Artificial Intelligence (AI) technologies mediate our interactions with the world and our daily decision making, ranging from shopping to hiring to surveillance. The development of rich AI algorithms able to process and learn from unparalleled amounts of data holds promise for making impartial, well-informed decisions. However, such systems also absorb human biases, such as gender stereotyping of activities and occupations. Left unchecked, they will perpetuate these biases on an unparalleled scale. A steady stream of press confirms that this is a widespread problem in real-world applications. This research brings together an interdisciplinary team to develop the science of AI bias. The findings will impact AI researchers and developers (through novel methodologies), computational social scientists (through a deeper study of human biases at web scale), educators and policy makers (through the comprehensive analysis of bias), and downstream users of AI technology.
Compared to applications such as criminal risk scoring where fairness has traditionally been studied, modern AI systems are characterized by massive datasets, complex deep models and an unprecedented breadth of applications. This results in a wider spectrum of biases with complex propagation pathways, requiring an in-depth scientific investigation. The project develops the tools and techniques for recognizing, mitigating and governing bias in AI by combining expertise in deep learning, crowdsourcing and dataset curation, AI ethics, analyzing inference risk, web measurement, and science and technology studies. The component on recognizing bias includes an application of the Implicit Association Test combined with zero-shot learning to understand the societal bias of web corpora. Mitigating bias includes bridging active learning with research on adversarial examples for AI models. Governing bias includes a qualitative and quantitative study of downstream bias effects. The research is designed to be tightly connected, as for example when the recognition of curation bias in datasets leads to techniques in mitigating bias through enforcing group fairness in deep learning to governing bias in deployed system through developing bias observatories. The study will include advancements in machine learning (decomposing deep architectures, adapting reinforcement learning, exploring domain adaptation), human-computer interaction (developing novel active learning techniques, studying model interpretability), and digital ethnography (studying the effect of AI bias on culture, establishing an AI bias taxonomy). It will serve as a bridge between these fields, establishing tighter connections between them.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
