, permission to release that is based on the government blessing a particular metric or particular performance. And I think in our view right now, I think we have faith that if CAISI came up with a metric, that would be workable. But I think we have concerns that if you make CAISI the bottleneck for that, that will lead to regulatory capture over time. It will be a small group of companies who will shape what that benchmark looks like. They’ll shape it in a favorable way. It will be easier for them to test against it and easier for them to pass that test. And again, I think a lot of people don’t think of that as licensing. I think it’s important to say that that’s how we hear it.
Collin McCune (35:00)
Well, in addition to that, what about, almost like a self regulatory organization, like an SRO. Where you have, now you’ve got 10 companies, the 10 biggest labs in the country, or maybe in the world that are coming up with their own benchmarking ideas. They have to be blessed by a regulator. Maybe it’s CAISI, maybe it’s NIST, maybe it’s, you know, pick your agents, maybe it’s commerce. That has to then bless those. We do this a lot actually in the financial services space.
Then those top 10 companies are basically just deciding exactly what they want. So if you have an enterprising young startup that is trying to also build, they can basically shape the rules to such that no one else is going to be able to compete. That is how the regulatory capture game works.
Matt Perault (35:46)
So there’s a version of this proposal that we hear repeatedly, which is, it’s just going to be voluntary. Don’t worry about it. It’s not regulatory, just voluntary. And so you can choose to participate, but you don’t have to. Any voluntary participant gets some benefit by going through this regime. So you have to hit a certain performance benchmark or you have to get certified by a verification organization. Once you’re certified, it’s not that there’s a penalty. Once you’re certified, you get this benefit. You get some liability protection. And there are different ways that people think about what liability protection might be. You can think of the carrot in a lot of different ways. And then you can think of the verification system, the benchmarking, in a bunch of different ways. But I think it’s important for us to be clear. We hear this as also a form of licensing.
That’s probably not exactly the right term for it, but we hear that as an approach that would be really bad for startups because let’s say that the verification process is disproportionately hard on Little Tech relative to Big Tech. If they can’t do it without some competitive disadvantage, then even if it’s purely voluntary, if the carrot’s a really good one, like a really meaningful one, like some amount of liability protection, and they’re disadvantaged in going through that process to get that carrot, then they are at a significant disadvantage in that situation.
Collin McCune (37:02)
I think it’s a really important point and one that’s not understood. I go back to this. We’ve mentioned this a couple of times, but it’s a really important point to drive home. There are these bills and these shells and these frameworks out there that look very innocuous. They look very common sense. Yeah, but they are wolves in sheep’s clothing. And this is the way that people sort of use these as beachheads to get things through. And ultimately they’re very, very consequential, not only for American consumers, but certainly for startups, the smallest of the small.
Matt Perault (37:37)
I want to get to state stuff in a minute, but before we do, let’s ask the big question, which is your world, is this gonna get done?
Collin McCune (37:45)
I think it will get done in all of legislating, right? You go through the phase where people are just throwing ideas at the wall. You’re throwing ideas, you’re throwing ideas, you’re throwing ideas. And I think the last year and a half has effectively been that. There’s been a couple of real attempts. There’s even the moratorium play, which I think ultimately was the right thing to do, but at the time was completely misunderstood. And we’ll talk about that. But it almost got done at the end of the day. And I think that that would have been sort of a temporary reprieve but ultimately wouldn’t have solved the entire issue we need.
That’s why I think this comprehensive proposal that’s come out from the White House, stuff that we’ve published, that’s why you’re seeing a move towards, okay, we need something that’s bigger than this. But those are all steps in a process, right? And now we have moved into more meat on the bone. We’re starting to have real conversations within committees about what could go in these packages. And above all, I think the politics are pretty good right now. The politics on both sides of the aisle, people want to get things done. People want to protect the American public, which is ultimately the goal of these legislative proposals.
But I will go back to my crappy analogy here, which is legislating is like big wave surfing. I’ve never surfed a big wave, but you spend all of this time putting together a team and surfing the area and you get the jet skis, you gotta raise money and sponsors, and things like that. But ultimately you cannot ride a big wave until the swells and the weather are correct. And politics is the same. Politics makes good policy. Policy doesn’t make good politics. We need a political environment by which we can get something going and get it over the finish line. And I think that we’re fast approaching that. And it could happen between now and the end of the year, certainly.
Matt Perault (39:36)
I think if we do a documentary on how an AI bill becomes law, it will be just as popular as a hundred foot wave.
