EXPANDING THE HAISP DATASET: AI’S IMPACT ON SONGWRITING
ACROSS TWO AI SONG CONTESTS
Lidia Mo is1Michele Newman1Xinya Tang1
Renee Singh1Ma cel Vélez Vásquez2Rebecca Lege 3Jin Ha Lee1
1In o ma ion School, Uni e si y o Washing on
2Uni e si y o Ams e dam
3F aunho e Ins i u e o In eg a ed Ci cui s
[email p o ec ed], [email p o ec ed]
ABSTRACT
As a i icial in elligence (AI) con inues o shape c ea i e
p ac ices, unde s anding i s ole in human-AI songw i ing
emains c ucial. This pape expands he Human-AI Song-
w i ing P ocesses (HAISP) da ase by inco po a ing da a
om he 2024 AI Song Con es , building upon he o iginal
2023 da ase . By analyzing new submissions, we p o ide
u he insigh s in o AI’s e ol ing impac on songw i ing
wo k lows, c ea i e decision-making, and con ol. A com-
pa a i e s udy o AI ool usage and pa icipan s a egies
be ween he 2023 and 2024 con es s e eals shi s in col-
labo a ion pa e ns and ool e ec i eness. Addi ionally, we
assess he di e ences be ween gene al-pu pose AI sys ems
and pe sonalized, ine- uned ools, highligh ing hei im-
pac on c ea i e agency. Ou indings o e key design im-
plica ions o AI-assis ed songw i ing ools, p o iding ac-
ionable insigh s o AI de elope s and music p ac i ione s
seeking o enhance co-c ea i e expe iences.
1. INTRODUCTION
A i icial In elligence (AI) has apidly become an in eg al
componen o c ea i e ields, eshaping a is ic exp ession
ac oss a ious domains. F om isual a s o li e a u e, AI-
powe ed ools a e being le e aged o augmen human c e-
a i i y, aising new ques ions abou au ho ship, o iginali y,
and he e ol ing na u e o co-c ea ion [1–3]. Nowhe e is
his ans o ma ion mo e e iden han in music composi-
ion, whe e AI sys ems a e inc easingly employed o gen-
e a e melodies, ha monies, ly ics, and en i e song s uc-
u es [4, 5]. These ad ancemen s ha e gi en ise o new
o ms o collabo a ion be ween human musicians and AI,
necessi a ing a deepe unde s anding o he dynamics o
human-AI co-c ea ion in songw i ing.
The s udy o human-AI collabo a ion in music is pa -
icula ly impo an due o he complex, o en subjec i e
© L. Mo is, M. Newman, X. Tang, R. Singh, M.A. Vélez
Vásquez, R. Lege , and J.H Lee. Licensed unde a C ea i e Commons
A ibu ion 4.0 In e na ional License (CC BY 4.0). A ibu ion: L. Mo -
is, M. Newman, X. Tang, R. Singh, M.A. Vélez Vásquez, R. Lege , and
J.H Lee, “Expanding he HAISP Da ase : AI’s Impac on Songw i ing
Ac oss Two AI Song Con es s”, in P oc. o he 26 h In . Socie y o
Music In o ma ion Re ie al Con ., Daejeon, Sou h Ko ea, 2025.
na u e o he c ea i e p ocess. While AI can accele a e
composi ion wo k lows and gene a e no el musical ideas,
i s ole in enhancing e sus eplacing human c ea i i y e-
mains a c i ical a ea o in es iga ion [6, 7]. Fu he mo e,
ques ions ega ding ou changing de ini ion o compu a-
ional c ea i i y and he ole AI can play in he c ea i e
p ocess as a ool o collabo a o con inue o loom la ge
o e he ield [8, 9]. Add essing hese conce ns equi es
quali a i e da a ha cap u es no jus empi ical in o ma-
ion on he use o gene a i e AI, bu he li ed expe iences
o c ea o s wo king wi h AI in music p oduc ion.
To con ibu e o his g owing ield o s udy, he Hu-
man–AI Songw i ing P ocesses (HAISP) da ase was in-
oduced in 2024 as a cu a ed esou ce designed o ex-
plo e he in e ac ion be ween human musicians and AI sys-
ems [10]. The da ase was de i ed om submissions o
he AI Song Con es 2023, an annual compe i ion ha in-
i es eams o musicians, da a scien is s, and esea che s
o explo e he c ea i e po en ial o AI in songw i ing [11].
