Affective Bond Between AI-Generated Music and Human Listeners

Abstract

Generative artificial intelligence is being applied to create music in an ever-widening range, but it has not been determined whether the results can effectively convey human emotions. A controlled-variable experimental design is used to explore the causal relationship between information density in prompts and listeners' affective recognition of AI-generated music. Six AI-generated musical works were selected as the experimental stimuli under three increasing levels of prompt conditions: minimal emotional descriptions, prompts that added harmonic progression information, and detailed prompts including parameters such as instrumentation, vocal processing and dynamic changes. The results showed that listeners could not understand the creators' purposes in either the minimal-prompt or harmony-augmented situations. All 30 listeners were able to feel the emotion and structure that were desired, and they could do so after repeated instructions. The study also found that the audio outputs of AI music systems generally have logical discontinuities and perceptual incongruity without fine-grained human encoding intervention. This paper proposes that AI functions as a medium for music creation, and its affective communication performance is based on a closed-loop human intervention system at two stages: prompt engineering (information encoding) and generative evaluation (signal decoding). This study offers empirical support and practical paths for human-AI collaboration models in AI-assisted music creation.