An “AI” label fails to trigger negative bias in new pop music study

A study in Singapore found that, contrary to expectations, study participants rated pop songs labelled as AI-generated more highly in positive emotions compared to pop songs (which were also AI-generated) that were labelled as human-composed. The positive emotions included happiness, interest, awe, and energy. Consequently, this study found no evidence of negative bias towards AI-generated music. The paper was published in Computers in Human Behavior: Artificial Humans.

In recent years, the use of artificial intelligence (AI) systems to generate music has been gradually transforming the way in which music is composed. AI music generation systems use machine learning models to create melodies, harmonies, rhythms, and full musical pieces based on patterns learned from large collections of human-made music.

These systems can compose in many styles, from classical orchestral arrangements to electronic beats or film-score atmospheres. Many musicians have started using AI as a creative tool to quickly explore ideas or generate variations they might not think of on their own.

However, some people view AI-generated music with skepticism because they fear it lacks genuine emotion or human intention. Others worry that AI tools could replace human artists or reduce appreciation for traditional craftsmanship.

This leads to a theoretical “negative bias” where people judge AI music more harshly than they would judge the same piece if they thought a human composed it. Sometimes, people rate a piece of music as less creative when told it was made by AI, even if the sound is identical to a piece they rated more favorably.

Study author Suqi Chia and their colleagues wanted to explore whether there really is a negative bias towards AI-generated music. They conducted a study in which they used the same set of AI-generated pop songs but randomly decided which to label as human-composed and which to label as AI-generated.

They reasoned that if people have indeed learned to associate AI-generated music with being less emotional, less immersive, or less meaningful, then labelling a song as AI-composed should activate these expectations and influence the evaluations.

The study participants were 64 university students. Forty-six of them were women, and the participants’ average age was 20 years. Eighty-three percent ranked pop music among their top three favorite genres. While the initial number of participants was higher (90), some were excluded because they correctly guessed the purpose of the study or because they were not paying attention.

The study authors used Suno AI, a program designed to generate songs with vocals and instrumentation, to create eight pop songs. All songs were created using the prompt “pop genre, happy and chill”. Half of the songs featured female vocals and the other half featured male vocals. In this experiment, participants listened to song excerpts that were 30 seconds long, taken from the chorus section of each song.

Participants listened to all eight excerpts in a randomized order. However, for four of the excerpts, they were told that the song was composed by a human, while for the other four, participants were told that the songs were AI-generated.

The decision of which songs were labeled as human-composed and which as AI-generated was randomized. For each song, the study authors also used a made-up composer name, sounding either like an AI (e.g., TuneSoft) or like a human composer (e.g., Victoria Moore).

After listening to each piece, the participants’ task was to provide ratings on how much they liked the music, their emotional responses (happiness, interest, awe, and energy), their sensorial responses, whether the music created pictures in their minds, their experiential responses (e.g., “I felt as if I were part of the song”), their need to re-experience the song (e.g., “I would enjoy listening to this song again”), and how likely they would be to purchase the song.

Contrary to expectations, the results showed no negative bias towards songs labelled as AI-generated. Moreover, participants rated songs labelled as AI-generated more highly in positive emotions—including happiness, interest, awe, and energy—compared to those labelled as human-composed.

“These results suggest that while the perception of AI authorship does influence listeners, the effects are primarily affective rather than sensorial, imaginal, experiential, or behavioral. Notably, considering that listeners rated pop songs labelled as AI-generated more positively in emotions, the findings imply that AI-generated music may be more readily accepted than previously assumed,” the study authors concluded.

The study contributes to the scientific understanding of people’s perceptions of AI music. However, it should be noted that study participants were a small group of university students from Singapore. Results for other demographic groups or people from other cultures might differ.

The paper, “Do listeners devalue AI-generated pop music? Exploring negative biases in listeners’ responses to AI-labelled vs human-labelled pop music,” was authored by Suqi Chia, Andree Hartanto, and Eddie M.W. Tong.