How should states regulate AI? Andreessen Horowitz weighs in.

Summary

POLITICO Tech Podcast Summary: "How Should States Regulate AI? Andreessen Horowitz Weighs In"

Release Date: February 4, 2025

---

Introduction

In the February 4th episode of the POLITICO Tech podcast, host Stephen Overle delves into the pressing issue of artificial intelligence (AI) regulation within the United States. The episode centers on insights from Matt Peralt, the Head of AI Policy at the venture capital firm Andreessen Horowitz (a16z). The discussion navigates the complexities of state versus federal regulation of AI, the implications of China's advancements in the AI sector, and the potential impact of a fragmented regulatory landscape on American startups.

---

Context: The AI Competitive Landscape

The episode opens with a brief overview of the AI landscape, referencing the emergence of a Chinese AI model named Deepseek. This development has heightened concerns among Silicon Valley, Wall Street, and Washington lawmakers about the United States' position in the global AI race. The conversation underscores that while federal deliberations on AI restrictions have been ongoing for over a year, the substantive legislative actions are occurring at the state level, with nearly every state proposing some form of AI legislation in the past year. Colorado notably passed the first comprehensive AI law in the country.

Key Quote:

"Nearly every state in the nation considered some type of AI bill last year, including Colorado, which passed the first comprehensive law in the country."
— Stephen Overle [00:36]

---

Andreessen Horowitz's Stance on AI Regulation

Matt Peralt articulates Andreessen Horowitz's concerns regarding the potential overregulation of AI, particularly focusing on AI model development. He warns that excessive regulatory burdens could stifle innovation, especially for smaller tech companies that lack the resources of larger corporations to navigate a patchwork of state laws.

Key Quotes:

"When you're regulating AI model development, you're essentially just focused on creating regulatory burdens that apply to the science of AI development. In our words, you're regulating math."
— Matt Peralt [03:34]

"Larger companies might be able to have large compliance teams that could figure out how to navigate product development in the face of differing restrictions across state lines, but little tech companies are going to really struggle to do that."
— Matt Peralt [01:41]

Peralt emphasizes that regulating the development process itself could hinder the pace of innovation, making it challenging for American firms to maintain leadership in AI.

---

Federal vs. State Regulation

A significant portion of the discussion revolves around the merits and drawbacks of federal versus state-level regulation. Peralt argues for a federal approach to AI regulation to prevent a fragmented legal environment that could disadvantage startups and smaller tech firms. He suggests that the federal government should oversee AI model development, while allowing states to focus on regulating the uses of AI, particularly harmful applications.

Key Quote:

"If you are trying to make it more difficult for models to develop in new ways, if you're making it more difficult for companies to develop the most innovative, interesting frontier models, then you're inevitably going to slow the pace of what American companies can do."
— Matt Peralt [04:05]

Peralt also highlights the role of states as "laboratories of democracy," citing Utah's introduction of a regulatory sandbox for AI companies as a positive example of state-level experimentation that can inform federal policy.

Key Quote:

"States are often referred to as the laboratories of democracy. And states are doing some experimental things in AI that I think are compelling and would probably be good for Little Tech."
— Matt Peralt [08:26]

---

Regulatory Sandboxes: A Path Forward

The concept of regulatory sandboxes is explored as a means for states to experiment with AI governance without imposing blanket restrictions. These sandboxes allow companies to operate under relaxed regulations in a controlled environment, providing policymakers with valuable insights into the practical implications of AI technologies.

Key Quote:

"There might be some areas of law that are not enforced against them once they're participating in this bounded environment."
— Matt Peralt [09:41]

This approach aims to balance innovation with consumer protection, enabling tech firms to develop cutting-edge AI applications while ensuring oversight and accountability.

---

China's AI Advancements and National Competitiveness

The episode draws parallels between the current AI race and the historical Sputnik moment of the 1950s, reflecting concerns that China's advancements in AI could challenge American technological supremacy. Peralt underscores the urgency for the U.S. to foster a conducive environment for AI innovation to prevent China from dominating the field.

Key Quote:

"Do we want the future of AI to be determined by China with Chinese products? Or do we want it to be, do we want the leading companies in this field to be American companies?"
— Matt Peralt [02:56]

He stresses that maintaining competitiveness requires prioritizing the development of AI technologies without imposing restrictive regulations that could slow progress.

---

Avoiding Overregulation: Focus on Harmful Uses

Peralt advocates for a regulatory framework that targets harmful uses of AI rather than the development process itself. This approach would protect consumers from misuse of AI without hampering the creation of innovative technologies.

