Page 142 - Weiss, Jernej, ur./ed. 2025. Glasbena interpretacija: med umetniškim in znanstvenim┊Music Interpretation: Between the Artistic and the Scientific. Koper/Ljubljana: Založba Univerze na Primorskem in Festival Ljubljana. Studia musicologica Labacensia, 8

glasbena interpretacija ... | music interpretation ...
to (yet not referencing) a prediction that by 2025 more than 90 % of all web
30
content may be AI-generated to some extent. In September 2024,
[t]he creator of an open source project that scraped the internet to de-
termine the ever-changing popularity of different words in human lan-
guage usage says that they are sunsetting the project because generative
AI spam has poisoned the internet to a level where the project no longer
has any utility. 31
How can model collapse affect AI-generated music, and in particular
issues of musical interpretation? Taking the example of AI-generated Frank
Sinatra songs, if an AI system trains itself for the task of creating a “new” Si-
natra by checking any Sinatra song that has been uploaded to the web this
is fine as long as those are all “real” Sinatra recordings. If, however, some
of them already are “AI Sinatras”, this means model collapse is starting to
set in. If the number of AI Sinatras keeps increasing while the number of
“real” Sinatra songs remains the same (after all, he is not producing new re-
cordings anymore), we enter the spiral of AI training itself more and more
on its own previous creations, with the likely result that a homogenised,
“standard” Sinatra – the golden retriever of Sinatras, so to speak – is going
to emerge, and the specific idiosyncrasies or traits that only occur in a few
songs will be more and more suppressed. A “one and only” AI way of how
Sinatra sings will be the result, ultimately followed by Constantino’s “com-
plete devolution to nonsensical pablum.”
Some AI-created songs – usually those meant to demonstrate the abil-
ities of the AI systems – are publicly designated as such, although I doubt
whether AI systems are able to distinguish those from the recordings by the
real artists. However, most AI-generated music is not identified as such an-
yway. This particularly applies to functional music used in areas like adver-
tising, or background music of any kind such as contemporary music. Here
the risk of music entering the model collapse spiral is real, and if research-
ers such as Shumailov are correct we will all witness the ever-increasing
blandness and eventual degeneration of these musics in the not-too-distant
future (yes, those musics are already bland today, yet it can still get worse).
Even in the case of AI imitating well-known artists like Sinatra it is likely
30 Constantino, “Is AI Quietly Sabotaging Itself.”
31 Jason Koebler, “Project Analyzing Human Language Usage Shuts Down Because
‘Generative AI Has Polluted the Data,’” 404 Media, September 19, 2024, https://www.
404media.co/project-analyzing-human-language-usage-shuts-down-because-gen-
erative-ai-has-polluted-the-data/.

142