Computational creativity has passed through two eras defined by an inverse relationship between generative reach and interpretability. In the formal era, creativity was built from explicit representations — bisociation operationalized as conceptual blending, Boden's conceptual spaces formalized as creative systems, category-theoretic invention in COINVENT — yielding systems interpretable by construction but narrow in scope. Large language models brought the bargain to its extreme: trained on web-scale text, they produce fluent creative output across every domain, yet generate without exposing the mechanism linking input to output. This survey reads the field through that trade-off, tracing three foundational questions — how creativity is generated, measured, and explained — across both eras, and surveys the field's open challenges: reliable evaluation, output homogenization, ethics of creative AI, human-AI co-creativity, the status of machine creativity, and multimodal reach. We argue that one promising direction is to recover the interpretability of the formal era within the generative power of learned models.
W Sebastian (Thu,) studied this question.