Why Liability Alone Won't Make AI Safe
Daniel Schwarcz and Josephine Wolff
As AI systems are being incorporated into more elements of our daily lives, from driving to writing code to generating music, movies, and pictures, the risks associated with these systems are also proliferating. AI risks vary widely—from chatbot-enabled suicide to copyright infringement to car accidents—that it can be difficult to come up with a one-size-fits-all approach to addressing and mitigating these risks. That’s why some scholars and policymakers have advocated for relying principally on tort liability to manage a broad range of AI risks. Doing so, the thinking goes, allows the companies creating these AI systems to figure out for themselves how to handle any risks associated with their technology, rather than relying on outmatched regulators to impose ex-ante rules and requirements for AI.
In an article forthcoming with the Fordham Law Review, we argue that it would be a mistake for governments to let the private sector and tort liability entirely dictate the future direction of AI safety and security controls. We base this argument largely on the work we’ve done looking at cyber risks, and more specifically, cyber insurance, because any attempt to make AI safer through liability will necessarily result in insurance products that provide companies with coverage for that liability. Many existing insurance products, ranging from cyber to CGL to D&O, already provide silent coverage for risks associated with AI. As the liability regimes associated with AI become clearer and more sprawling, it seems likely that insurers will face more pressure to sell coverage for an even wider variety of AI-related risks.
As insurers begin offering more AI-related coverage, they will also take on primary responsibility for dictating the safeguards that companies must implement to limit AI risks. Rather than figuring out for themselves how best to test AI systems for copyright infringement or faulty decision-making or security vulnerabilities, companies that buy insurance to protect themselves from the liability associated with these types of risks will look to their insurers to tell them what to do to secure affordable coverage. This will be especially true of small and medium-size companies that don’t have resources or expertise to invest in developing their own security and safety programs in-house.
That could be great if insurers know what advice to give their policyholders about how to make their AI systems safer and more secure, the same way they know to require fire insurance policyholders to have sprinklers and smoke detectors. But on the whole, insurers have struggled to figure out the equivalent of sprinklers and smoke detectors for computer systems—even systems less complex and varied than advanced AI systems.
Fifteen years ago, when policymakers were wrestling with some of the same questions about how best to mitigate cyber risks, many of them were similarly drawn to the idea that liability and insurance could do a better job of solving these problems than regulation. They hoped that insurers would collect data about cybersecurity incidents and then analyze that data to understand how best to mitigate these types of incidents. They also expected insurers to communicate these lessons directly or via pricing signals to their policyholders. But as the cyber insurance industry developed, it became increasingly clear that even as insurers helped cover the costs associated with a wide range of cyber incidents, they were not well positioned to actually identify, require, or price effective mitigation measures.
The failure of insurers to effectively mitigate cyber risks resulted from a variety of different factors—they struggled to collect granular data on their policyholders’ computer systems and security configurations, both because such data was often difficult and time consuming to gather and because, in some cases, collecting it might mean making it easier for victims of data breaches to sue their clients. Related to their inability to collect reliable data, insurers have had trouble identifying which cybersecurity controls work best and therefore don’t always know what they should require of their policyholders. And on top of that, the potential for a single cybersecurity incident to compromise computers across companies in different locations and industry sectors has made it difficult for them to diversify their risk pool, leading to fears about interconnected and catastrophic risk that could force them to pay out claims to all their policyholders for a single incident.
As a result, while cyber insurance is now a fairly robust market, insurers have not managed to raise the overall level of cybersecurity in the way that policymakers once hoped they would. Instead, insurers have focused largely on reducing the liability their policyholders face for cybersecurity incidents, for instance by involving lawyers in incident response who then prevent reports from being written about those incidents so there’s less discoverable evidence for plaintiffs to use against their policyholders. As a result of this failure by insurers to make computers more secure, or even collect good data about security incidents, regulators are now starting to require cybersecurity incident reporting and cybersecurity controls for specific high-risk sectors.
We believe that many types of AI risks will pose similar challenges for insurers. AI systems, like computer systems, are difficult to assess and require considerable technical expertise to identify vulnerabilities and security issues. There is also no consensus on the best way to protect AI systems from different types of risks, so it’s not clear what insurers would require of their policyholders or how they would try to improve the safety of such systems. And the potential for catastrophic risks seems, if anything, even more significant in the case of AI systems that are being embedded in every element of infrastructure and decision-making, creating additional headaches for insurers.
None of this is to say that there won’t be a robust market for AI insurance, or that liability for AI risks won’t play an important role in shaping some of the ways that companies test and safeguard their systems. But it does suggest that these tools, on their own, will not be sufficient for us to learn about the AI threat landscape, how best to protect ourselves against these risks, or how to require that companies implement those protections. The history of cyber risks and cyber insurance points to the need for more proactive steps to regulate the reporting of security and safety incidents so we can better understand harms and improve system safety, not just protect ourselves from liability.
Daniel Schwarcz is the Fredrikson & Byron Professor of Law at the University of Minnesota Law School. Josephine Wolff is a professor of cybersecurity policy at The Fletcher School at Tufts University. This post is based on their forthcoming paper, The Limits of Regulating AI Safety Through Liability and Insurance: Lessons from Cybersecurity.