Current AI-generated music has been found to be inferior to music that has been composed by humans, according to researchers at the University of York.
They have also demonstrated that there are flaws in AI music generation algorithms that could violate copyright, and they have created guidelines to assist others in assessing the systems they are using.
In the study, 50 highly musically knowledgeable participants listened to musical excerpts, some of which were taken from authentic human-composed works and others of which were produced by deep learning (DL), a kind of artificial neural network, and non-DL algorithms.
To ensure that the study’s findings went beyond just musical expression, experts in note content and stylistic success were sought out as participants.
“On examination, the scores for human-composed snippets are much higher and stylistically more effective than those for any of the systems responsible for computer-generated excerpts.”
Co-author Dr. Tom Collins, from the School of Arts and Creative Technologies at the University of York,
Musical standards.
The listeners were not informed as to whether the excerpts were created by humans or computers and were instead asked to rate them based on six musical criteria: stylistic success, aesthetic pleasure, repetition or self-reference, melody, harmony, and rhythm.
Dr. Tom Collins, a co-author from the University of York’s School of Arts and Creative Technologies, stated, “On analysis, the ratings for human-composed excerpts are noticeably higher and stylistically more successful than those for any of the systems responsible for computer-generated excerpts.”.
Additionally, the study’s conclusions raise questions about possible ethical transgressions caused by deep learning methods that directly copy data. It has been demonstrated that the output of a common type of DL architecture known as transformer—the same type of architecture powering OpenAI’s ChatGPT—copies significant portions of the training data.
Legal and moral.
Dr. Collins explained, “If Artist X uses an AI-generated excerpt, the algorithm that generates the excerpt might accidentally copy a section of a song in the training (input) data of Artist Y. Artist X will unknowingly violate Artist Y’s copyright if they release their song.
It is a worrying finding that might imply that companies that create algorithms ought to be subject to oversight or ought to police themselves. They are aware that these algorithms have flaws, so the emphasis should be on fixing them so that AI-generated content can still be created, but in a morally and legally acceptable manner.”.
Seven guidelines have been offered by the study’s researchers for conducting a comparative analysis of machine learning systems. The results might address current ethical concerns, advance the development of AI-generated music, and help steer clear of copyright infringement legal snags in the future.
The research was published in the journal Machine Learning.
More information: Zongyu Yin et al, Deep learning’s shallow gains: a comparative evaluation of algorithms for automatic music generation, Machine Learning (2023). DOI: 10.1007/s10994-023-06309-w