I comp ises 34 coded en ies documen ing how eams
used AI ools in hei songw i ing p ocesses. I p o ides a
s uc u ed amewo k o analyzing a ious aspec s o AI-
assis ed music c ea ion, including:
• The speci ic AI ools and models used in composi-
ion
• The songw i ing me hodologies employed by
human-AI eams
• Re lec ions on e hical conside a ions and challenges
ela ed o AI in music
• Teams’ assessmen s o hei collabo a i e expe i-
ence wi h AI
The indings om he HAISP da ase highligh ed he di-
e se ways in which AI is in eg a ed in o songw i ing, wi h
eams using AI o asks anging om melody gene a ion
o pe o mance syn hesis. Howe e , he da ase also unde -
sco ed he limi a ions o AI ools, such as lack o c ea i e
con ol, echnical limi a ions, and conce ns abou o igi-
nali y. In addi ion, e hical conce ns ega ding he p o e-
nance o da a and he anspa ency o AI-gene a ed con en
eme ged as key hemes.
28
Building on his ounda ion, he cu en s udy ex ends
he HAISP da ase by inco po a ing new da a om he
2024 AI Song Con es , o e ing a longi udinal pe spec i e
on he e olu ion o human-AI collabo a ion in songw i ing.
By compa ing da a om 2023 and 2024, his expanded
da ase enables a deepe analysis o ends, eme ging ech-
nologies, and shi ing a i udes owa d AI in c ea i e wo k.
Th ough his esea ch, ou goal is o p o ide aluable in-
o ma ion o musicians, AI de elope s, c ea i i y schol-
a s, and beyond.
2. BACKGROUND
The in e sec ion o AI and music composi ion ep esen s
a apidly e ol ing ield ha has a long his o y o explo e.
AI-assis ed music c ea ion has p og essed om ea ly algo-
i hmic expe imen s and academic elec onic music cen-
e s [12] o sophis ica ed machine lea ning models capa-
ble o composing comple e musical pieces by he b oade
public [13]. As hese echnologies become mo e accessi-
ble, hey no only in luence he way music is made bu also
aise c i ical e hical, cul u al, and a is ic ques ions abou
human-AI co-c ea ion.
2.1 E olu ion o AI in Music Composi ion
The applica ion o compu a ional echniques in music
composi ion can be aced back o he mid- wen ie h cen-
u y, when ea ly pionee s expe imen ed wi h algo i hmic
app oaches o sound gene a ion [13], such as he wo k
comple ed a he Columbia-P ince on Elec onic Music
Cen e , which laid he g oundwo k o a a ie y o com-
pose s ca ee s and echnological inno a ions [12, 14]. By
he ea ly 2000s, de elopmen s in machine lea ning acili-
a ed he c ea ion o models ha could au onomously gen-
e a e melodies, ha monies, and song s uc u es [15,16].
The inc easing sophis ica ion o deep lea ning and gen-
e a i e AI in he pas decade has u he ans o med he
landscape o music composi ion. No able ad ances in-
clude OpenAI’s MuseNe , Google’s Magen a, and Me a’s
MusicGen, all o which employ ans o me -based a chi-
ec u es o p oduce di e se composi ions [17–19]. These
ools enable musicians o collabo a e wi h AI in a ious
ways, om gene a ing musical ideas o assis ing wi h a -
angemen , and mo e [19]. The g owing accessibili y o
hese echnologies has been showcased in pla o ms such
as he AI Song Con es (AISC) [11].
2.2 AI Songw i ing Tools and Me hods
Va ious AI-powe ed ools ha e eme ged o acili a e
human-AI collabo a ion in songw i ing. OpenAI’s
MuseNe [20] is a deep neu al ne wo k capable o com-
posing mul i-ins umen al pieces ac oss mul iple gen es,
while Google’s Magen a p ojec p o ides open-sou ce ap-
plica ions o AI-assis ed melody gene a ion, cho d p o-
g ession, and hy hm c ea ion [18]. While hese ools o e
new c ea i e possibili ies, hey also in oduce challenges
ela ed o a is ic con ol, o iginali y, and he implica ions
o AI as a co-c ea i e en i y, especially when he use is
an inexpe ienced music c ea o [2, 11, 21]. AI-gene a ed
music is o en cons ained by i s aining da a, leading
o conce ns abou p edic abili y, s ylis ic homogeniza ion,
and he po en ial o AI o ein o ce exis ing musical con-
en ions a he han os e ue inno a ion [22, 23]. Fu -
he mo e, he ex en o which AI-gene a ed composi ions
can be conside ed “c ea i e” in he same sense as human-
au ho ed wo ks is s ill being ques ioned, especially when
i comes o jus how much o a ole he AI plays in he
composi ional p ocess [24,25].
