Recu si e Misin e p e a ion II
Ca los G. Román
Expe imen al Music & Digi al Media
Louisiana S a e Uni e si y
c [email p o ec ed]
Robe o Moche i
College o Music & D ama ic A s
Louisiana S a e Uni e si y
[email p o ec ed]
Jesse Allison
Expe imen al Music & Digi al Media
Louisiana S a e Uni e si y
[email p o ec ed]
Abs ac
This pa icipa o y pe o mance piece builds on a se ies o p ocess-based
composi ions explo ing he c ea i e ension be ween human in en ion and
AI-gene a ed ou comes in ex - o-audio gene a ion. The o iginal wo k employs a
ecu si e s uc u e whe e p omp s guide al e na ing con ibu ions om a human
compose and a gene a i e AI model, emb acing misin e p e a ions as c ea i e
d i e s. Fo his i e a ion, he p ocess is ex ended in o a li e, dis ibu i e o ma
inco po a ing eal- ime audience pa icipa ion. Audience membe s submi
ex ual desc ip ions, which a e syn hesized by a la ge language model in o a
single p omp ha is hen used o gene a e new audio ia a ex - o-sound model.
This AI ou pu becomes he ounda ion o he nex cycle o p omp s, c ea ing a
eedback loop be ween human pe cep ion, language in e p e a ion, and machine
gene a ion. The piece emphasizes he a is ic alue o seman ic gaps and
misalignmen s in AI in e ac ion, p oposing gene a i e miscommunica ion
be ween humans and AI models as e ile g ound o collabo a i e sonic
explo a ion.
1 O igins and concep ion o he piece
This dis ibu i e pe o mance piece o igina es om an idea explo ed in a se ies o wo ks called
Recu si e Misin e p e a ion (2025) by B azilian compose Robe o Moche i, which we e ini ially
s uc u ed as p ocess pieces, guided by a se o ins uc ions a he han a ixed sco e. The
concep ual basis o hese pieces explo es he in e p e i e gap be ween human in en ion and
machine-gene a ed ou comes in ex - o-audio gene a ion. In he piece, he compose and an AI
model al e na e in gene a ing musical mo emen s. S a ing om a simple echnical p omp , he
compose c ea es a i s sec ion, hen uses he p omp o gene a e audio ia an AI model. The
compose hen lis ens o he AI ou pu and w i es a new p omp based on hei in e p e a ion o
i — his guides he nex human-composed mo emen . The p ocess epea s, wi h each new p omp
de i ed om he p e ious AI ou pu , con inuing un il he esul s eel comple e o he AI ou pu
becomes unin e es ing. The piece ends wi h an AI-gene a ed sec ion. The e o e, he ini ial p omp
se es no as a bluep in , bu as a poin o depa u e o i e a i e, dialogic c ea ion. This i e a i e
use o p omp s has been employed as a p ocess o explo a ion and expe imen a ion in di e en
a eas o knowledge (e.g. Hu son & Co oneo, 2023), allowing non-expe use s o be e g asp he
c ea i e po en ial and bounda ies o gene a i e AI models.
Fo his con e ence, we p opose ex ending his composi ion p ocess in o a li e, in e ac i e o ma
whe e he audience con ibu es p omp s in eal- ime h ough a web in e ace. We a e also inspi ed
by he in eg a ion o mul imodal gene a i e ools in di e en scena ios (e.g. S u m e al, 2024).
Fi s ly, audience membe s desc ibe he ini ial agmen o a po en ial piece in hei own wo ds.
