Design and Governance of Human-Facing Algorithms
7Sarah Cen (Massachusetts Institute of Technology)
https://simons.berkeley.edu/talks/design-and-governance-human-facing-algorithms
AI and Humanity
Data-driven algorithms have increasingly wide and deep reach into our lives, but the methods used to design and govern these algorithms are outdated. For example, most data-driven algorithms assume that humans report information about themselves truthfully, but this assumption rarely holds in human-facing applications, which has significant implications on the algorithms and their performance because they use the information humans provide as training data. Similarly, current methods for auditing data-driven algorithms are often too brittle to apply as state-of-the-art algorithms evolve. In this talk, we'll explore data-driven algorithms along two axes: (1) designing vs. governing data-driven algorithms, and (2) whether they are used for repeated vs. one-off decisions. We'll examine the design of algorithms for repeated decisions using recommender systems (e.g., Yelp, Facebook, Google) as a case study, focusing on the role of *trust* between users and recommenders. We'll examine the governance of algorithms for repeated decisions by looking at how to *audit* social media. Finally, we'll examine both the design and governance of one-off decisions by proposing a new legal right---the right to be an exception in data-driven decision-making---that tackles the problem of using averages (in almost every part of the algorithmic pipeline) to make high-risk decisions on *individuals*.