This paper proposes a new theory of reduplication that is based on redefining phonological copying as distribution of underlying activity within the framework of Gradient Symbolic Representations (Smolensky & Goldrick 2016). This assumption correctly predicts that copied elements are more likely to undergo phonological reduction and that reduction in a non-copying context implies reduction in a copying context but not vice versa. These predictions are summarized in the new empirical generalization of the 'Copying-Weakening-Implication'. It is illustrated with case studies of two language types that show different thresholds for reduction: languages can show reduction only for copied material or only for material that is copied twice. Another important result of the proposal is the fact that copying symmetrically weakens all elements that are involved in the copying and does not single out the 'reduplicant' as the best target for reduction processes. This prediction is an important difference to accounts of reduplication based on BR-faithfulness (McCarthy & Prince 1995).