2.3 C ea i i y S udies
The inc easing adop ion o AI in c ea i e domains has
spa ked deba es abou he na u e o c ea i i y and he ole
o machines in a is ic exp ession [11, 26–28]. T adi ional
iews o c ea i i y emphasize human in ui ion, cul u al
con ex , and emo ional dep h [29–31] - quali ies ha AI,
as a s a is ical modeling sys em, does no inhe en ly pos-
sess. Can AI uly be conside ed a c ea i e agen , o is i
me ely an ad anced ool o pa e n ecogni ion and ecom-
bina ion? E hical conce ns also ex end o he implica ions
o AI’s inc easing ole in he c ea i e wo k o ce [32, 33].
As AI-gene a ed composi ions become mo e sophis ica ed,
he e is a po en ial o au oma ion o displace human mu-
sicians in ce ain comme cial con ex s [34]. In esponse o
hese challenges, schola s and indus y p o essionals ha e
called o g ea e anspa ency in AI aining da a, e hical
guidelines o AI-assis ed composi ion, and policies o en-
su e ha human a is s emain cen al o he c ea i e p o-
cess [35].
3. DATASET EXPANSIONS: METHODOLOGY
The HAISP Da ase is accessible as a .cs and .xlsx on he
Open Science F amewo k (OSF) unde a C ea i e Com-
mons A ibu ion-NonComme cial 4.0 In e na ional (CC
BY-NC) license, which allows o b oad access and u i-
liza ion o esea ch pu poses [36].
We gene a ed he da ase ia consensus coding [37].
One esea che coded a selec ion o he da a en ies, col-
lec ing hem in o he da ase . A second code hen e-
iewed he ini ial coding, alida ing he coding by ei-
he ma king ag eemen o disag eemen wi h he cod-
ing choices wi hin a commen on he code in he da ase ,
adding wha hey el he code was missing wi hin hei
codes om he da a. In he case o disag eemen , a hi d
esea che helped decide on he inal code as a ie-b eake .
3.1 Da a Collec ion
Simila ly o he 2023 p ac ice, each eam had o ill in he
AI Song Con es 2024 Submission Fo m ia Google Fo ms
o pa icipa e in he con es [10]. The o m consis s o en y
ields ha co e all he basic in o ma ion abou eams and
songs:
• eam (bio o he websi e, loca ion, le el o expe -
ise, mo i a ion o pa icipa e, how hey hea d abou
he AISC);
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
29
• song ( i le, leng h, link o music ideo/soundcloud/
blogpos , concep /idea, ly ics, li e pe o mance).
Each eam was addi ionally asked o gene a e a p ocess
documen and sa e i as a PDF ile, which mainly includes
mo e de ailed mo i a ion, songw i ing wo k low and col-
labo a ion p ocess, e alua ion o co-c ea ion, and e hical
conside a ions. In his way, each eam had mo e space o
elabo a e on hei collabo a ion p ocess han in p e ious
submissions.
Fu he mo e, all eams had o gi e hei consen o hei
esponses o be published in a scien i ic pape . The an-
swe s we e collec ed in o a Google Shee wi h links o ou -
s anding PDF iles. In o al, 67 submissions we e collec ed,
including 34 submissions ha used ex - o-music models
such as Udio and/o Suno in he human-AI collabo a ion
p ocess and hus we e conside ed disquali ied due o he
judges inabili y o "assess he le el and manne o he use
o AI in each en y" and an inabili y o ob ain "a desc ip-
ion o he da a used o ain he AI model." [38] A e
hese 34 submissions we e excluded, 33 e ec i e pa ic-
ipa ing eams emain in he 2024 edi ion. The comple ed
ques ionnai es we e hen handed o e o he esea ch g oup
excluding pe sonal da a.
