Statistical Monoculture: How Text-to-Image AI Flattens Cultural and Aesthetic Diversity

AI image generators produce outputs that look strikingly alike, flattening cultural and aesthetic difference into a generic "AI aesthetic". This paper develops a critical account of that flattening as a statistical monoculture: a regime in which a single aesthetic centre of mass dominates all representations through the mathematics of the generative objective. We build on a recent measurement study of text-to-image diversity collapse: the mean-squared-error denoising objective computes a probability-weighted average of aesthetic modes (the Mode Averaging Principle), and commercial models accordingly compress within-prompt outputs into roughly ten of 768 effective dimensions, render minority cultures 17% less diversely than majority ones under CLIP embeddings (attenuating to about 8% under DINOv2), and converge across two independently developed systems to a shared aesthetic centre. Taking this mechanism as given, we ask what it means for visual culture, representation, and governance. Reading the generative process through theories of platform realism and colonial mimicry, we show that model outputs are "almost the same, but not quite" as authentic cultural expression, and that the gap of difference grows with a culture's marginalisation in the training data. We situate the "ground truth" of these systems within actuarial logics of risk and standardisation, and argue that aesthetic homogenisation is a corollary of the models' structure and not merely of biased data. We close by drawing out implications for sovereign-AI initiatives and for regulation that operationalises fairness at the level of who appears, not how diversely they are rendered.