Why AI's biggest players might want regulation

You say you want a regu-LAY-shun…well, you know

In 1971, George Stigler famously argued that regulation tends to be “acquired by the industry and designed and operated primarily for its benefit.” The shock value of that claim has dissipated over the years. Today, business leaders are often the ones to propose regulation, and policy analysts openly ridicule the concept of “neutral” agencies acting in the “public interest.” Whether the result of cynicism or accumulated evidence, we appear some intellectual distance from the era when regulation was understood purely as a public service designed to restrain private sector excesses.

This literature doesn’t paint a picture of scheming businesses looking to “capture” pliant governments. It simply treats public intervention as a valuable service that can increase the risk-adjusted return on invested capital through entry barriers, price regulation, protective standards, exclusive franchises, tariffs, subsidies, procurement preferences, or rules that increase rivals’ compliance burdens. Historical treatments often link business’ demand for regulation to some form of market instability, often stemming from foreign entry, technological disruption, price cutting, or other forms of “ruinous competition.”

Even when industries face unwanted public intervention due to fraud, banking crises, safety accidents, or pollution externalities, the most common advice from Washington-based consultants is for management teams to be “for” something. That affords a “seat at the table” so that businesses can shape the new regime rather than have it imposed on them. When public safety is at issue, regulation can transfer reputational exposure and legal liability from firms to the government or relevant agency, offsetting some of its direct cost.

Businesses’ participation in the policy process not only allows them to shape regulation but also the discourse that produces it. What is the nature of the risk or potential harm that necessitates public intervention? How did it arise? What statistics are relevant for its measurement? These questions typically get answered through strategic, semiotically laden messaging campaigns that target journalists, elected officials, investors, and the broader public. When successful, they can distill issues of high-order complexity into an industry-friendly conceptual framework.

One sees that today in the AI policy discourse, which has imprinted two concepts into our collective consciousness. The first is the awesome power of “frontier” models. What it means to be frontier is generally based on model capabilities, which are said to be advancing so rapidly that they may soon introduce intolerable cybersecurity (or even existential) risks. The second is a high-stakes, zero-sum international competition between the US and China. While the precise nature of this competition can be hard to discern, the implication is that a loss would prove catastrophic for US interests.

“Frontier” could also be defined in economic terms. This is the premium AI industry segment, which charges customers higher prices in exchange for a qualitatively superior service. To maintain that superiority, frontier labs must train new models, which require massive upfront outlays that have been increasing at a 3x-to-5x annual rate. And as the pace of innovation accelerates, the time newly released models enjoy at the frontier shrinks, requiring the development of even newer models on a faster timeline to maintain price differentials and avoid the most dreaded fate for premium brands: commoditization.

This sounds precisely like the “ruinous competition” that spurs affected businesses to demand public intervention. But don’t assume that the risks are therefore fabricated. It could be that frontier lab executives anticipate that the next iterations of their models may pose national security risks of a sort that would limit their prospective customer base. After all, Lockheed Martin is not permitted to sell Joint Air-to-Surface Standoff Missiles (JASSMs) to whomever would like to buy one.

As we alluded in our recent paper (Rules of relevance), the market focuses on AI model capabilities even as returns depend on monetization. These can be very different things. Regulation may increase the odds that the returns on the installed capital base reach levels commensurate with investors’ expectations, but it may not arrive in time. It was 14 years after the Panic of 1873 that federal railroad regulation came into existence, and the Federal Power Act was enacted six years after the 1929 crash in the price of electric holding company and equipment stocks (Bubbles as a feature not a bug).