3.2 Me hodology and Valida ion
Due o he change in pa icipan en y me hods and sub-
mission o ma , he da a dic iona y om he ini ial HAISP
da ase was adap ed by he code s o be e ex ac he da a
essen ial o he da ase om he w i en en ies. The ca e-
go ies we e e alua ed one by one by he ou esea che s,
i e a ing h ee imes, wi h es ing o each new adap a ion
o he dic iona y be o e eaching inal consensus. Fo each
i e a ion o he da a dic iona y, wo code s es ed i on wo
sample en ies o ensu e ha he ca ego ies we e p ope ly
de ined and applicable o he new da a.
3.3 Da a S a is ics
The HAISP da ase o he 2024 edi ion consis s o da a
om 33 eams, ep esen ing 22 di e en coun ies and e-
gions. The Uni ed S a es has he highes ep esen a ion,
wi h 12 eams pa icipa ing. The Uni ed Kingdom ollows
wi h ou eams, while Swi ze land has h ee. Ge many
and Spain a e each ep esen ed by wo eams.
O he ep esen ed coun ies include he Ne he lands
(NLD), Colombia (COL), Japan (JPN), I aly (ITA), B azil
(BRA), Thailand (THA), Canada (CAN), F ance (FRA),
Tunisia (TUN), Denma k (DNK), Hunga y (HUN), Tu key
(TUR), China (CHN), and Chile (CHL).
The ype o a ilia ion o he HAISP da ase 2024 edi-
ion e lec s a signi ican shi compa ed o he 2023 edi-
ion. The mos no able change is a clea end away om
pa icipan s ypically coming om academic backg ounds
owa d hose who wo k in he c ea i e indus y, sugges ing
a g owing engagemen o p o essional a is s and c ea i es
wi h AI-d i en music composi ion, in a is ic and comme -
cial sec o s a he han academic esea ch se ing. In 2023,
58.8% o pa icipan s we e a ilia ed wi h academia, mak-
ing i he dominan ca ego y. Howe e , in 2024, academic
a ilia ion d opped o jus 19.57%, while he c ea i e in-
dus y su ged o 58.7%, making i he la ges ep esen ed
ca ego y in his yea ’s da ase .
The HAISP da ase o he 2024 edi ion o he AISC
showcases a di e se a ay o AI models and ools em-
ployed by pa icipa ing eams. Compa ed o he 2023 edi-
ion, which saw he usage o 74 di e en AI ools, he 2024
da ase e lec s an e en b oade spec um o AI applica-
ions wi h 82 di e en AI ools, excluding he o he ools
used by he disquali ied pa icipan s. These ools include
AI-powe ed music gene a ion models, oice cloning so -
wa e, AI-d i en mixing and mas e ing ools, AI-assis ed
composi ion pla o ms, and mo e. Some eams u ilized
publicly a ailable AI ools like Music y o Ki s.ai, while
o he s employed cus om-buil AI models ailo ed o hei
speci ic c ea i e needs, like Pu Da a.
A signi ican po ion (42%) o eams in 2024 con in-
ued o use Cha GPT and OpenAI’s GPT-based models o
ly ics, s uc u e, and c ea i e assis ance. Addi ionally, he
ise o S abili y AI models, such as S able-Audio-Open 1.0
and S able Di usion XL, sugges s a g owing eliance on
AI o bo h music gene a ion and isual con en c ea ion.
4. COMPARATIVE ANALYSIS
The HAISP da ase e eals ha while AI-assis ed song-
w i ing can enhance c ea i i y and e iciency, musicians
equen ly encoun e challenges ela ed o con ol, ans-
pa ency, and p ocess in eg a ion when wo king wi h AI
ools. Se e al ecu ing hemes, as p esen ed below,
eme ge om he da ase ha highligh he limi a ions o
cu en AI models.
4.1 Con ol
Twel e pa icipa ing eams exp essed us a ion o e he
lack o ine-g ained con ol o e AI-gene a ed ou pu s, es-
pecially when i comes o using he mo e popula and
easily accessible AI sys ems. Use s speci ically choose
sys ems ha allow o g ea e con ol and lexibili y o e
he ou pu s, highligh ing sys ems whose a o dances allow
hem o con ol “...mechanisms such as ex /audio p omp -
ing and loop gene a ion” (Team 63). Wi h sys ems ha
do no allow o such use modi ica ions, many eel ha
he esul s a e limi ed, and “...mos ly based on seed luck
and good p omp ing” (Team 28). One eam in pa icula
no ed ha hey chose o use an AI ool c ea ed by Ele en-
Labs [39] no only because hey el i allowed o “g ea e
con ol,” (Team 52) bu because o hei e hical s ance as
a eam, no ing ha Ele enLabs was mo e e hically ans-
pa en in he c ea ion o i s music da abase, using only li-
censed con en om Shu e s ock [40]. Decisions abou
ools a e no only based on con ol o e he p ocess o ou -
pu , bu also on he eam’s con ol o e how o accommo-
da e o apply hei e hical posi ions.