Collin McCune (39:44)
But it’s all progress, right? I mean it takes a long time, we think about legislating here in 10-year chunks. That is mind-blowing especially in the tech space. But especially if you’re working at a startup, oh my gosh, ten years it takes that long, but that is reality. I mean there are plenty of arguments, both from a philosophical perspective, but also from people who are just big history buffs, on whether or not it was designed to be that way and I think it is, like if you read what the founding fathers were pushing for, it is supposed to be difficult. It’s supposed to take a long time and politics are afoot. We’re also in an election year. Things get really difficult. It feels like the Spider-Man meme. Everyone’s kind of pointing at each other and saying, you didn’t do this, you didn’t do this. But the quiet thing that’s happening behind the scenes is a lot of very, very productive policy conversations. I remain bullish. You could tell me that we’re going to get something done between now and the end of the year, or you could say that it’s going to take another two years, and I would not be surprised either way. But the point is, there’s positive momentum and everyone knows that we need to get something done.
Matt Perault (40:52)
There’s a low pressure system in the North Atlantic starting…
Matt Perault (40:59)
All right. So we’ve talked a lot about DC. But I think it’s arguable and it’s hard to have a counter argument that when you think about AI policy, what’s really happening in AI policy now, it’s principally about state activity. We’re talking about the possibility of a bill becoming a law in DC. In the states, it’s really happening, over a thousand bills, I think the fifteen hundred roughly is the count now at the state level in this session.
Collin McCune (41:27)
Back to back sessions, back to back years.
Matt Perault (41:29)
Yeah, like 1,000 plus last year, 1,500 plus this year. And that’s not just bills that are being introduced. I think last year there were somewhere around the 200 range that were actually enacted into law. This year we’ve seen significant chat bot bills in Oregon, Washington. We’ve seen legislation in Utah. Stuff’s moving in Tennessee. California and New York are always typically a little bit later in terms of what we see actually get enacted. And I’m sure there will be activity there as well.
So how do you size up the dynamic now between what’s happening in DC and what’s happening in state capitals?
Collin McCune (42:03)
Yeah, the way I think about it, it’s two sides of the coin, right? In one sense, it’s really productive. I would say that first and foremost, in all 50 states, there will be an AI bill that’s passed this year, right? And some of them can be relatively low level, I think probably low impact to startups or good governance things. To the extent of what we were talking about earlier, which is banning under 18s from using chatbots, or very, very robust safety frameworks like we have seen in both California and New York that were passed last year.
Matt Perault (42:39)
Yeah. Can I talk about the Tennessee bill? It’s actually has gotten much better. But I think it’s actually just a good starting point to understand where some state lawmakers are starting. The initial draft of this bill that we saw would make it a class A felony in Tennessee. That means murder. Class A felony. That it means like first degree murder, not like negligent manslaughter. You know, it’s first degree murder to use an AI system for an activity that you would require a license for, like practicing law or medicine. Maybe there should be penalties for that, we generally think you shouldn’t be able to use AI in a way that you couldn’t if you were not using AI, but that prohibition should be at parity. But the idea that it would be first degree murder to essentially misrepresent the status of a license feels an extreme starting point.
Collin McCune (43:40)
It feels like a very extreme starting point. And I would say that there’s probably 25 states with something that is almost similar to that, right? So the conversation is muddled because I think there’s a lot of really good work being done by the states, but there’s also some bills out there that are incredibly extreme, that would be crippling to the industry.
And in addition to that, we now have both federal and state data center moratoriums. And make no mistake about it, a moratorium or a ban on building data centers in the United States is effectively a ban on AI, on the development. They can’t exist without compute power.
Matt Perault (44:25)
And particularly the ban component of it will skew toward a harsher penalty on Little Tech. Like the hyperscalers, they will be constrained, it will be problematic for them, but they will likely in a world of constrained supply probably do relatively well.
Collin McCune (44:41)
Right, exactly. And we can talk about, we talked about this on the last podcast, but we can talk about it again, where this political advocacy and this push is coming from. I think what gets reported is much different than reality. In reports, it’s like a small community of people who really care about their community or pushing back on this. And it’s a very coordinated national campaign that has been crippling in a lot of situations, they’ve been quite effective. That’s my general feeling.
This also comes up in the context of the preemption conversation that we’re having, which we should discuss because it’s important. There was a lot of really good work going on in the states. And there’s a lot of really, really unfortunate things that I think are happening that would have really major impacts. So why don’t we talk about preemption?
