David Manheim

David Manheim coauthored Categorizing Variants of Goodhart’s Law with Scott Garrabrant. The manuscript develops a framework for separating failures of proxy optimization in public policy, machine learning and AI Alignment.^[1]

Contribution

The joint paper formalizes a regulator selecting states by a metric rather than its true goal. It distinguishes noise, optimization outside a model’s useful range, causal intervention and other actors’ responses. These distinctions make Goodhart’s Law more specific than a blanket claim that targets always fail.^[1]

The manuscript does not assign individual responsibility for particular equations or sections. This entry attributes the framework jointly and makes no claim about current employment or the authors’ positions beyond this work.