The 2024 HAISP da ase expansion ein o ces many o
he hemes om he 2023 da ase , pa icula ly in ega ds o
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
30
Figu e 1. Ba Cha compa ing he a ilia ion o AI Song Pa icipan s in 2023 and 2024, showcasing an inc ease in c ea i e
indus y a ilia ion and dec ease in academic a ilia ion.
main aining con ol o e hei c ea ion p ocess. In 2023,
eams equen ly encoun e ed igidi y and limi a ions when
wo king wi h AI ools, as exempli ied by 2023’s Team 13’s
expe ience wo king wi h Me a’s MusicGen. Ini ially “ as-
cina ed” by he ou comes ha came om his ool, hey
soon ealized he ool p oduced epe i i e esul s, and u -
he a emp s o e ine he ou pu esul ed in “...an unwel-
come su plus o noise, leading o a sense o limi a ion.”
This led hem o swi ch o Google’s Magen a so hey could
ha e mo e con ol o e he MIDI ou pu s. O he 2023
eams such as Team 16 also no ed ha ex - o-music mod-
els equen ly needed mo e speci ic and de ailed p omp ing
ha included in o ma ion on he key and empo in o de
o a oid he “...incohe en and some imes noisy ou pu s.”
O e all, eams om bo h yea s show a s onge p e e ence
o AI ools ha allow o eal- ime adjus men s, i e a i e
p omp ing, and clea e hical s ances so hey can ac i ely
and con inuously make choices ha gi e hem he con ol
hey desi e o e he c ea i e p ocess.
4.2 Applica ions o AI
In he 2024 da a, we no ed ha pa icipan s on a e age
used 2-3 imes mo e AI ools han he 2023 pa icipan s.
2024 AISC pa icipan s le e aged AI o melody and ha -
mony gene a ion, using models o p oduce ini ial musical
ideas ha we e la e e ined h ough human music p oduc-
ion s ages like mixing and mas e ing. Voice syn hesis was
ano he key applica ion, wi h AI ools ans o ming ocal
pe o mances o gene a ing syn he ic oices ha could be
adjus ed o i he song’s a is ic ision. Unlike 2023, he
majo i y o he ools used we e openly a ailable ools, and
no cus om-buil and ained models. Tools like S able Au-
dio, Ele enLabs, and Cha GPT 3.5/4.0 we e amongs he
mos commonly used ools in he 2024 da ase .
In con as , 2023’s eams o en employed AI ools i -
e a i ely, using hem o e ine composi ions and ly ics
h oughou he p ocess, which is desc ibed mo e as a ecu -
si e wo k low han s ep-by-s ep, wi h hei AI ool “...p o-
iding c ea i e sugges ions and helping us i e a e mo e
e icien ly” (Team 26). The i e a i e p ocess means ha
eams could lis en o hei AI-gene a ed song elemen s and
add on o hem wi h human elemen s as hei submission
de eloped, a he han ha ing he elemen be a gene a ed
piece ha canno be ec ea ed exac ly, e en wi h he same
p omp s. 2024’s Team 38 desc ibed hei us a ion wi h
his issue, w i ing ha "The p oblem wi h all o his, and
wha makes his wo k low so g anula , is he AI s a s o
d i o e ime, losing sigh o one aspec o he p omp in
a o o ano he ; ou pu s begin o di e in leng h, imb e,
language ( o some eason) bu mos impo an ly empo."
Addi ionally, many o he ools used by he 2023 pa ici-
pan s we e ei he buil by he eams o we e open sou ce
models ha we e ained by he eam. In 2023, la ge-scale
AI sys ems like Cha GPT we e mos ly used o c ea e sug-
ges ions o gene a e ideas o song elemen s such as he
melody, which was hen played and eco ded on eal in-
s umen s by pa icipan s, o ly ics, which we e hen sung
by a sepa a e AI ool o human eam membe .