My perspective is this conversation has been completely lost on what is actually reality. But do you want to talk about the divide?
Matt Perault (45:41)
Actually, I think we should own it, because I think this is a thing we really have believed in. I feel like we have lost it. I was excited about the way that we approached it, and I feel like the general perception has moved against us, not in the direction that we’ve been pushing in. There are a lot of people who have skepticism about the technology, the term moratorium is not one that people are responsive to. So I get that there are things about it that have made our arguments hard to make and that have been headwinds for us.
But I feel like the core concepts here, I don’t know, I am hesitant to use the term unobjectionable, but I feel like the general thrust should be, I would hope would be, one that would be between the 40 yard lines, which is we should have an AI policy, what the constitution says we should have in lawmaking generally, which is there’s Congress, there are state legislatures, they each have an important role, but they’re not equivalent. And Congress has primary responsibility for regulating interstate commerce. States have the ability to police harmful conduct within their jurisdictions. It’s very clear that states have police powers. So we’re not talking about a world where it’s just Congress or just states. Often the rhetoric that we hear is Congress hasn’t acted, so states need to act. And states should act, but consistent with their constitutional authority. We’ve also heard some state lawmakers celebrate when they pass a bill and they say, we’re setting a national standard for AI development, to some extent, will run into legal headwinds. In many cases, that viewpoint, I think, will be held to be unconstitutional because states cannot interfere in interstate commerce in a way that creates a burden that is clearly excessive relative to its in-state benefits.
And so, again, the specifics here are complicated. They’ll play out over a long period of time, but I think it is very important that there be this distinction between state authority and federal authority.
Collin McCune (47:38)
Absolutely. And we should be very clear. The states have a role to play.
When we’re talking about preemption, we are not talking about preempting all state law. We are talking about ensuring that there is both a federal framework that governs the space and certain state frameworks that govern in the space.
It’s the conversation around the moratorium, you know, the meme that was created around it was what Congress was attempting to do is basically make it effectively illegal for any states to come forth with any sort of AI legislation or regulation. And that’s just simply not true.
If the states did not pass a single additional bill from this moment forward, they would have all kinds of authorities to regulate this. And they are actively regulating this.
Matt Perault (48:24)
And so there was an exemption in the moratorium for generally applicable law, which would have meant, you want to punish fraudulent activity using AI, you want to update your criminal law to make sure it can account for AI, it would be fine as long as the law is generally applicable. And typically in public policy, it’s a best practice to not make something tech-specific. Like why would you want your new fraud statute to just single out AI? In theory, you should make it about harmful fraudulent conduct generally and cover AI.
And so even the moratorium would have left a lot of room for states to act. Our view has been that states could go after AI related harm specifically. So our view is not exactly in line with the moratorium. If the state wanted to have not just a generally applicable fraud statute, but a fraud statute related to AI specifically in our view, that’s workable. That’s not a problem for Little Tech. But as you’re saying, the perception of the moratorium was really different.
I think the White House in the federal framework also took some steps to really clarify what this would look like in a way that was positive. So the preemption idea they have is Congress should preempt state AI laws that impose undue burdens to ensure a minimally burdensome national standard. It was focused more on preemption related to development than harmful use. And I think that’s a really positive thing.
And then there’s another bullet that says this national standard should respect key principles of federalism and not preempt. And then goes into a few different things, traditional police powers, as we’ve discussed, state zoning laws, obviously seem like that’s an area where states should continue to act. So that seems like a valuable starting point, I think, for these congressional discussions and making clear that when Congress acts, it needs to do it consistent with its constitutional authority. It can’t interfere with state police powers.
Collin McCune (50:17)
Yeah, I agree with you completely. And I would also just say, just to put it more simply, I think what we’re asking for, and I think what maybe, you know, I’ll speak for the rest of the industry here, we’re asking for is a very similar setup to what every other major industry minus the insurance industry has in the United States, which is we have to have a federal framework and we have to have state framework.
That is what happens in financial services and energy, et cetera. You go all the way down the list. I think it would be a really, really strange thing to have the state of California or the state of Utah govern all financial services in the United States. That’s just not how we’re set up. Would we be in favor of abolishing the SEC and the CFTC and their jurisdiction or the FDIC or the OCC and saying, oh no, we’re just going to let California set the standards for everyone in the United States? I don’t think that’s good government. I think that’s a bipartisan issue by the way. The more and more that we’re talking to people, the more and more they understand the intricacies of this, especially when we’re talking about a framework that includes more than just preemption.