These p omp s a e no ed di ec ly in o he gene a i e model, ins ead, all submi ed p omp s a e
P oceedings o he 6 h Con e ence on AI Music C ea i i y (AIMC 2025),
B ussels, Belgium, Sep embe 10 h-12 h
sen o an open la ge language model (LLM), which summa izes and syn hesizes hem in o a
single cohe en p omp ha desc ibes wha he consensus o he audience pe cei es a a pa icula
momen . This collec i e p omp is hen sen o he ex - o-sound open model o gene a e a new
audio piece, which ge s mixed in eal- ime wi h he p e ious agmen . The p omp s submi ed and
he collec i e p omp s gene a ed a e also isualized and p ojec ed in eal- ime o he audience. The
p ocess can be epea ed i e a i ely: he newly gene a ed piece becomes he basis o ano he ound
o audience p omp s, o ming a eedback human-in- he-loop be ween human desc ip ion, LLM
in e p e a ion, and AI sound gene a ion. This wis opens possibili ies o unp edic abili y,
au ho ship dispe sion, and collec i e c ea i i y. Fo he con e ence, we p opose an 8- o-10-minu e
e sion o he piece.
2 Though s abou he p ocess and ole o he compose /audience
The in en ion wi h his dis ibu i e pe o mance is ha he audience in e ac s wi h AI gene a i e
ools ( ex - o-music models), using s uc u ed p omp s o elici musical ma e ial. Ra he han
co ec ing o ejec ing AI “e o s” o “miscommunica ions” in he gene a ion, he p ocess
emb aces hem: di e gences and unexpec ed sonic esul s a e in e p e ed as c ea i e oppo uni ies
o mo e he piece o wa d. These “misdi ec ions” be ween he p omp and he sonic ou come a e
ecu si ely in eg a ed in o he composi ion, in o ming he nex ins uc ion o sec ion o he
piece. The compose and he audience assume a hyb id ole as au ho s, edi o s and cu a o s o he
AI gene a ed pieces, engaging in a eedback loop wi h he AI sys em.
Ano he key ques ion explo ed h ough his se o p ocess pieces: Can we na iga e and
meaning ully unde s and he la en space o a ex - o-music model using p omp s alone? Cen al o
he p oposed pe o mance is he acknowledgmen o a seman ic gap (Allison e al, 2024) be ween
he audience's in en ion, he w i en p omp , and he AI's audio ou pu . We also wan o highligh
he con as be ween u ili a ian uses o gene a i e AI music ools and expe imen al, p ocess-based
c ea i e p ac ices like he one p esen ed in his piece. This piece builds on op o he o iginal wo k
by using gene a i e AI and p omp s o allow a synch onous collabo a ion be ween audience
pa icipan s, ins ead o he AI/compose collabo a ion explo ed in he ini ial piece.
3 Conclusion
This pe o mance highligh s he gene a i e po en ial o human-AI misalignmen in c ea i e
p ocesses. I aims a demons a ing how AI ex - o-music models can unc ion less as a ool o
ealizing ixed ideas and mo e as a collabo a o o sonic explo a ion in an e ol ing a is ic
dialogue. Because o he limi a ions o cu en AI gene a i e models, di e en c ea i e app oaches
can be used o add a new dimension o he human-in- he-loop dialogue. The me hodology sugges s
possibili ies o collabo a i e c ea ion o dis ibu ed au ho ship in u u e pe o mances. The
app oach in i es e lec ion on au ho ship, con ol, and in e p e a ion in he age o gene a i e music
echnologies.
Acknowledgmen s
Re e ences
Allison, J., Fa a , D., Nash, T., Román, C. G., Weeks, M., & Ju, F. X. (2024). Play Me Some hing Icy:
P ac ical Challenges, Explainabili y and he Seman ic Gap in Gene a i e AI Music. P oceedings o he 2nd
In e na ional Wo kshop on eXplainable AI o he A s, ACM C ea i i y and Cogni ion Con e ence, Chicago,
Illinois, June 23, 2024 a Xi p ep in a Xi :2408.07224. h ps://a xi .o g/pd /2408.07224
Hu son, J., & Co oneo, P. (2023). Gene a i e AI ools in a educa ion: Explo ing p omp enginee ing and
i e a i e p ocesses o enhanced c ea i i y. Me a e se, 4(1).
S u m, B., Ame o i, M., Dalmazzo, D., C os Vila, L., Casini, L., & Kanho , E. (2024). S ochas ic Pi a e
Radio (KSPR): Gene a i e AI applied o simula e comme cial adio. In P oceedings o he 5 h Con e ence on
AI Music C ea i i y, Ox o d, England, 2024.
2