Key Quote:

"Our focus is actually that regulation should focus on harmful uses of the technology as opposed to model development."
— Matt Peralt [12:53]

He argues that policymakers should enhance their ability to enforce existing laws against harmful applications of AI by building technical capacity and understanding.

---

Conclusion: Balancing Innovation and Regulation

The episode concludes with a reaffirmation of Andreessen Horowitz's position: to foster a thriving AI ecosystem in the U.S., regulation should be carefully calibrated to avoid unnecessary burdens on model development while effectively addressing the risks associated with AI misuse. Peralt envisions a unified federal approach complemented by state-level initiatives that encourage experimentation without creating insurmountable obstacles for smaller tech firms.

Key Quote:

"The key thing is to try to enable companies to deliver the most innovative models possible while simultaneously focusing more on protecting consumers directly as opposed to trying to do that through the bank shot of AI model development regulation."
— Matt Peralt [14:39]

---

This episode of POLITICO Tech provides a comprehensive examination of the current AI regulatory landscape in the United States, offering valuable perspectives from a leading venture capital firm. Matt Peralt's insights highlight the delicate balance between encouraging technological innovation and ensuring responsible AI deployment, emphasizing the need for cohesive federal policies to support American competitiveness in the global AI arena.

Transcript

[Advertiser]

This episode is brought to you by Shopify. Forget the frustration of picking commerce platforms when you switch your business to Shopify, the global commerce platform that supercharges your selling. Wherever you sell with Shopify, you'll harness the same intuitive features, trusted apps and powerful analytics used by the world's leading brands. Sign up today for your $1 per month trial [e-mail] tech. All lowercase. That's shopify.com tech.

[Stephen Overle]

Hey, welcome back to POLITICO Tech. Today's Tuesday, February 4th. I'm Stephen Overle. We talked last week about a Chinese company called Deep Sea and how its new AI model is causing Silicon Valley and Wall street to freak out about America's standing in the global AI race. Well, it stoked concerns in Washington, too, where lawmakers have been debating for over a year whether to impose restrictions on the most powerful AI models. But when it comes to AI policy, the real action is happening in the states where legislatures have been moving faster than Congress to regulate. Nearly every state in the nation considered some type of AI bill last year, including Colorado, which passed the first comprehensive law in the country. The venture capital firm Andreessen Horwitz warns that Deepseek should not lead lawmakers to overregulate and that a medley of state laws would be bad for startups.

[Matt Peralt]

Larger companies might be able to have large compliance teams that could figure out how to navigate product development in the face of differing restrictions across state lines, but little tech companies are going to really struggle to do that.

[Stephen Overle]

Matt Peralt is making that case to lawmakers as the head of AI policy at Andreessen Horwitz. And it's worth noting that Andreessen Horwitz isn't just any investment firm. The principals made headlines for donating to President Donald Trump's campaign, and several allies and former employees now have plum roles inside the new administration. On the show today, Matt outlines where Andreessen Horwitz wants federal and state regulators to get involved when it comes to AI and where it doesn't. Here's our conversation. Matt, welcome back to POLITICO Tech.

[Matt Peralt]

It's great to be on again.

[Stephen Overle]

So we're in this deep seq moment. I want to start there. Your boss, investor Marc Andreessen, referred to deepseek as AIs, quote, Sputnik moment, referring to kind of the crisis moment in the US Space with Russia. Should Washington think of the AI race now as kind of a cold war of sorts with China?

[Matt Peralt]

So I think it is really important that Washington focus on how American companies can be as competitive as possible. I think the question is, do we want the future of AI policy? Do we want the future of AI to be determined by China with Chinese products? Or do we want it to be, do we want the leading companies in this field to be American companies?

[Stephen Overle]

And so is this kind of a crisis moment as you see it? Because I've heard that argument right for a long time that we need to be investing in AI to stay ahead of China. Many are kind of interpreting this deep seq AI model to be a sign that we need to move even faster or that China is catching up.

