Subjective risk decomposition provides new unified framework for AI uncertainty

A new paper from researchers Raghad Alamri, Michele Caprio, and Gavin Brown introduces a paradigm shift in uncertainty quantification (UQ) for machine learning. Rather than treating uncertainty measures as primitive constructs requiring axioms and justification, the authors propose that they arise naturally from a decomposition of subjective risk based on a strictly proper loss function. This

For example, using reverse cross-entropy as the loss recovers classical information-theoretic uncertainty terms such as entropy and mutual information. More importantly, the same mathematical framework reproduces numerous UQ measures previously scattered across the literature, providing them a common theoretical foundation. The authors also extend their view to learning theory—defining subjective risk analogues of excess risk, approximation error, and estimation error—and identify connections between these concepts and UQ. This is a first step toward a complete learning-theoretic treatment of uncertainty in machine learning models.