Generative AI refers, in the context provided, to the kind of output system commonly called “generative AI” today: a machine-learning-based system, especially one built on a large deep learning model trained with large-scale data. Although AI is originally a very broad research field, the passages describe generative AI more specifically as a system that produces outputs by optimizing them according to given data, objective functions, and evaluation criteria.
This makes the generation process strongly top-down, because a human typically supplies the desired direction from above, such as what kind of output should be obtained. The model’s internal representations and decision-making process are often difficult for humans to interpret, so it frequently functions as a black box. In the artistic examples provided, AI is used for sound generation by training on paired datasets of real train-window video and recorded ambient sound. The output is then adjusted through parameter tuning to balance fidelity, synchronization, and naturalness while avoiding overly arbitrary human intervention.
At the same time, the passages note a tension: once the AI output is fine-tuned, human value judgments inevitably re-enter the process, which can weaken the original question the artist wanted to ask of the AI. The provided context does not offer a broader technical history or a full general definition beyond these points.