Peter Salid (Harvard Law School) has posted “Machine Learning and Class Certification” on SSRN. Here is the abstract:

Class actions are supposed to allow plaintiffs to recover for their high-merit, low-dollar claims. But current law leaves many such plaintiffs out in the cold. To be certified, classes seeking damages must show that, at trial, “common” questions (those for which a single answer will help resolve all class members’ claims) will predominate over “individual” ones (those that must be answered separately as to each member). Currently, many putative class actions in important doctrinal areas—mass torts, consumer fraud, employment discrimination, and more—are regarded as uncertifiable for lack of predominance. As a result, even plaintiffs with valid claims in these areas have little or no access to justice. This state of affairs is exacerbated by a line of Supreme Court cases beginning with Wal-Mart Stores, Inc. v. Dukes. There, the Court disapproved of certain statistical methods for answering individual questions and achieving the predominance of common ones.

This Article proposes a first-of-its kind solution: A.I. class actions. Advanced machine learning algorithms could be trained to mimic the decisions of a jury in a particular case. Then, those algorithms would expeditiously resolve the case’s individual questions. As a result, common questions would predominate at trial, facilitating certification for innumerable currently-uncertifiable classes. This Article lays out the A.I. class action proposal in detail. It argues that the proposal is feasible; the necessary elements are precedented in both complex litigation and computer science. The Article also argues that A.I. class actions would survive scrutiny under Wal-Mart, though other statistical methods have not. To demonstrate this, the Article develops a new, comprehensive theory of the higher-order values animating Wal-Mart and its progeny. It shows that these cases are best explained as approving statistical proof only if it can deliver accurate answers at the level of individual plaintiffs. Machine learning can deliver such accuracy in spades.