Frontier Technologies, Complex Externalities, and the Evolution of Private Order: Implications for AI Governance

When a new technology arrives, it brings with it a promise of benefit and a shadow of risk. Often, the true nature of that risk is hidden until the technology is actually used in the real world. This is especially true for artificial intelligence, a tool that can reshape how we live, work, and communicate. Because the dangers are so complex and depend on specific situations, it is difficult for anyone, including government officials, to write the perfect rules before the technology is even deployed. The central question facing society is not just how to stop bad outcomes, but how to discover the right ways to manage these risks as we learn more about them. This involves understanding "externalities," a term economists use to describe costs or benefits that affect people who are not part of the original deal. When a new technology creates problems for neighbors or strangers, those problems are externalities. The challenge is that these externalities are often so tangled and unpredictable that no single authority can map them out in advance.

Two researchers, Ali Palida and Nicholas Thielman, set out to explore a different angle on this problem. Instead of asking how the government should regulate artificial intelligence, they asked how government regulation might stop private groups from figuring out their own solutions. They argue that the process of using a new technology is a form of discovery. By trying things out, people learn what works, what fails, and who needs to be protected. This learning process allows private groups—like insurance companies, safety certifiers, or even neighbors negotiating with each other—to build flexible rules that fit the specific situation. The authors suggest that if the government steps in too early with a strict ban or a rigid rule, it might block the very experiments needed to find the best way to govern the technology.

To test this idea, the researchers looked at three historical examples where new technologies faced similar uncertainties. The first example involves electricity in the late 1800s. When electric lights and power lines were first introduced, they caused frequent fires, but no one knew exactly which devices were safe. The government did not immediately step in to set safety standards. Instead, a private organization called Underwriters Laboratories emerged. Started by insurance companies worried about fire losses, this group began testing electrical equipment and creating their own safety standards. Over time, as they tested more devices and saw what caused fires, they refined their rules. Their private system became so trusted that cities adopted their standards into law, and manufacturers sought their approval to sell products. The authors note that this system worked because it was flexible; it could change as new hazards were discovered through actual use, rather than being stuck with rules written before anyone knew what the risks were.

The second example comes from the world of cyber insurance. As computers and the internet grew, hackers became a major threat, but the nature of these threats changed constantly. One year, the main risk might be stolen credit card numbers; the next, it might be ransomware that locks up hospital records. Because the risks were so fluid, government regulators struggled to write rules that would stay relevant. Private insurance companies, however, adapted quickly. They began offering policies that required companies to meet specific security standards to get coverage. If a company failed to update its security, the insurance company would raise the price or cancel the policy. This created a feedback loop: as new types of attacks appeared, insurers learned from the claims data and updated their requirements. This private system allowed for a constant, real-time adjustment of safety rules, something a slow-moving government bureaucracy might struggle to match.

The third story takes place in the early days of radio broadcasting. Before the government created a central agency to manage radio frequencies, stations often interfered with each other, creating a chaotic mix of signals. The common belief was that this chaos could only be fixed by a central authority. However, the researchers found that before the government stepped in, radio stations were actually negotiating with each other. They would trade frequencies, agree on broadcast times, or buy out competitors to avoid interference. This private system of bargaining and property rights allowed the radio industry to grow and become valuable. It was only later that the government replaced this flexible, private system with a rigid, centralized plan. The authors argue that this shift didn't happen because the private system failed, but because politicians and powerful broadcasters wanted to control the valuable airwaves for themselves. The result was a system that was less adaptable and slower to innovate.

To understand the mechanics of this process, the researchers built a simple model of how decision-making works when the future is unknown. In their model, a new technology is used, and the results are observed. This observation teaches everyone something new. If a group of people affected by the technology can meet and renegotiate their rules whenever they learn something new, they can adjust their approach to fit the new reality. They can decide to stop using the technology if it is dangerous, or change how it is used if it is safe but needs tweaks. However, if an outside authority imposes a permanent ban or a fixed rule that cannot be changed by private agreement, the learning process stops. The ban prevents the experiment that would have revealed the truth. Even if the group later learns that the technology is safe and beneficial, they are still stuck with the old rule because they cannot bargain around the government's veto.

The researchers found that this creates two specific problems. First, a permanent ban stops the learning that is necessary to understand the risks. Second, even if the group learns that a different arrangement would be better for everyone, the fixed government rule prevents them from making that change. The authors emphasize that this does not mean government rules are always bad. Sometimes, a ban is necessary to prevent a disaster that is too severe to risk. But the study suggests that we must weigh the immediate safety of a ban against the long-term cost of stopping the discovery of better, more flexible solutions. The goal is not to leave dangerous technologies completely alone, but to avoid locking in rules before we have had a chance to learn what works.

The paper concludes that effective governance for artificial intelligence may not be something we can design perfectly from the start. Instead, it might be something that emerges as we use the technology and learn from our mistakes. Private groups, through contracts, insurance, and certification, are often better positioned to learn and adapt quickly because they have a direct stake in the outcome. They can change their rules as soon as new information arrives. Government intervention, while sometimes necessary, carries the risk of freezing the system in place before the right rules have been discovered. The authors suggest that the best path forward is to allow space for this private experimentation to happen, ensuring that we do not block the very process that will teach us how to manage the technology safely. The challenge for policymakers is to protect people from immediate harm without cutting off the path to better, more effective governance in the future.