4.3 Co-C ea ion s. Au oma ion
The da ase indica es a s ong p e e ence o collabo a i e
AI ools o e ully au oma ed music gene a o s. Many
a is s wan AI o unc ion as an assis i e ool a he han
an au onomous compose . Use s exp ess in e es in AI
sys ems ha espond dynamically o hei inpu s, a he
han gene a ing s a ic musical pieces ha equi e ex ensi e
manual e ision. As Team 28 no ed, i was no jus ha
u ilizing an AI ool ha made he wo k co-c ea i e, bu he
combina ion o hei own musical aining and skills and
he wo k o he AI ools which allowed o "... a seamless
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
31
blend [o ] human a is y and echnological inno a ion."
One issue wi h he 2024 da ase ha highligh s he dis-
inc ion be ween co-c ea ion and au oma ion was he 34
eams ha we e disquali ied due o hei use o Suno o
Udio, ex - o-music AI model. Because Suno and Udio
gene a e ully o med musical pieces - bypassing he c e-
a i e decision-making p ocess and c i ical disce nmen o
da a sou cing me hods ha he AISC expec s eams o en-
gage in - he use o he applica ion, and o he ex - o-
music models, was highly discou aged unless pa icipan s
could show hey used hei AI ou pu s in a c ea i e man-
ne . This highligh s he shi ing pe cep ions o owne ship
when i comes o AI-assis ed songw i ing, a leas in some
pa s o he MIR communi y. While use s o en see AI
as a ool o enhance hei c ea i e p ocess, compe i ions
like he AI Song Con es uphold a s ic e in e p e a ion
o human-AI collabo a ion, emphasizing ha AI should
be used as an assis i e echnology a he han a eplace-
men o human c ea i i y. The ques ion o owne ship be-
comes especially complex in his con ex : pa icipan s en-
gaging wi h co-c ea i e AI ools see hemsel es as he p i-
ma y au ho s, shaping and cu a ing AI-gene a ed elemen s,
whe eas models like Suno, which p oduce nea -comple e
composi ions, blu he bounda ies o au ho ship. The con-
es ’s decision o disquali y Suno use s e lec s an e ol ing
discou se a ound he ole o AI in music-making—-one
ha inc easingly p i ileges human c ea i e agency o e
AI-d i en au oma ion.
In compa ison, while some pa icipan s in he 2023
con es did use ex - o-music gene a i e models, hei use
was dis inc om he ully au oma ed app oach, and s ill
conside ed co-c ea i e by he con es judges. Fo exam-
ple, 2023’s Team 10 inco po a ed AudioLDM in o hei
p ocess o “gene a e backg ound and a mosphe e noises.”
Howe e , hey delibe a ely cu a ed, selec ed, and shaped
he gene a ed sounds, using hem as elemen s wi hin a
much b oade c ea i e p ocess. 2023’s Team 16 also u i-
lized a ex - o-music AI model o gene a e music segmen s
co esponding o s anzas om he poem “Visi o he ob-
se a o y” by Ha y Ma inson, bu also included an AI-
gene a ed oice and an imp o isa ional song om one o
he eam membe s. The dis inc ion be ween hese uses o
ex - o-music models and hose ha we e disquali ied om
2024 sugges s he e is an e ol ing discou se on AI’s ole
in music-making, whe e ools ha ely on i e a i e human-
AI in e ac ion a e seen as mo e legi ima e c ea i e aids,
while hose ha au oma e he p ocess en i ely unde mine
a is ic au ho ship. This shows ha he bounda ies o AI-
assis ed c ea i i y a e being con inuously nego ia ed as he
echnology e ol es, pa icula ly in con ex s whe e owne -
ship, au ho ship, and human in e en ion emain cen al o
c ea i e legi imacy.
5. DESIGN PRINCIPLES FOR MUSIC
GENERATIVE AI
5.1 In e ac i e and Adap i e AI Sys ems
De elope s should ocus on c ea ing AI ools ha allow o
i e a i e co-c ea ion, enabling musicians o p o ide eal-
ime eedback and e ine AI ou pu s [41, 42]. Ins ead o
gene a ing en i e songs in a single pass, AI models o mu-
sic c ea ion could unc ion mo e as collabo a o s, espond-
ing o use inpu and adap ing o changes in s yle, emo-
ion, and s uc u e [43]. This may also con ibu e o human
c ea o s’ inc eased sense o agency, con ol and owne ship
o e he inal c ea ion, and p omo e expe imen a ion o ex-
plo a ion o di e en c ea i e s yles.
