The Phrase That Alarms a16z’s AI Policy Chief: “It Only Covers a Handful of Companies”

Matt Perault is head of AI policy at Andreessen Horowitz, with side affiliations as a fellow at the Abundance Institute and at the technology-policy center he once directed at UNC Chapel Hill, now housed at NYU. He tells Nathan Labenz that he is not an engineer or a computer scientist, and that he has never been good at the forecasting exercise people always ask of anyone in tech. The people who actually build the tools sit on the other side of the house; his work is the policy agenda that lets them keep building.

Pressed for a picture of abundance, he does not lead with robots or a post-work society. Both of his parents and his sister are mental-health professionals, and he has heard since childhood how valuable support is and how hard it is to get: cost, scarcity of clinicians, stigma. Breaking down those access barriers, without lowering the quality of care, is one of the futures that actually animates him. The other is writing. Competence on the page has been the foundation of his academic and professional life, and the machines are now chasing that skill.

At Facebook he once told a boss he understood his job was to deliver a bad first draft. The boss laughed, then asked whether every now and then he could develop a good one. Today there are tools that can produce that bad first draft, and maybe, eventually, a very good one. Perault finds that both exciting and a little scary: it comes for a skill that has defined him, and it also lowers the barrier for people who never had that skill in the first place. Labenz, who calls himself classically ambivalent about AI, shares the thrill of new products and worries more about control than about his own job.

Labenz notes that a16z is known for techno-optimist leadership and for a running theme: leave room for “little tech,” not just big tech, to experiment and ship. Perault wants to correct a common reading of that stance. Plenty of people assume his job is a success only if no AI law is passed. That is not the case, he says, and the reason is the economics of the firm. An Andreessen Horowitz fund lives for ten years, longer than a typical public-company vesting cycle and longer than the horizon he saw in private equity.

They are not trying to juice tomorrow’s stock price. They are trying to back healthy, ambitious companies over a long stretch. If the public falls in love with AI products next year and the market then craters because of a cascade of real-world harms, a16z does not get the returns it is aiming for. Startups need room to get off the ground, but the ecosystem also has to produce tools that people like using and feel safe using. Big-tech interests sometimes overlap with that and sometimes do not; the firm’s focus, he insists, is little tech.

The first-order agenda, then, is not “no rules.” It is to push policymakers toward regulating harmful use and away from regulating AI development. That, in his view, is also how software has historically been treated: there is little regulatory intervention in the writing of code, but if the software you ship hurts people, existing law can usually find you, prosecute you, and deter the next builder. Leave the development process open; put the teeth on the harm.

Labenz presses on the high end of “development.” Fine-tuning a model for an app is one thing; pushing the frontier is another. He analogizes frontier training to biological research: labs ask what would happen if a virus could do X, often with the sincere hope of learning how to treat it, and yet lab leaks are not theoretical. At sufficient power, he argues, you may not even need to deploy the system, or intend to, for something to get out of control. That, for much of the AI safety community, is the number-one concern.

Perault says he does not know enough about how biological research is regulated to certify the analogy, but it feels like a good one on the surface. Society gives researchers broad latitude, including on work that would be catastrophic if released unlawfully, because the research process itself is not good or bad. He reaches for a case he knows: Cambridge Analytica. Data moved from Facebook to an academic researcher under the company’s terms; the researcher then misused it. The transfer looked, at the time, like the kind of academic access the research community is always asking for. The abuse was the problem, and that is where the penalties should sit.

Mapped onto AI, development is science. It is math. Criminalizing the math, he argues, makes it harder for scientists to do science and does little to address harmful uses. He thinks this is why so many people hear a16z as purely deregulatory: they absorb “don’t regulate development,” then the ears close before “and regulate harmful use.” He does not treat the second half as a diversion. As a law-school intern in the criminal section of the Justice Department’s Civil Rights Division, he watched attorneys spend their days building cases. Violations do not arrive with a clean fact pattern. Someone has to find them.

