Claudio Zucca (European Union – European Investment Bank) has posted Personhood for AI Entities: Fragmentation, Correlation, and Capital on SSRN. Here is the abstract:
Proposals to grant legal personhood to AI systems raise a basic economic question: who bears the cost when an AI entity causes harm? This paper treats AI personhood as a grant of asset partitioning and limited liability and studies when it is compatible with efficient care. Three features of AI entities distinguish them from ordinary firms. First, they can be created at almost no cost, so their owners can split activity into many thinly capitalized entities. Per-entity capital requirements then lose their force, and even entities of a few units preserve most of the liability shield. Second, many entities run on the same underlying model, so a single defect can cause harm across all of them at once. This defeats the pooling of risk and shifts the decisive incentive to the model developer, who under current liability rules bears only a small fraction of the harm its defects cause. The arrangement works like a securitization in which the developer is the originator and the victims hold the senior claims without being able to price them. Efficient care would require the developer to retain nearly all of the loss that deployers cannot pay, and, if its safety work cannot be observed, to bear the full harm its defects cause. The minimum retention of 5% set for securitizations, which serves a different purpose, shows how far this lies from existing practice. Third, capital rules based on a solvency standard, such as the 99.5% rule used in insurance regulation, can leave the prevention of rare, shared failures almost entirely unrewarded. Limited liability for AI entities can be made compatible with efficient care if liability is aggregated across entities under common control, developers are liable for harm caused by defects in their models and hold capital to pay it, and capital rules are judged by the incentives they create as well as by solvency.
Highly Recommended! For an early treatment of the legal personhood question, see my Legal Personhood for Artificial Intelligences, 70 North Carolina Law Review 1231 (1992).
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum