Edwin Eboigbe (University of Illinois Urbana-Champaign) has posted Disaggregating Code, Data, and Design in the Law of Algorithmic Discrimination on SSRN. Here is the abstract:
Commentators, regulators, and even courts tend to talk about ‘AI bias’ as though it were one phenomenon with one legal answer. It is not. A discriminatory outcome produced by an algorithm can come from at least three different places: a rule someone deliberately wrote into the code, a pattern the system absorbed from historically skewed training data without anyone writing a discriminatory rule at all, or a design choice, such as which features to use or which outcome to optimize for, that was never tested for its discriminatory potential before deployment. These three causes are not legally interchangeable. They map onto disparate treatment law, disparate impact law, and a still-underdeveloped duty-to-investigate theory respectively, and a defendant’s best argument in one category can be irrelevant or even self-defeating in another. This author works through that taxonomy using real cases and enforcement actions primarily from the United States, with the European Union drawn in for comparative contrast. After which we turn from doctrine to consequence, tracing how each category of bias actually affects an ordinary person across different sectors. We close by asking what would narrow these gaps, considering the doctrinal and human stakes, and by analyzing the statistical definitions of ‘fairness’ that audits rely on whether or not they are all simultaneously satisfiable.
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum