What can machine learning discover about the nature of emotion?

The science of emotion increasingly relies on machine learning (ML) in the hope of making objective discoveries that resolve century-long debates. We examine research strategies in two main research traditions and show how confirmation bias can arise when theory-based hypotheses are evaluated using theory-laden sampling strategies, model specifications, and inferences. Each tradition manifests its theoretical differences in its ML methods, and, therefore, ML methods cannot settle those differences. The reciprocal relationships between starting assumptions, findings, and conclusions illustrate that ML investigations of emotion, like all scientific investigations, are never free from human experience and beliefs; however, they can be made less biased through a theory-aware research strategy to generate more justified knowledge about emotion.