Representation with a capital 'R'

Ambiguity surrounding the term 'representation' in biological and artificial neural systems hampers our ability to assess their alignment. In this paper, we draw a critical distinction between two notions of representation: the conventional 'representation' as basic encoding and 'Representation-with-a-capital-R', which entails functional use within a system. We argue that while current methods in neuroscience and artificial intelligence often focus on the former, advancing our understanding requires a shift toward the latter. We critique existing linking methods such as representational similarity analysis and encoding models, highlight key limitations in their ability to capture functional correspondence, and propose an updated paradigm that relies on explicit models of information readout. This framework, which treats neural networks as a new species of model organism, may help reveal the principles governing functional representations in both biological and artificial systems.