Rethinking Robot Liability

Today a growing chorus of voices in the tort-law literature sings a siren song. AI harms must be redressed through across-the-board strict liability, they croon—the same standard we apply to dynamite users, lion tamers, and faulty-chainsaw makers. Their rationale is understandable: at first blush, many AI harms appear impervious to other traditional theories of tort. In what sense might a self-driving car be negligent? Or how might one prove that an AI’s neural network—a “black box” filled with billions or trillions of inscrutable numbers—is defective under a products-liability theory? But tempting though strict liability may be, categorically applying it to AI harms would be a mistake, for at least three reasons. First, just like human activities, AI activities are not monolithic. An AI-enabled treadmill does not pose the same risk as an AI demolitions robot. Second, our default tort rules will work far better than the chorus suggests. Negligence and products liability are high flexibility, high context doctrines that have a long history of redressing harms caused by novel activities and technologies. Plus, they aren’t our only tools. Where an AI activity presents the potential for catastrophic harm, tailored, domain-specific legislation will be the right tool for the job. And where AI activities cause non-tortious harms, data-driven insurance markets (old and new) will be well suited to fill most compensatory voids. Third, even if something like categorical strict liability just for AI harms were desirable, no current theory of tort law could justify the bifurcated tort system that kind of rule would create.