Page 138 - 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 ...
Turing test – and it has by now become very good at that. However, it is in-
capable of determining whether what it says is actually correct or true. This
criterion is just not built into the system, and there is no way that the statis-
tical method it is based on could achieve this result. Most of what it says is
likely to be correct, yet any text on a topic I know something about virtually
always includes one or several so-called “hallucinations”, that is statements
which are incorrect (I do not like the term “hallucinations” very much as it
seems to indicate the presence of a conscious “something” in the machine,
which is not the case). Generative AI systems can thus be described as the
ultimate bullshitters – they sound incredibly convincing, yet create their
19
output without any concern for what is correct or true. Another poten-
tial issue with generative AI is that it is difficult – if not impossible – for it
to be really creative or original. This is because what it creates is based on
what already exists. Postmodern combinations of unusual and distant sty-
listic traits can be regarded as original and creative, of course, yet genera-
tive AI does not go for this kind of “outlier” – instead the texts, pictures and
music generated by AI often appear to be cursory and stereotypical. If they
are given the same prompt repeatedly the results are never exactly identi-
cal, yet often very similar with the differences being of a superficial nature
that affect (in the case of texts) wording more than content. Texts as well
as pictures often come in a very anodyne style lacking the quirks and idio-
syncrasies that most human authors and creators have, and that give a text
a personal note. The nature of the algorithm ensures that such a personali-
sation cannot occur – instead it is at risk of evolving towards the so-called
“model collapse” which will be discussed in the next section.

Model Collapse And The Standardisation Of AI Results
How can AI relate to issues of musical interpretation? Since AI is still rela-
tively new, the answer has to some degree be speculative, particularly with
regard to music. While the number of AI-generated texts is already vast, the
amount of AI-generated music is much lower – partly because there is more
demand for AI-generated texts, yet also because LLMs started generating
pictures and music later than texts. Yet since, as we have seen, all LLMs

19 I use the term “bullshitting” here in the sense in which the philosopher Harry G.
Frankfurt defined it in his essay On Bullshit (Princeton, NJ: Princeton University
Press, 2005). He distinguishes bullshitting from lying by outlining that bullshitting
doesn’t require any knowledge of the truth and isn’t concerned with it at all. This
makes bullshitters more dangerous than liars; populist politicians are prime exam-
ples of bullshitters in today’s world.

138