Physics

Meteorology

Work (physics)

Pure mathematics

Mathematics

Knowledge management

Epistemology

Engineering

Philosophy

Ground truth

Artificial intelligence

Psychology

Training (meteorology)

Data science

Quality (philosophy)

Common ground

Need to know

Computer science

Field (mathematics)

Computer security

Social psychology

Mechanical engineering

MIS Quarterly

Organizational decision-makers need to evaluate AI tools in light of increasing claims that such tools out-perform human experts. Yet, measuring the quality of knowledge work is challenging, raising the question of how to evaluate AI performance in such contexts. We investigate this question through a field study of a major U.S. hospital, observing how managers evaluated five different machine-learning (ML) based AI tools. Each tool reported high performance according to standard AI accuracy measures, which were based on ground truth labels provided by qualified experts. Trying these tools out in practice, however, revealed that none of them met expectations. Searching for explanations, managers began confronting the high uncertainty of experts’ know-what knowledge captured in ground truth labels used to train and validate ML models. In practice, experts address this uncertainty by drawing on rich know-how practices, which were not incorporated into these ML-based tools. Discovering the disconnect between AI’s know-what and experts’ know-how enabled managers to better understand the risks and benefits of each tool. This study shows dangers of treating ground truth labels used in ML models objectively when the underlying knowledge is uncertain. We outline implications of our study for developing, training, and evaluating AI for knowledge work.

Is AI Ground Truth Really True? The Dangers of Training and Evaluating AI Tools Based on Experts’ Know-What