Loading...
Description
Artificial Intelligence (AI) and Machine Learning (ML) algorithms are being developed with ever higher accuracy. However, the use of ML also has its dark side. In the recent past, examples have repeatedly emerged of ML systems learning discriminatory and even racist or sexist patterns and acting accordingly. As ML systems become an integral part of both private and economic spheres of life, academia and practice must address the question of how non-discriminatory ML algorithms can be developed to benefit everyone. This is where our research in progress paper contributes. Using a real-world smart living case study, we investigated discrimination in terms of ethnicity and gender within state-of-the-art pre-trained ML models for face recognition and quantified it using an F1 metric. Building on these empirical findings as well as on the state of the scientific literature, we propose a roadmap for further research on the development of non-discriminatory ML services.
Recommended Citation
Kortum, Henrik; Fukas, Philipp; Rebstadt, Jonas; Eleks, Marian; Nobakht Galehpardsari, Marjan; and Thomas, Oliver, "Proposing a Roadmap for Designing Non-Discriminatory ML Services: Preliminary Results from a Design Science Research Project" (2022). Wirtschaftsinformatik 2022 Proceedings. 3.
https://aisel.aisnet.org/wi2022/human_rights/human_rights/3
Proposing a Roadmap for Designing Non-Discriminatory ML Services: Preliminary Results from a Design Science Research Project
Artificial Intelligence (AI) and Machine Learning (ML) algorithms are being developed with ever higher accuracy. However, the use of ML also has its dark side. In the recent past, examples have repeatedly emerged of ML systems learning discriminatory and even racist or sexist patterns and acting accordingly. As ML systems become an integral part of both private and economic spheres of life, academia and practice must address the question of how non-discriminatory ML algorithms can be developed to benefit everyone. This is where our research in progress paper contributes. Using a real-world smart living case study, we investigated discrimination in terms of ethnicity and gender within state-of-the-art pre-trained ML models for face recognition and quantified it using an F1 metric. Building on these empirical findings as well as on the state of the scientific literature, we propose a roadmap for further research on the development of non-discriminatory ML services.