Disposable People: Data cleaning and power in AI ethics

Presented by Pinar Barlas, LIS PhD candidate, Faculty of Information and Media Studies.

Attend in person: FNB 4110
Attend virtually: Zoom link

Abstract: From employment discrimination to wrongful arrests, many Artificial Intelligence (AI) applications have been found to produce harmful results. These harms affect marginalized groups – such as Black and other non-White people, people with disabilities, and non-men – more frequently and more intensely, in line with existing systematic injustices in society. In order to minimize or counteract these negative effects, researchers have been investigating the results of deployed systems as well as the datasets used to train them. Decades of findings show that biases, stereotypes, and other values become embedded in “raw” data, and such “raw” data must be processed before it can be used to train models and create AI systems. Yet, research in the AI ethics and fairness fields has long overlooked this processing – or “cleaning” – stage, during which ‘dirty’ data such as outliers and  missing values must be fixed. The cleaning process can therefore erase or otherwise remove the reality of marginalized groups from the dataset so that the AI system can be created. This presentation will outline an argument for using the Discard Studies framework and ethnographic methods to critically examine data cleaning practices, which will deliver new insights into ethical considerations in technology development. 

This talk is part of the 2023/24 Mediations Lecture Series.