106 Iowa L. Rev. 775 (2021)
Artificial intelligence and machine learning represent powerful tools in many fields, ranging from criminal justice to human biology to climate change. Part of the power of these tools arises from their ability to make predictions and glean useful information about complex real-world systems without the need to understand the workings of those systems.
But these machine-learning tools are often as opaque as the underlying systems, whether because they are complex, nonintuitive, deliberately kept secret, or a synergistic combination of those three factors. A burgeoning literature addresses challenges arising from the opacity of machine-learning systems. This literature has largely focused on the benefits and difficulties of providing information to lay individuals, such as citizens impacted by algorithm-driven government decisions.
In this Essay, we explore the potential of machine learning to clear opacity —that is, to help drive scientific understanding of the frequently complex and nonintuitive real-world systems that machine-learning algorithms examine. Using examples drawn from cutting-edge scientific research, we argue machine-learning algorithms can advance fundamental scientific knowledge and that deliberate secrecy around machine-learning tools restricts that learning enterprise.
Our argument is more than a general plea for the innovation-related benefits of open science, or even a call for special attention to the unusually strong competitive protection secrecy can provide in the arena of machine learning. Rather, because the counterintuitive results machine learning can produce must be scrutinized particularly closely to distinguish exciting new hypotheses from spurious or otherwise misleading correlations, openness is particularly critical.
Turning to practical questions of law, institutions, and economics, we examine why developers are likely to keep machine-learning systems secret. We then draw on the innovation policy toolbox to suggest ways to reduce secrecy so that machine learning can help us not only to interact with complex, nonintuitive real-world systems but also to understand them.