Dead rats, dopamine, performance metrics, and peacock tails: proxy failure is an inherent risk in goal-oriented systems.

When a measure becomes a target, it ceases to be a good measure. For example, when standardized test scores in education become targets, teachers may start 'teaching to the test', leading to breakdown of the relationship between the measure--test performance--and the underlying goal--quality education. Similar phenomena have been named and described across a broad range of contexts, such as economics, academia, machine-learning, and ecology. Yet it remains unclear whether these phenomena bear only superficial similarities, or if they derive from some fundamental unifying mechanism. Here, we propose such a unifying mechanism, which we label proxy failure. We first review illustrative examples and their labels, such as the 'Cobra effect', 'Goodhart's law', and 'Campbell's law'. Second, we identify central prerequisites and constraints of proxy failure, noting that it is often only a partial failure or divergence. We argue that whenever incentivization or selection is based on an imperfect proxy measure of the underlying goal, a pressure arises which tends to make the proxy a worse approximation of the goal. Third, we develop this perspective for three concrete contexts, namely neuroscience, economics and ecology, highlighting similarities and differences. Fourth, we outline consequences of proxy failure, suggesting it is key to understanding the structure and evolution of goal-oriented systems. Our account draws on a broad range of disciplines, but we can only scratch the surface within each. We thus hope the present account elicits a collaborative enterprise, entailing both critical discussion as well as extensions in contexts we have missed.

Conviction Narrative Theory gains from a richer formal model.

Conviction Narrative Theory (CNT) is a convincing descriptive theory, and Johnson et al.'s formal model is a welcome contribution to building more precise, testable hypotheses. However, some extensions to the proposed model would make it better defined and more powerful. The suggested extensions enable the model to go beyond CNT, predicting choice outcomes and explaining affective phenomena.