AI Governance Under Political Constraints: The Case for Ex Post Liability
This article argues that ex post strict liability is the most robust AI governance instrument under political constraints because it avoids the capture and consensus failures plaguing ex ante regulations, while offering a framework to classify complementary tools and evaluate efforts to weaken civil liability.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Artificial intelligence is moving faster than the rules meant to guide it. As these systems grow more powerful, they promise to solve difficult problems in medicine and science, but they also carry the potential for serious harm, from widespread fraud to the loss of human control over critical machines. Governments around the world are trying to decide how to manage this power before it causes damage. The central question is not just what rules to write, but how to write them in a way that actually works. The challenge is that no one agrees on exactly how dangerous these systems are, and the people who build them have far more influence over politicians than the general public does. This imbalance makes it difficult to pass strict laws that require companies to prove their technology is safe before they release it, because the politicians cannot agree on what "safe" means or how much risk is acceptable.
A new paper by Gabriel Weil, a researcher at the University of Houston, argues that the best way to handle this uncertainty is to stop trying to predict the future and start paying for the past. Instead of forcing companies to get government approval before they build or release a new artificial intelligence, the paper suggests a system where companies are held financially responsible for any harm their technology actually causes. This approach, known as "ex post liability," shifts the burden of safety onto the developers themselves. If the technology works well and causes no trouble, the companies pay nothing. If it causes a disaster, they must pay for the damage. Weil's analysis suggests that this method is more robust against political pressure and better suited to the unknown nature of the technology than trying to set strict rules in advance.
The paper begins by explaining why the usual methods of regulation are likely to fail in the world of artificial intelligence. Typically, governments try to manage risk by setting rules, requiring licenses, or taxing companies for the risks they create. For example, a government might say a company must prove its system is safe before it can be used, or it might charge a tax based on how much computing power the system uses. However, these methods require a clear agreement on how dangerous the technology is. To set a tax rate or a safety rule, politicians must first decide exactly how much harm to expect. Weil points out that experts cannot agree on this. Some believe the risk of catastrophe is nearly certain, while others think it is negligible. Because the technology is so new and complex, there is no solid data to settle this debate. Without a shared understanding of the risk, any rule or tax set by politicians will likely be either too weak to stop a disaster or so strict that it stops all progress.
Furthermore, the paper argues that the political process itself is skewed against effective regulation. The people who would suffer from a mistake made by an artificial intelligence system are the general public, who are scattered and unorganized. The companies that build these systems are few in number, wealthy, and highly motivated to protect their interests. In a political system, organized groups with money usually have more influence than the diffuse public. If a government tries to set strict rules, these companies can lobby to make the rules weaker or easier to bypass. They can also use their technical expertise to confuse regulators, who often rely on the companies themselves to explain how the technology works. This creates a situation where the rules end up protecting the companies rather than the public.
Weil proposes that holding companies liable for the actual harm they cause avoids these political traps. Under a system of strict liability, a company does not need to get permission to build a system, nor does it need to agree on a specific risk level with the government. Instead, the company simply knows that if its system hurts someone, it will have to pay. This works well even when no one knows exactly how dangerous the technology is. If the technology turns out to be safe, the company pays nothing. If it turns out to be dangerous, the company pays for the damage. This creates a powerful incentive for the company to figure out how to make the system safe on its own, because they are the ones who will have to foot the bill if it goes wrong.
The paper also addresses a common concern: what if a disaster is so huge that the company cannot afford to pay? In cases where the harm is too large for a single company to cover, the author suggests using punitive damages. This means that if a company takes a risk that could lead to a catastrophe, even if the catastrophe hasn't happened yet, they could be fined heavily if they come close to causing one. This fine would be calculated based on the potential danger they created, not just the damage they actually caused. This ensures that companies still have a reason to be careful about the biggest risks, even if they think they might not get caught. To handle the problem of companies not having enough money to pay for smaller accidents, the paper suggests requiring them to buy insurance. Insurance companies would then act as private regulators, checking the safety of the systems and charging higher prices for riskier ones.
The author contrasts this approach with other ideas that are currently being discussed, such as giving companies a "safe harbor" or immunity from lawsuits if they follow certain government rules. The paper argues that this is a bad deal. If a company is allowed to avoid paying for harm just because they followed a checklist, they have less reason to go beyond the minimum requirements to make their systems truly safe. The paper also looks at recent legislative proposals in the United States and Europe that try to trade liability for regulation. Weil finds that none of these proposals offer a good value. They tend to weaken the system of holding companies accountable without providing a strong enough alternative to keep people safe.
The paper concludes that while other tools like transparency requirements and safety research are useful, they should support a system of liability, not replace it. The most effective way to govern artificial intelligence is to let the market and the courts do the work of sorting out the risks. By making developers pay for the harm they cause, the system forces them to internalize the costs of their mistakes. This approach does not require politicians to agree on the future, nor does it give them the power to stop progress arbitrarily. It simply ensures that those who build the technology are the ones who bear the cost if it fails. In a world where the risks are unknown and the politics are difficult, this method offers the most reliable path to safety.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.