Shepard’s (1987) universal law of generalization holds that the probability of generalizing between two stimuli decays as a concave function of their distance in psychological space. While there is widespread evidence for the law in human perception, its relevance to artificial neural networks remains unclear, despite the importance of generalization for these systems. Here, we find that the representational spaces of models that vary in their architecture, objective, and training data yield a concave generalization gradient with respect to human judgments of naturalistic images (Peterson et al., 2018), consistent with Shepard’s law. Our results suggest that the representational spaces of deep vision networks serve as compelling, but imperfect, proxies for classic psychological spaces derived from behavioral data. This highlights the strengths and weaknesses of deep vision models as contributors to cognitive theories of perceptual generalization, while adding further evidence for the generality of Shepard’s law.