Matt Perault (51:23)
Sure, I also think Democrats like it when it’s New York and California who are acting. But if you said, look, this theory of authority would mean Texas and Florida can try to set a national standard in issues that are, it would advance Republican priorities. I think Democrats become much more anxious about that.
So I think our goal here is to kind of establish a principle that again, it’s not our principle, it’s a constitutional principle to delineate between state and federal authority and do it in a way that has staying power that would work for Republican states, for Democrat states would work for Republican Congress, Democrat Congress, Republican administration, Democratic administration. And I think if we do that well, that will be good for Little Tech.
Collin McCune (52:03)
I also think that we’re kind of framing this in terms of fairness, but if we do not have a federal standard, my firm belief is we will lose this AI race to China. We will not have the capability of creating a national AI market in the way that we need to. And then ultimately we will lose to our competitor, our number one foreign adversary.
Matt Perault (52:28)
I have wondered about why components of this concept are as controversial as they are. So there are people who are skeptical of preemption, which I think is fine. But sometimes I’ve tried to start those conversations by saying, I get you don’t like the solution or particular solutions, but are you saying there’s not a problem? Like, are you saying that for startups looking at, we have resources to devote to tracking state policy for the two of us, it’s still overwhelming. It’s still really hard. We have more resources to devote to it than most of the startups in our portfolio. They’re in a position where it’s much harder to figure out how do you comply with this? How do you track this bill that might affect you?
We brought to one of our portfolios this thing that would have essentially put them out of business and they weren’t even tracking it. We were like, there’s this thing moving that you need to take account of and they weren’t even following it. And it would have really crippled their business in a state that’s a significant state that would have been impactful for their business.
And so it seems to me like there’s a problem. I have wondered even, you again, I have found there to be just good guidance and looking to existing law. One of the principles that we’ve looked to is a dormant Commerce Clause principle. So the idea that it’s not just that Congress needs to take steps to preempt state action in order to limit state authority. State authority is limited even if Congress doesn’t act because of the interstate commerce clause. And there’s a principle within the dormant commerce clause concept known as the excessive burden principle, which is the idea from this test called Pike, or Pike balancing. So out of state costs, the cost that a law would impose on the 49 states that didn’t pass the law and on companies trying to operate throughout the country, are those clearly excessive relative to the local in-state benefits?
And we recently wrote a post articulating that that sets up this balancing test with costs on one side and benefits on the other, local benefits costs. And typically, and there are legal scholars who have written about this, judges aren’t well positioned to do this balancing. And so what we were calling for was more of giving judges tools to do this cost benefit analysis. So just giving them data so that they can help to understand the costs and the benefits and giving them tools to do this weighing, cause they’re gonna be doing it. And it’s better for them not to just do it on vibes. There’s a lot of like judicial delineation that is just kind of like, we know when we see it, we have a general feeling about it. Shouldn’t this process be more informed by data? More data might result in some laws getting struck down, but it might actually provide an evidentiary record that helps some laws get upheld. So to me, it feels totally non-controversial.
It got push back from certain components of the law and AI community. And so some of that just doesn’t quite make sense to me. You know it better than I do because you’re not surprised to see the politics play out in all these different ways. But it feels to me like there is an approach here that’s constitutional, that gives judges real tools to make empirical assessments that will be good for states, be good for Congress, good for companies that we care about, smaller companies with smaller legal teams. So I hope that’s the direction we go in.
Collin McCune (55:44)
That is incredibly helpful. We’re going to leave everyone with a little bit of homework, which is to go and research Pike and figure out what Pike is.
I would also, maybe just to conclude here, I do feel very, very bullish in the direction that we’re heading. And I do think that the right people are now starting to get to the table and push in the right direction.
The one call to action here is, it’s an unfortunate reality, but especially in tech, most people don’t take action until something really, really negative has happened. And we’re probably within a year of a state bill passing, that would be pretty awful for the industry, collectively. And I also think that is probably pretty awful for the industry, but also cloaked in a way that’s probably good for the American public, but probably not very good for the American public. Now, I don’t have a crystal ball, I can’t predict exactly what that is, but something is going to happen.
So we need people to continue to come to the table. We need people to be active on this and share ideas and concepts, especially from the Little Tech community. And that’s what we’re principally working on on a regular basis. The time is now. The time is now to act and be there because as we say all the time, if you’re not at the table, you’re on the menu.
Matt Perault (57:07)
Colin, this is super fun.
Collin McCune (57:08)
Thanks, Matt. Talk soon.
This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.