[Matt Peralt]

Well, in the last couple years there's been a really strong focus on regulating AI model development, and that's been a concern for Andreessen Horowitz, because when you're regulating AI model development, you're essentially just focused on creating regulatory burdens that apply to the science of AI development. In our words, you're regulating math. And the result of that is that you literally slow the pace of innovation. Because if you are trying to make it, if you're making it more difficult for models to develop in new ways, if you're making it more difficult for companies to develop the most innovative, interesting frontier models, then you're inevitably going to slow the pace of what American companies can do. In our view at Andreessen Horowitz, the burdens of that are not born proportionately throughout the AI ecosystem. They're really born disproportionately by little tech companies, which are the companies that we are focused on, the startups that are trying to get off the ground and compete with larger platforms. It's incredibly difficult to do that if there are significant regulatory compliance burdens that are placed upon you. I think the release of Deepseek makes clear we should instead concentrate on how AI is used. If there are different things that need to be strengthened about consumer protection law, for instance, so that, so that consumers are protected when AI is used, then we should do those things. But if the way that we go about trying to protect consumers is making it harder for models to be developed and innovative models to be developed, then we're going to make it less likely that American firms will lead the way in AI model development.

[Stephen Overle]

I guess with that in mind, what do you see as the worst overcorrection or the worst response policymakers could make to this deep sea moment and these fears about China catching up?

[Matt Peralt]

I would say if we looked at this as an opportunity to make it harder for American companies to release open source technologies, or if we thought that this showcases significant risk associated with AI and therefore we needed to actually increase the regulation of AI model development. That's the kind of thing I think that would take us backwards. Our view is that we need to do really sort of three primary things. Continue to make it possible for firms to develop open source technologies, focus regulation on harmful uses as opposed to model development, and then make sure that we are orienting the creation of AI policy around developing a national market in AI as opposed to a patchwork state by state framework where model development and the production of AI models is regulated potentially differently in California and Texas and New York and Vermont and Nebraska. In our view, it's really important that the federal government and the state governments play their traditional roles that they play in lots of areas of policy where the federal government is responsible for regulating interstate commerce and transactions that occur across borders and states, police activity within their jurisdictions.

[Stephen Overle]

Got it. I do want to talk a bit about the state federal breakdown. I'm curious if you worry at all about the federal government kind of doubling down on big tech on companies like OpenAI or Meta or Google in response to Deepseek, rather than kind of the competition of the US market and the little tech that I know Andreessen Horwitz aims to represent.

[Matt Peralt]

What we're really focused on is ensuring that little tech has the ability to compete and that there's a level playing field between little tech and larger platforms. If the federal government did things that made it harder for little tech to compete, such as creating lots of hoops that companies need to jump through in order to develop cutting edge AI models, that in our view would be really problematic. And of course we don't think that there should be an exception to antitrust law simply because a company is building an AI related product. If there are companies that are engaged in anti competitive conduct, then obviously traditional antitrust law should apply and should be enforced against those companies. But we're not looking for special treatment for little tech. What we're really focused on is ensuring that little tech has the ability to compete.

[Dave]

Oh, such a clutch off season pickup, Dave. I was worried we'd bring back the same team. I meant those Blackout motorized shades. Blinds.com made it crazy affordable to replace our old blinds.

[Matt Peralt]

Hard to install?

[Dave]

No, it's easy. I installed these and then got some for my mom. She talked to a design consultant for free and scheduled a professional measure and install hall of Fame son. They're the number one online retailer of custom window coverings in the world. Blinds.com is the goat shop.

[Advertiser]

Blinds.com right now and get up to 45% off select styles, rules and restrictions may apply.

[Stephen Overle]

You and I have spoken a lot in the past about how state legislatures have really introduced a record number of tech bills in recent years, including a number of bills aimed at regulating artificial intelligence. What does the state landscape look like for 2025?

[Matt Peralt]

So there are a range of different types of bills that are being introduced at the state level. What we're most concerned about are bills that are trying to regulate AI model development and to do it at the state level. And that I think is going to be problematic because the patchwork that would result from state by state regulation will be really challenging for Little Tech. So larger companies might be able to have large compliance teams that could figure out how to navigate product development in the face of differing restrictions across state lines. But Little Tech companies are going to really struggle to do that. And so we think that the regulation of models should really occur at the federal government. There are a range of other things that states are doing that I think are interesting and exciting. States are often referred to as the laboratories of democracy. And states are doing some experimental things in AI that I think are compelling and would probably be good for Little Tech. Utah, for instance, has introduced a regulatory sandbox for AI companies. And that's the kind of experimentation and innovation in policy structure that is, I think, interesting and compelling and is consistent with the traditional experimental roles that states have played with respect to policy development.

[Stephen Overle]

What is a regulatory sandbox? That's a new term for me.

