The Dangers of Human-Like Bias in Machine-Learning Algorithms

Missouri S&T’s Peer to Peer Volume 2 | Issue 1 Article 1 May 2018 The Dangers of Human-Like Bias in MachineLearning Algorithms Daniel James Fuchs Missouri University of Science and Technology Follow this and additional works at: https://scholarsmine.mst.edu/peer2peer Part of the Computer Sciences Commons Recommended Citation Fuchs, Daniel J.. 2018. "The Dangers of Human-Like Bias in Machine-Learning Algorithms." Missouri S&T’s Peer to Peer 2, (1). https://scholarsmine.mst.edu/peer2peer/vol2/iss1/1 This Article - Journal is brought to you for free and open access by Scholars' Mine. It has been accepted for inclusion in Missouri S&T’s Peer to Peer by an authorized administrator of Scholars' Mine. This work is protected by U. S. Copyright Law. Unauthorized use including reproduction for redistribution requires the permission of the copyright holder. For more information, please contact . Fuchs: Dangers of Human-Like Bias in MLAGs Machine learning (ML), frequently used in constructing artificial intelligence, relies on observing trends in data and forming relationships through pattern recognition. Machine learning algorithms, or MLAGs, use these relationships to solve various complex problems. Applications can range from Google's "Cleverbot" to résumé evaluation, to predicting the risk of a convicted criminal reoffending (Temming 2017). Naturally, by learning through data observation rather than being explicitly programmed to perform a certain way, MLAGs will develop biases towards certain types of input. In technical problems, bias may only raise concerns over efficiency and optimizing the algorithm's performance (Mooney 1996); however, learned biases can cause greater harm when the data set involves actual humans. Learned biases formed on human-related data frequently resemble human-like biases towards race, sex, religion, and many other common forms of discrimination. This discrimination and the question of the fairness of artificial intelligence have received increasing public attention thanks to the numerous social media-based AIs launched in recent years. Microsoft's "Tay", an AI made to resemble a teenage girl, became anti-Semitic, racist, and sexist; Tay was shut down a mere "16 hours into its first day" (Wiltz 2017). Following in Tay's footsteps, Microsoft's "Zo" exhibited similar problematic biases despite additional precautions (Shah 2017). Other MLAGs, such as Beauty.AI's "robot jury," have demonstrated learned biases towards physical properties like skin tone and facial complexion (Pearson 2016). In these three popular cases, though the biases were quickly identified, the designers were unable to simply remove the learned biases. Despite the intention of their designers, many ML implementations have developed harmful human-like biases that cannot be easily removed. While much research is being done to improve performance speed, create more efficient implementations, and create more powerful MLAGs to solve more difficult problems, much of Published by Scholars' Mine, 2018 1 Missouri S&T’s Peer to Peer, Vol. 2, Iss. 1 [2018], Art. 1 this research does not concern bias control or correction. This is to be expected, as many ML implementations are applied to solve purely technical problems. While ML implementations might not have to enforce any form of fairness when dealing with strictly technical data, the growing usage of MLAGs that operate on human data reveals a need to better regulate bias to ensure fairness. The purpose of this study is to show the effects of these human-like biases in MLAGs across a variety of scenarios and to analyze the results of both current and emerging methods of bias correction. Human-like biases in MLAGs have many harmful effects, and there is a need for greater control over and the correction of these learned biases. Research Design To study the effects of human-like bias in MLAGs, I used the ACM Digital Library, IEEE Xplore, and Scopus. These three databases provide numerous articles on observations of learned biases in MLAGs and records of correctional efforts and methods to manipulate biases. The search keywords machine learning, correctional, artificial intelligence, and bias were used to browse these databases. Articles that concern observations of learned bias in MLAGs and articles that concern bias correction or avoidance methods are included in this study. Articles that focus on solving purely technical problems with MLAGs or statistically evaluating the performance of an MLAG have been excluded. A variety of ML implementations across different fields were studied to provide a more thorough understanding of the effects of humanlike learned biases in different circumstances. I also refer to recent incidents of bias-driven discrimination by ML implementations that garnered noteworthy public attention. These incidents, such as Microsoft's AIs "Tay" and "Zo" or Beauty.AI's artificial jury, while having relatively well-documented results thanks to the great public outcry, tend not to have their technical details revealed to the public. Some of the parties https://scholarsmine.mst.edu/peer2peer/vol2/iss1/1 2 Fuchs: Dangers of Human-Like Bias in MLAGs responsible for these incidents, such as Microsoft, offered statements explaining the behavior of their ML implementations (Wiltz 2017) but still did not disclose specific technical details. In discussing these events then, since research and academic journal articles are generally unavailable, I relied on popular articles on the subject. These sources are used here to discuss the behavior and actions of each MLAG. Machine Learning Training and Bias Origin MLAGs generally require two components before they can be applied to a particular problem. First, the underlying ML framework must be constructed. While the algorithm's designer may understand the framework itself, as Maria Temmings writes, "it’s often unclear — even to the algorithm’s creator — how or why [the algorithm] ends up using data the way it does to make decisions" (2017). It is difficult to directly observe learned biases to see why they formed or how they affect data; the complex network of relationships that compose the learned bias exist as an effectively abstract object. Therefore, rather than attempting to directly detect a learned bias, observers can identify bias by observing trends in the MLAG's decisions. The second component to creating a functional MLAG is proper "training." Training refers to exposing an MLAG to a special set of inputs with specific desired outputs to teach the algorithm how to solve a problem (Osaba and Welser 2017). This particular style of training, commonly known as "supervised training," sets up an MLAG to deal with future cases by using the training data as a reference. The MLAG then extrapolates from the training data to make future decisions. If the training data accurately represents the population that algorithm is to operate in, the behavior of the algorithm will (...truncated)