UC Berkeley

Machine learning is transforming the economy, reshaping operations in communications, law enforcement, and medicine, among other sectors. But all is not well: Many machine-learning-based applications harvest vast amounts of personal information and yield results that are systematically biased. In response, policy makers have begun to offer a range of incomplete solutions. In so doing, they have overlooked the possibility —suggested intuitively by scholars across disciplines—that these systems are natural monopolies and have thus neglected the long legal tradition of natural monopoly regulation. Drawing on the computer science, economics, and legal literatures, I find that some machine-learning-based applications may be natural monopolies, particularly where the fixed costs of developing these applications and the computational costs of optimizing these systems are especially high, and where network effects are especially strong. This conclusion yields concrete policy implications: Where natural monopolies exist, public oversight and regulation are typically superior to market discipline through competition. Hence, where machine-learning-based applications are natural monopolies, this regulatory tradition offers one framework for confronting a range of issues—from privacy to accuracy and bias—that attend to such systems. Just as prior natural monopolies—the railways, electric grids, and telephone networks—faced rate and service regulation to protect against extractive, anticompetitive, and undemocratic behaviors, so too might machine-learning-based applications face similar public regulation to limit intrusive data collection and protect against algorithmic redlining, among other harms.