When is a chair not a chair? Big data algorithms, disparate impact, and considerations of modular programming

• Published in: Computer and Internet Lawyer Vol. 34; no. 8; p. 6
• Main Author: Sherer, James A
• Format: Journal Article Trade Publication Article
• Language: English
• Published: Frederick Aspen Publishers, Inc 01.08.2017
• Subjects: Algorithms; Analysis; Artificial intelligence; Auditing; Big Data; Consumer behavior; Data collection; Data integrity; Data management; Datasets; Decision making; Disparate impact; General Data Protection Regulation; Marketing; Privacy; Quality management; Redlining; Social networks; Software; United States > US
• ISSN: 1531-4944

Objective data "quality" has therefore emerged-unsurprisingly-as a core component of modern algorithm development.1 Taking marketing algorithms as a proxy for human behavior-directed algorithms designed to reveal preference, such algorithms are built with the view that "historical consumer behavior" and "an individual's or a group's ... past choices" are the best predictors of future actions.2 True historical data often is considered the most important portion of the equation,3 and bigger is better, where "very large data sets can improve even the worst machine learning algorithms. Online sellers who "have large amounts of data about users' past purchases" subsequently "use this data as input" for their marketing algorithms as a matter of course.6 These data sets support algorithm development premised on the idea that truly rich data sets that identify an individual as well as certain things about that individual that the individual might prefer were private or "forgotten," will provide superior granularity when trying to sell a product or service. Some views supporting real data algorithm development argue that this process eliminates human biases from the development decision-making process, but algorithms ultimately comprise the data they work with.11 Algorithms "trained on historical data" will, for example, "know" that "poor uneducated people (often racial minorities) have a historical trend of being more likely to succumb to [things such as] predatory loan advertisements. If there is a discriminatory issue recognized on the back end, because such an issue is an "emergent property of the algorithm's use" rather than a design consideration, it can...