[Matt Peralt]

So a regulatory sandbox. There are lots of different definitions, there are different permutations, but the basic idea is that companies can apply to participate, and once they're participating, there typically is a higher level of transparency that's required. So companies that are participating have to provide information that will enable policymakers to learn about how a policy framework is functioning in practice. And while they're participating, there's some kind of regulatory forbearance. So there might be some areas of law that are not enforced against them once they're participating in this bounded environment. And so the idea is that it functions kind of like a testing environment, almost like clinical trials in medicine, where there is regulatory oversight, but so long as there's that oversight, companies are given a little bit more room to experiment.

[Stephen Overle]

Got it.

[Matt Peralt]

They're usually for a bounded period of time. So to the extent they aren't working well, like if companies were engaged in practices that were problematic, or if the regulatory regime were too burdensome or didn't sufficiently address potential consumer harm, because it's within that bounded period of time, it will end at a certain point. And at that point you could figure out an alternative mechanism for governance.

[Stephen Overle]

Got it. So it sounds like you do see some role for states to play in this kind of AI oversight. You're not saying states set this one out completely. This is federal territory.

[Matt Peralt]

I think a lot of times when people talk about preemption, the way that they talk about it is the federal government should act and states should not. And in the post that we published today, that's not the framing that we adopt. There's robust law that exists currently at the state level that should obviously be enforced when it comes to AI. There are state unfair and deceptive trade practices laws, and there are state civil rights laws, and there are state antitrust laws. There's no exception in existing law for AI. The question is whether states should also be engaged in what's traditionally been the domain of the federal government, which is regulating interstate commerce. And so when you have companies in places like China that are able to develop cutting edge models with the support of their government, it doesn't make sense for American companies to have to create models which will exist across state lines and then outside the country as well across international boundaries with respect to different legal regimes in one state versus another. And that's consistent with the, with the framing in the Constitution around the role that the federal government and the state governments would play. The federal government is responsible for regulating interstate commerce.

[Stephen Overle]

We've seen in many other areas in tech policy where states have moved to regulate because Washington really hasn't. Data privacy, I think is a prime example where we now have a number of states that have state data privacy laws. Kids online safety is another area where you see a lot of states moving to act. It cynically makes me wonder like is calling for Congress to regulate or saying this is the domain of the federal government, essentially the same as calling for no regulation at all.

[Matt Peralt]

We're not calling for no regulation at all. Our focus is actually that regulation should focus on harmful uses of the technology as opposed to model development. And so in a piece we actually released last week, we laid out how we think about that. It's not a piece that calls for no regulation. And in fact, we say that in order to enforce existing, even if you were just to focus on enforcement of existing law, even just to do that, there are important things that need to be done to ensure that lawmakers can do that effectively. They don't just get handed cases on a conveyor belt. They actually need to build cases to understand how to enforce existing law when new technologies are used to potentially violate it. And so that means they need the technical capacity and technical training to understand how AI might be used in harmful ways. And so I think if the orientation is around how do we regulate harmful uses of AI, that is not at all saying do nothing. There's an important policy agenda behind that and an important set of activities that policymakers need to be engaged in to make sure that that happens. But that's very different from putting a regulatory focus on AI development.

[Stephen Overle]

When I've heard from a number of AI executives in recent days talking about Deep Seq, including Sam Altman from OpenAI, for instance, they say this competition is a good thing. And I understand, you know, in a free market, you know, competition should, in theory, drive businesses to work harder or more efficiently, I guess I wonder how that squares with this notion that Chinese AI is bad for national security. And I'm curious what your take is on whether competition like this is ultimately a good thing or not.

[Matt Peralt]

So I don't think it would be a good thing if 20 years in the future, what we're saying is that China won the battle to create the best AI companies in the world. I think the reason that some people are viewing this as a good thing, and I think framing it as a Sputnik moment is consistent with that, is that there's an opportunity to shift policy direction when we're still at a nascent phase of the development of AI tools. And this is a moment to change the trajectory of the regulatory environment. And in our view, the key thing thing is to try to enable companies to deliver the most innovative models possible while simultaneously focusing more on protecting consumers directly as opposed to trying to do that through the bank shot of AI model development regulation.

[Stephen Overle]

Got it, Got it. Interesting. Well, listen, Matt, appreciate you being here on Politico Tech.

[Matt Peralt]

Thanks a lot, Steven.

[Stephen Overle]

That's all for today's Politico Tech. If you enjoy Politico Tech, please subscribe. And for more tech news, subscribe to our newsletters, Digital Future Daily and Morning Tech. Our managing producer is Annie Reiss. Today's show was produced by Kara Taber. I'm Steven Overle. See you back here tomorrow.