Notes
Teach me to be fair - An Experimental Study on Fairness, Self Efficacy and Human Involvement in ADM Systems
219Algorithmic Decision Making systems (ADMs) are becoming ubiquitous, posing new threats for society by using opaque processes making them hard to assess and control.
220Especially lay people struggle to understand the complex functionality of ADMs and are therefore often subject to misconceptions by believing ADMs to be objective while in fact they can be highly biased. This overconfidence in the system’s recommendation carries the risk of unreflectively adopting suggested decisions, thereby unknowingly fostering existing biases.
221In the present work, a job candidate selection scenario was designed to confront lay people without a background in machine learning algorithms with discriminatory algorithmic recommendations. This experimental mixed-design study (N=59) explores whether different levels of human involvement in algorithmic decision-making, i.e. believing to be contributing to the learning process of the ADM (being in-the-loop) or merely receiving recommendations (being out-of-the-loop, between subjects factor)), influences actual decision adoption, perceived self-efficacy as well as perceptions of the ADM regarding ethicality, objectivity, and performance. Additionally, we explored the application of fairness dimensions to ADM systems and reasons for corrective behavior in questionnaires and semi-structured interviews after the experiment.
222The study consisted of a pre-assessment of the constructs “Attitudes towards Algorithms”, “Algorithmic Self Efficacy” and “Dimensions of Fairness”, the interaction with the biased ADM system, a post assessment of the constructs, and a qualitative interview used to gain a deeper understanding of participants’ fairness evaluations. During the interaction phase, participants “out-of-the-loop” were only able to rate the fairness of the four decision cases, having been told that the evaluation could potentially lead to a revision by the software team. Participants “in-the-loop” additionally had the opportunity to rerank the job applicants to add to the training data set of the algorithm, thereby having higher control over the outcome as well as future performance of the ADM.
223Algorithmic decision making systems / algorithmic fairness / algorithmic self efficacy / artificial intelligence ethics / fairness dimensions / fairness perception / human-in-the-loop / machine learning perception / responsible AI