Machine Neuroaesthetics Is Advancing Computational Creativity

Computational creativity has long been constrained by a persistent evaluation problem: machines can generate artifacts, but they lack an operational account of aesthetic value. Existing systems often rely on external proxies such as human ratings, engagement metrics, learned preference models, or hand-coded novelty scores, leaving aesthetic judgment opaque, unstable, and difficult to audit. This position paper argues that machine neuroaesthetics can advance computational creativity: the mechanistic study of aesthetic, affective, and interestingness representations inside generative models. Machine neuroaesthetics inverts computational neuroaesthetics: instead of using machine learning to explain how human brains respond to art, it uses mechanistic interpretability to examine how artificial neural networks internally represent and process aesthetic structure---structure inherited from the human culture compressed into them during training. Recent findings on sparse autoencoder features, functional emotion concepts, musical representations, audio-model features aligned with human neural data, poetry-planning circuits, and foundation-model-guided open-ended discovery suggest that large generative models already encode causally accessible structures relevant to aesthetic evaluation. We argue that these structures offer a new substrate for computational creativity: aesthetic judgment can become not only predicted, but localized, intervened upon, compared across models, and validated against human response. This does not imply that machines possess subjective aesthetic experience. Rather, it means that their learned aesthetic machinery can be made inspectable. Machine neuroaesthetics therefore reframes creative AI from artifact generation toward causal, auditable, and human-grounded aesthetic evaluation.