Ex Ante fails

(Yifat Nahmias Senior Research Fellow, Center for Cyber Law and Policy (CCLP), University of Haifa; Researcher, The Heth Academic Center for Research of Competition and Regulation, College of Management. & Maayan Perel*+ Associate Professor, Netanya Academic College, Faculty of Law; Senior Research Fellow, Haifa Center for Law & Technology, University of Haifa Faculty of Law, “THE OVERSIGHT OF CONTENT MODERATION BY AI: IMPACT ASSESSMENTS AND THEIR LIMITATIONS,” pg lexis//pa-ef)

****Footnote 132*****

In order to illustrate this point, consider environmental law, where impact assessments are often used as a tool for addressing the effects of technology and the possible harm caused by it to the public. See Kathleen Waugh & Gary E. Marchant, Collaborative Voluntary Programs: Lessons from Environmental Law, in The Growing Gap Between Emerging Technologies and Legal-Ethical Oversight: The Pacing Problem 183, 184 (Gary E. Marchant et al. eds., 2011). The environmental-impact-assessment model shows that an impact assessment can guarantee that critical information will be made available to a larger audience, thus rendering the decision-making process less opaque. See Selbst, supra note 77, at 169; Robertson v. Methow Valley Citizens Council, 490 U.S. 332, 349 (1989); Kleppe v. Sierra Club, 427 U.S. 390, 409 (1976).

Examining a system before it is implemented could theoretically provide meaningful insights into how some algorithms operate and could certainly help to detect faulty nodes or other inherent issues with the algorithm. 136However, even if the algorithm is free of defects, if the data fed back into the system present a distorted or biased picture or perpetuate human biases against certain groups, then the outcome of the automated decision-making system will still be biased or discriminatory. 137An ex ante evaluation of the system alone will not identify a problem. 138To illustrate, consider the following hypothetical: a company has recently implemented a new credit scoring system that is based on a machine-learning algorithm, using previously underutilized, internal, and legitimate third-party data to perform smarter credit scoring. The company evaluated the system before implementation and submitted an impact assessment to the relevant agency, indicating its benefits, costs, and risks. During the training period, the company fed the system with data previously made by the company's human  [*166] employees. 139Clearly, if the training data presents a distorted or biased picture against a certain group of individuals, then the result generated by the scoring system will be erroneous. 140For instance, in our hypothetical example, if the vast majority of higher income individuals turn out to be male, while almost none of them were women, the overrepresentation of males in this sample may affect the learning of the algorithm. Although the algorithm itself is not faulty, discriminatory, or biased, the AI-based system might learn to discriminate against women. 141Thus, the AI system could amplify and perpetuate bias. 142

Scholars commonly differentiate between ex ante and ex post approaches to oversight and accountability. The ex ante approach generally aims at determining whether the automated decision-making process works as expected. The ex post approach is designed to support review and oversight once a decision has been made. See Kroll et al., supra note 66, at 637.