Paper Type
ERF
Description
Crafting ground truth data labels is instrumental but challenging in AI development. In contrast to the prevailing dominant objective view on ground truth labels and human-centered data labeling approaches, we adopt a conjoined agency perspective to theorize how the complementarities between humans and AI play out in organizing the data labeling process for AI development. We conceptualize ground truth data labeling as a highly iterative process involving reflection in action between human agency and AI agency. We propose that the level of ground truth uncertainty determines the composition of conjoined agency and the degree of reflection in action necessary to get the appropriate labels, which can lead to two different organizing principles emphasizing either accuracy or divergence. Our theoretical framework and propositions are expected to contribute to unpacking the composition and interactive dynamics of humans and AIs in constructing ground truth data labels and how learning occurs within human-AI interactions.
Paper Number
1599
Recommended Citation
Tian, Hengqi; Rai, Arun; Kostis, Angelos; and Holmström, Jonny, "Ground Truth Mechanisms in AI Development: A Conjoined Agency Perspective" (2024). AMCIS 2024 Proceedings. 24.
https://aisel.aisnet.org/amcis2024/ai_aa/ai_aa/24
Ground Truth Mechanisms in AI Development: A Conjoined Agency Perspective
Crafting ground truth data labels is instrumental but challenging in AI development. In contrast to the prevailing dominant objective view on ground truth labels and human-centered data labeling approaches, we adopt a conjoined agency perspective to theorize how the complementarities between humans and AI play out in organizing the data labeling process for AI development. We conceptualize ground truth data labeling as a highly iterative process involving reflection in action between human agency and AI agency. We propose that the level of ground truth uncertainty determines the composition of conjoined agency and the degree of reflection in action necessary to get the appropriate labels, which can lead to two different organizing principles emphasizing either accuracy or divergence. Our theoretical framework and propositions are expected to contribute to unpacking the composition and interactive dynamics of humans and AIs in constructing ground truth data labels and how learning occurs within human-AI interactions.
Comments
SIGAIAA