Page 143 - 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

ai and musical interpretation
that model collapse will negatively affect songs the later they are being pro-
duced. In those cases model collapse will probably occur at a much slow-
er pace since songs of this kind are created in much smaller numbers than
advertising or background musics, yet it will eventually become noticea-
ble too. And if eventually a homogeneous AI standard of Sinatra’s singing
emerges there is even a risk that people who did not hear his recordings be-
fore the emergence of AI (necessarily an ever-increasing cohort over time)
may take it to be the real thing. All of this can, of course, be applied not just
to voices, yet also to the sounds and styles of instrumentalists or ensembles
(such as specific big bands, for example) in all genres.
Imitating the voice of a specific singer runs into copyright problems,
and Tennessee has recently become the first US state to restructure its copy-
right law and ban the unauthorised use of AI to imitate the work and style
32
of copyright-protected artists. However, there will always be states and
territories which have not signed up to international copyright agreements.
Servers located there can be used to distribute unauthorised imitations that
are illegal elsewhere, and attempts to block them will probably be ultimate-
ly unsuccessful.

Conclusion
Generative AI has the potential to limit the range of musical interpreta-
tion. At least for the time being “deepfake” imitations of performers’ voic-
es are often not entirely convincing. While this aspect is likely to get better
over time, two other problems cannot be overcome that easily. Firstly, the
AI systems “compose” music by “averaging out” the data they study while
operating, meaning that a certain homogenisation of the results is virtual-
ly inevitable. This applies both when the style of specific pieces or perform-
ers is imitated, yet also when music in certain genres or styles is request-
ed. Secondly, it is very likely that – as the number of AI-created pieces of
music increases – an onset of model collapse, that is a spiral of self-refer-
entiality, will be set in motion that will initially reinforce the homogenisa-
tion effect before potentially leading to a collapse of the systems in the long
run. Generative AI has parameters that give it a certain leeway when select-
ing results, yet comparing different texts generated in response to the same
32 Rebecca Rosman, “Tennessee Becomes the First State to Protect Musi-
cians and Other Artists Against AI,” NPR, March 22, 2024, https://www.npr.
org/2024/03/22/1240114159/tennessee-protect-musicians-artists-ai. The new law is
known as the ELVIS act, an acronym for Ensuring Likeness Voice and Image Secu-
rity Act.

143