Unfair treatment by artificial intelligence toward protected groups has become an important topic of discussion. Its potential for causing harm has spurred many to think that legislation aimed at regulating AI systems is essential. In the US, laws have already been proposed both by Congress as well as by several key states. However, a number of challenges stand in the way of effective legislation. Proposed laws mandating testing for fairness must articulate clear positions on how fairness is defined. But the task of selecting a suitable definition (or definitions) of fairness is not a simple one. Experts in AI continue to disagree as to what constitutes algorithmic fairness, which has led to an ever-expanding list of definitions that are highly technical in nature and require expertise that most legislators simply do not possess. Complicating things further, several of the proposed definitions are incommensurable with one another, making a cross-jurisdictional regulatory regime incorporating different standards of fairness susceptible to inconsistent determinations. On top of all this, legislators must also contend with existing laws prohibiting group-based discrimination that codify conceptions of fairness that may not be suitable for evaluating certain algorithms. In this article, I examine these challenges in detail, and suggest ways to deal with them such that the regulatory regime that emerges is one that is more effective in carrying out its intended purpose.
Greg Demirchyan (Fri,) studied this question.