The same will be true of AI. Using a model to violate anti-discrimination law will not necessarily be obvious. You need people, budget, and technical literacy to see what happened, and in some cases you need to tweak existing statutes so they can actually cover the new harm. Most U.S. criminal law lives at the state level, so tightening rules around criminal uses of AI is largely work for state lawmakers, not a federal monopoly. Workforce development, AI literacy, and transparency are jobs for both. That is why he calls “regulate harmful use” a legislative agenda rather than a slogan.

A pattern of bills has emerged that sponsors describe as a compromise with people like Perault. Set a high threshold so that only a high-single-digit or low-double-digit number of companies are covered. Those firms must write a safety plan, publish or share it, follow it, and in some versions report incidents or sit for an audit. New York’s RAISE Act, sponsored by Assembly member Alex Bores, is the latest example Labenz has covered on the show. Proponents say they are not trying to squash frontier research, only to make sure the serious players are actually doing the work.

Perault first wants agreement on the category: those bills target development, not harmful use. Then he names the thing that sounds reasonable and is, for a16z, the thing they fear most. “This is just going to apply to a handful of companies.” Five, ten, fifteen. The other way to say it is that the market will be concentrated, and the regulatory model will not merely tolerate that concentration; it will help create it. Only a small number of firms will be allowed, in practice, to build at the frontier. A founder and five engineers should still be allowed the ambition of building there.

Before he joined a16z, he thought that story was an exaggeration. He has now heard it on Labenz’s show and in conversation with someone working on European Union AI regulation. For a lot of people the line sounds moderate. For the firm it is strongly opposed to a healthy ecosystem. Labenz is sympathetic to the fear of concentrated power, and he thinks many safety-minded people are lifelong techno-optimist libertarians except on this one issue. His hard objection is timing: some harms, once done, cannot be walked back. A fine after a pandemic does not resurrect the dead.

Perault flips the COVID example. How many constraints did the United States actually put on biological research afterward? Gain-of-function work, as far as he knows, continues. Society lived through a pandemic and did not create the civil or criminal liability Labenz is describing, probably because the same research is also how you learn to fight the next virus. If the upside of AI runs from better conclusions to dramatically cheaper mental-health care, he does not want handcuffs on the development process. Publishing a safety-protocol document, he adds, has no direct correlation with whether a product is actually safe. Limited efficacy plus a real burden on builders is, in his view, the wrong trade.

Labenz agrees that safety-plan-and-disclose bills are not a strong barrier against the worst outcomes. What he wants is not to fly blind. He sees labs chasing a self-improving loop in which models take over ML research. When Anthropic launched Claude 4, someone there said, in substance, that they wanted Claude to take over all the ML research so everyone could go to the beach. Daniel Kokotajlo, a lead author of the AI 2027 scenario, expects the gap between frontier capability and what the public can use to widen as labs race for supremacy and stop showing their cards. Labenz asks whether firms should at least have to say what their in-house models can do.

Perault points to an a16z blog arguing that many transparency mandates are probably unlawful under the First Amendment and not particularly useful to consumers. Courts have been fairly clear that it is hard to survive scrutiny if you require companies to provide information that is controversial, non-factual, or burdensome. Speculating about long-run harm or scoring safety risk falls in that bucket. The alternative the firm proposes is “AI model facts”: a basic fact sheet of information that actually helps a user decide how to query a model and when to distrust it.

His favorite example is the knowledge cutoff date. If you ask an LLM who won last night’s basketball game and the training data stopped months ago, you should not trust the score it invents. Knowing the cutoff lets you ask better questions and raises your suspicion when the model answers anyway. The team ran existing disclosure proposals against four tests: lawful, useful to consumers, consistent with the First Amendment, and not unduly heavy for little tech. Very few items cleared all four. Labenz is trying to keep outsiders from being locked out of the lab; Perault is trying to keep disclosure from becoming a new R&D tax that does not make products safer.

