Listeners consider AI-made music less expressive than near-identical human-made content

This presentation considers why and how listeners view musical content differently when they believe it be generated by an AI. Empirical studies (e.g., Shank et al. 2023) and historical commentary (e.g., Benjamin 1968) suggest that consumers find AI-made content less useful and expressive than human-made content. However, much of this research assumes noticeable differences between AI-made and human-made content, a less tenable assumption in the last few years. I describe two experiments that test listeners’ reactions to technically comparable music made by humans and AI. In Experiment 1, half of participants were told a piece was composed by a human and half were told the same piece was by an AI. Participants found the piece less expressive when attributed to AI. In Experiment 2, 650 participants were given two excerpts and were asked which was created by a human versus by AI, justifying their choices using technical or expressive qualities. Half of participants was shown actual AI/human pairs, while the other half was presented with two excerpts both by AI or by a human. Analysis shows that even when there were no observable technical differences, listeners believed that the (supposedly) AI-generated content was less expressive. I argue that these sorts of distinctions will be increasingly important as AI-generated content becomes more sophisticated and that experiments like these can be used to advocate for good governance of AI companies.