AI, on Algorithmic Justice: A New Proposal Toward the Identification and Reduction of Discriminatory Bias in Artificial Intelligence Systems [Abstract

Abstract

As with all digitally transformational technologies, AI systems’ technical complexity and rapid adoption in society fast and far outpace the current knowledge and experience of most lawmakers and regulators and operate beyond the envisaged scope of many existing legal doctrines and frameworks. Even if existing law is potentially adequate to address discriminatory algorithmic bias in some domains, the contextual “translation” of those laws to this brave new AI world is indeterminate and uncertain. Further, due to the difficulty of enforcement against illegal algorithmic discrimination, jurisprudential guidance is sparse and slow in coming. These gaps between the currently limited status of algorithmic law and the realities of the Wild AI West create a void in which unscrupulous or merely unfettered and ambitious algorithm purveyors may pursue enriching, but societally corrosive opportunities. Uninformed persons making decisions about and using AI systems may adopt and deploy such systems without appropriate insight, preparation, or restraint. Further, they may use algorithmic systems in ways contrary to purveyors’ guidance, such with recidivism risk systems used in sentencing. All these factors coalesce to create the potential for discriminatorily-biased algorithms to do exponentially amplified, persistent, and irreparable harm to individuals, communities, and society. The need is urgent for a workable system of algorithmic justice by which to illuminate and eradicate discriminatory computational biases, or at least to more quickly identify them and reduce their incidence and duration. Fostering greater access to justice, public trust in AI technology, and other important policy goals, an algorithmic justice system also would provide empirical mechanisms by which to establish baselines and measure and communicate the status of progress toward eradicating discriminatory algorithm bias in State of the Algorithmic Nation reports, for example. In addition, this algorithmic justice system would provide a framework for crafting meaningful policy, legislative, and regulatory systems for AI and a more accessible and definitive means of enforcing and litigating against illegal algorithmic discrimination. This work offers a new model toward an algorithmic justice system. As a beginning to address the likely immense and certainly multiply complex problem of discriminatory algorithmic bias, this new model commences with two foundational processes.