This paper discusses correlation learning from data, which is the foundation of machine learning (ML) and artificial intelligence (AI) research. Modern methods can automatically discover complex patterns, but they are prone to failing to capture unintended correlations. This vulnerability has led to a growing body of research on spurious correlations, often seen as threats to model performance, fairness, and robustness. In this paper, we move beyond the traditional statistical definition of spurious correlations, which refer to non-causal relationships that arise due to chance or confounding variables, and investigate how their meaning is negotiated in ML research. Rather than relying solely on formal definitions, researchers evaluate spurious correlations through what we call a pragmatic frame. A pragmatic frame is a judgment based on what the correlation actually does: how it affects model behavior, supports or hinders task performance, or aligns with broader normative goals. Drawing on an extensive survey of the ML literature, this paper identifies four frames: relevance (“models should use correlations that are relevant to the task”), generalizability (“models should use correlations that generalize to unseen data”), human-likeness (“models should use correlations that humans would use to perform the same task”), and harmfulness (“models should use correlations that are not socially or ethically harmful”). These representations demonstrate that the desirability of a correlation is not a fixed statistical property, but a situational judgment informed by technical, epistemological, and ethical considerations. By examining how fundamental ML issues are problematized in the research literature, this paper contributes to the broader discussion of the contingent practices by which technical concepts such as spurious correlations are defined and operationalized.
