Emotional fidelity in AI-Generated music: A comparative analysis of conveyance and induction across models and modalities

The introduction of Generative Adversarial Networks has accelerated the development of AIGC, enabling creative assistance or even replacement in various fields including music. However, while AIGC-based music generation systems allowed users to input specific prompts to describe emotions, these models did not explicitly address or evaluate whether the generated emotions matched the emotion labels. This study aims to investigate the ability of AI music generated technology in emotion conveyance and induction. Three music generated software ( Google MusicLM , Stable Audio, MusicGen ) and two generated forms ( text-to-music, music-to-music ) were used to generate 75 AI music clips with four typical emotion labels (Energetic, Distressed, Sluggish, Peaceful) as prompts in 22 participants. The results show that Energetic AI music had significantly higher accuracy of conveyance than Sluggish and Peaceful in text-to-music , and its accuracy in text-to-music was higher than in music-to-music . The accuracy of Google MusicLM is remarkably higher than MusicGen . The Sluggish has the lowest accuracy in both forms. For emotion induction, interaction effects between Software and Prompt were observed, with Google MusicLM significantly higher than MusicGen for Energetic, and Energetic was significantly higher than Sluggish within Google MusicLM . These suggest that AI music can convey all emotions, but it may lead to significantly higher accuracy in conveying Energetic AI music as humans' higher sensitivity to positive-valence positive-arousal emotions. AI music's emotion conveyance ability may further influence its emotion induction. Compared to the other two software, Google MusicLM 's higher accuracy of conveyance allows it to more effectively induce emotional responses in both forms. Ultimately, these findings inform the design of more anthropomorphic and emotionally intelligent systems, where AI-generated music can serve as a crucial non-verbal cue to foster deeper and more effective human-machine emotional connections. • The emotional fidelity of AI-generated music varies significantly across different emotional categories. • Conveyance accuracy is highest for high-arousal, positive-valence emotions (‘Energetic’) and lowest for low-arousal emotions (‘Sluggish’). • A model's emotion conveyance accuracy is positively associated with its effectiveness in emotion induction. • AI music models exhibit significant variability in emotion conveyance and induction.