Whose Ground Truth? Modeling Ambiguity in Human-Centered AI

Summary

As AI systems increasingly interact with people and make decisions about them, the authors argue that human interpretation should be treated as a central design concern. Conventional machine-learning practice often assumes that each task has one definitive ground truth, aggregating differing annotations or treating them as noise. The position paper argues that many human-centered tasks legitimately support multiple interpretations of the same input. Compressing these interpretations into one target can erase meaningful differences in perception, judgment, and experience. It therefore calls for models of the space of plausible human judgments, with meaningful ambiguity distinguished from annotation noise. The authors propose that this perspective should shape AI representation, learning, evaluation, deployment, and governance, so systems better reflect the diversity of human interpretation.