5.2 Enhanced T anspa ency and Explainabili y
To build us and usabili y, AI ools should include ex-
plainabili y ea u es ha allow use s o see he easoning
behind AI-gene a ed sugges ions [44]. Visual in e aces
ha display cho d p og essions, ha monic ela ionships,
o AI decision pa hways could help musicians be e un-
de s and and manipula e AI-gene a ed con en . Ins ead o
“democ a izing” AI by ha ing he AI sys em gene a e con-
en beyond he use ’s c ea i e skill le el, his could allow
o he use o build domain knowledge o music and mu-
sic c ea ion [45]. As a esul , hey can engage mo e ully in
a deepe collabo a ion wi h he sys em, a he han simply
using i o ill in he gaps in a passi e manne .
5.3 S yle Cus omiza ion and Gen e Flexibili y
To o e come s ylis ic biases, de elope s should expand AI
aining da ase s o include a mo e di e se ange o musi-
cal adi ions, s yles, and composi ional echniques. This
could help ealize he goal o AI o b ing esh pe spec i es
and p omo e inno a ion in music c ea ion. Use s should
also ha e he abili y o ain AI models on hei own sound
lib a ies, allowing o g ea e pe sonaliza ion and gen e
lexibili y. While cus om buil o adap ed open sys ems do
ha e his abili y, gi ing use s o hese eady-made la ge
scale sys ems like S able Audio o Google Magen a he
abili y o cus omize hei model–e en in small ways–could
help make hem eel mo e in ol ed and connec ed wi h he
c ea ion p ocess and ou pu , a he han o cing hem in o
a limi ed se o musical s yles.
5.4 In eg a ion wi h Exis ing C ea i e Wo k lows
The compa ison highligh s he shi ing ole ha AI can
play wi hin he c ea i e p ocess, ac ing as ei he an assis-
i e ool o as a au oma ion o he c ea i e p ocess. In his
case, we sugges ha an impo an aspec o suppo ing he
i s case, allowing AI o unc ion as a co-c ea o a he
han some hing aking o e he p ocess, is o ensu e ha
AI ools a e adap able enough o wo k wi hin al eady es-
ablished music c ea ion wo k lows.
A ecu ing aspec o he AISC da ase is he use o mul-
iple ools o music c ea ion, sugges ing ha i is impo -
an o designe s o AI-based ools o ensu e ha hey a e
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
32
no dis up ing al eady es ablished wo k lows o indi idual
c ea o s, as well as pa e ns o c ea ion well u ilized by
communi ies. Fo example, AI ools ha wo k wi hin o he
so wa e such as digi al audio wo ks a ions (DAWs), MIDI
con olle s, and music no a ion so wa e may be mo e eas-
ily in eg a ed in o exis ing wo k lows. Fo example, he
abili y o expo po ions o loops o s ems om an AI ool
a e mo e use ul o he p oduc ion o music han ully gen-
e a ed pieces. P o iding plugin-based solu ions and lex-
ible API in eg a ions would make AI assis ance mo e ac-
cessible o musicians a all skill le els.
6. CONCLUSION AND FUTURE WORK
O e all, he expansion o he HAISP da ase o include
2024 da a allows o a mo e nuanced unde s anding o he
e ol ing ela ionship be ween musicians and he use o AI
o music c ea ion. Compa a i e analysis be ween 2023
and 2024 submissions showed a shi in AI Song Con es
submissions om p ima ily academia owa ds c ea i e in-
dus y, highligh ing inc easing in e es in he explo a ion
and adop ion o AI echnologies in p o essional music p o-
duc ion.
Submissions showed use o a b oad spec um o AI
ools, e lec ing a ying le els o adop ion o AI in he c e-
a ion p ocess. Pa icipan s showed a p e e ence o i e -
a i e eedback and cus omiza ion, using AI ools o sup-
plemen a eas hey el less expe ienced in o suppo hei
songw i ing p ocess while s ill main aining c ea i e con-
ol o e he inal composi ion. In gene al, pa icipan s’
use o AI is highly dependen on hei c ea i e philosophy
and exis ing echnical expe ise, and o some, hei e hical
s ances.
