Internet visual culture has produced a sprawling vocabulary of "aesthetics" such as Frutiger Aero, vaporwave, cottagecore, and liminal spaces, each a visual genre with a distinct relation to the past. Whether this folk vocabulary has a measurable affective architecture has been argued only qualitatively. Using naturally-occurring human behaviour, we reconstruct the aesthetics ecosystem as a network of 1,115 aesthetics from community cross-referencing. The network is strongly modular, organised by latent affect (p = 0.0005) and historical era (1920s to 2010s), and its unsupervised communities align with an a priori four-mode taxonomy (restoration, hauntology, anemoia, uncanny inversion; NMI = 0.50). A second, independent record of user co-curation matches the same communities above chance (NMI = 0.21 vs. a 0.17 null, p = 0.032). To test whether the structure is real, a frontier panel (Gemini 2.5 Flash, GPT-4o, Claude Sonnet 4.6) predicts the aesthetic that people assign each image to from affect ratings alone (accuracy 0.45 vs. 0.29 chance, z = 18, n = 1,727), matching a 29-feature perceptual baseline. The same models over-attribute unease: they reach near-ceiling recall on the uncanny (0.86) but read warmth and longing as eerie (restorative 0.35, anemoia 0.31). The bias is interpretive: it is absent from image statistics, it persists when the image is edited, and it falls sharply when the rating task is reframed; on standardized affective images the same models match human valence norms (r = 0.84 to 0.87). The over-attribution, its boundary, and its causal mechanism replicate across all three models. Internet aesthetics thus have a real, cross-source affective architecture, readable even to systems trained on none of it, which in reading it reveal a diagnosable and correctable bias toward the uncanny. We release the network, corpora, model ratings, and code.