Labenz shares the worry about regulatory capture, but he thinks threshold designers landed where they did because scaling laws already look like a natural oligopoly. When he asks people to name the top-tier AI companies, the list is sometimes three names and never longer than ten. Costs are falling, and some proposals escalate the line over time, including via revenue, which he knows Perault likes better. Still, he asks, is it realistic to expect a crowded frontier when power laws rule everything a venture investor touches?

If the goal is to capture large companies that can bear compliance costs, Perault says revenue is the right vehicle. Once hundreds of millions, or hundreds of billions, are coming in the door, it is hard to argue that a firm cannot devote some share to whatever disclosures regulators want. A company taking in $500 billion a year can pay for paperwork. Compute and training-cost thresholds are different: they measure the science of building the tool, not the capacity to handle regulatory complexity, and startups do build models with significant compute.

He took a $100 million training-cost threshold to someone on a16z’s technical team, because he was seeing that number in more bills, including RAISE. Is that a good way to carve out little tech? The colleague’s answer, which Perault trusts even if he cannot certify it himself, was that every model costs $100 million to build. The line would not separate anyone. RAISE also fails, in his reading, to say the cap cannot be cumulative. A startup shipping several $25 million models a year could, over a few years, trip the “large developer” definition. Technology will outrun the legislature’s ability to update the number.

He does not deny that compute and data are natural barriers. The firm has ideas for broadening access to both. The policy question is whether lawmakers should add a second, artificial hurdle on top. Do that, and you get something that looks like telecommunications: a handful of companies under heavy oversight, historically including an explicit government-granted monopoly. Some people think that is the right destination. a16z thinks it means less innovation and a frontier that startups cannot contest.

Labenz notes that a16z’s ten-year clock is now longer than many people’s AGI timelines, and he has long pitched sunset clauses: put rules in place for a critical three to five years, then let them expire if the thresholds were wrong. California’s SB 813 is a cousin of that idea. Companies could opt into a state-approved, privately administered compliance regime in exchange for a tort-liability shield, and potentially opt out later. Nobody, he says, wants the GDPR of AI. Perault’s reply is that plenty of lawmakers seem to want exactly that, judging by the bills they are moving.

He has seen reporting, which he cannot independently confirm, that the EU is considering pausing implementation of its AI Act, its own AI-shaped GDPR. The only U.S. state to enact comprehensive AI legislation is Colorado, and Colorado’s governor came out in support of a federal AI moratorium that would have made his own law unenforceable. People who have already put their hands on the statute are getting nervous. That is a louder signal than any op-ed. Experimentation, including sunsets, is something Perault generally likes. A stringent regime for even a short period is still, he says, like running a race as fast as you can while wearing a 20-pound backpack for the first mile. You can run. Your place at the finish may already have changed.

On SB 813 itself, a16z is still studying the bill. Perault credits it for facing tort liability, which most proposals never touch and which he thinks is unavoidable if ordinary people are going to use AI for small, useful things. The knife is the word “voluntary.” Immunity from tort law is a massive benefit. If the price of that benefit were $20 billion, large companies would pay and small ones would eat the legal risk. That is not a choice; it is a price that sorts the market. His test is simple: can little tech actually bear the compliance cost? If yes, the bill can move toward support. If not, it is an incumbent advantage dressed up as an opt-in.

The disagreement between the two men is mostly empirical, not moral. Both dislike concentrated power, both fear black boxes, both want the technology to deliver real benefits. What divides them is whether scaling laws will freeze the number of frontier players, how powerful the systems will get and how soon, and how much a published safety document actually reduces accidents. Those are claims evidence can move. Perault’s closing watch item is the Trump administration’s national AI action plan, expected in June or July. He does not predict the future. He tries to make sure the people who can build still have somewhere to build.