Fo u u e de elopmen o AI music c ea ion ools, de-
signe s should p io i ize adap i e, anspa en , and use -
d i en unc ionali y o be e align wi h a di e se ange
o needs. Fu u e AI ools should ocus on inc easing pe -
sonaliza ion and lexibili y, allowing musicians o in eg a e
AI-gene a ed ou pu s seamlessly in o hei exis ing wo k-
lows, and main ain anspa ency a ound e hical and e-
sponsible da a sou cing o ou pu s. To suppo his p o-
cess, u he esea ch could explo e he long- e m impac
o AI on music composi ion ac oss di e se musical gen-
es and cul u al con ex s. Addi ionally, he de elopmen
o open-sou ce and cus omizable AI ools could p o ide
musicians wi h mo e agency o e hei c ea i e p ocesses.
By con inuing o examine how AI shapes he a is ic land-
scape, we can os e inno a ion ha espec s a is ic in-
eg i y while le e aging he s eng hs o human-AI collab-
o a ion in music c ea ion.
As o he u u e o his p ojec , we plan o publish a
jou nal a icle ha includes a compa a i e analysis o he
da a om bo h yea s o u he explo e eme ging ends
in AI-suppo ed music c ea ion, as well as analyzing he
songs submi ed o he con es . In ligh o he shi in indus-
y pa icipa ion and he e ol ing quali y o submissions o
he AI Song Con es , we also aim o de elop mo e con-
c e e guidelines o u u e en ies ha encou age c ea i e
me hods and unique uses o AI. This will ensu e g ea e
consis ency in he expansion o he da a se and enhance
he o e all obus ness o he HAISP da ase .
7. ACKNOWLEDGMENTS
We hank [ edac ed o blind e iew] o hei con ibu ion
in o ganizing his e en .
8. REFERENCES
[1] P. Fo mosa, S. Bankins, R. Ma uliony e, and
O. Ghasemi, “Can cha gp be an au ho ? gene a i e ai
c ea i e w i ing assis ance and pe cep ions o au ho -
ship, c ea o ship, esponsibili y, and disclosu e,” AI &
SOCIETY, pp. 1–13, 2024.
[2] U. Messe , “Co-c ea ing a wi h gene a i e a i icial
in elligence: Implica ions o a wo ks and a is s,”
Compu e s in human beha io : a i icial humans,
ol. 2, no. 1, p. 100056, 2024.
[3] J. He, S. Houde, and J. D. Weisz, “Which con ibu ions
dese e c edi ? pe cep ions o a ibu ion in human-ai
co-c ea ion,” a Xi p ep in a Xi :2502.18357, 2025.
[4] P.-H. Chen and P.-L. P. Rau, “Using a i icial in elli-
gence in music c ea ion? a su ey based on copy-
igh consciousness,” in In e na ional Con e ence on
Human-Compu e In e ac ion. Sp inge , 2024, pp.
210–225.
[5] S. Li, “The impac o ai-d i en music p oduc ion so -
wa e on he economics o he music indus y,” In o -
ma ion De elopmen , p. 02666669241312170, 2025.
[6] N. Capo usso e al., “Gene a i e a i icial in elligence
and he eme gence o c ea i e displacemen anxie y,”
Resea ch in Psychology and Beha io , ol. 3, no. 1,
2023.
[7] K. E ickson, “Ai and wo k in he c ea i e indus ies:
digi al con inui y o discon inui y?” C ea i e Indus-
ies Jou nal, pp. 1–21, 2024.
[8] A. C oss, “Tool, collabo a o , o pa icipan : Ai
and a is ic agency,” B i ish Jou nal o Aes he ics, p.
ayae055, 2024.
[9] C. Anscomb, “Ai: a is ic collabo a o ?” AI & SOCI-
ETY, pp. 1–11, 2024.
[10] L. J. Mo is, R. Lege , M. Newman, J. A. Bu goyne,
R. G o es, N. Mangal, and J. H. Lee, “HAISP: A
Da ase o Human-AI Songw i ing P ocesses F om
he AI Song Con es ,” No . 2024, publishe : ISMIR.
[Online]. A ailable: h ps://zenodo.o g/doi/10.5281/
zenodo.14877357
[11] C.-Z. A. Huang, H. V. Koops, E. New on-Rex,
M. Dinculescu, and C. J. Cai, “Ai song con es :
Human-ai co-c ea ion in songw i ing,” a Xi p ep in
a Xi :2010.05388, 2020.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
33
[12] R. J. Gluck, “The columbia—p ince on elec onic mu-
sic cen e : Educa ing in e na ional compose s,” Com-
pu e Music Jou nal, ol. 31, no. 2, pp. 20–38, 2007.
[13] R. T. Dean and A. McLean, The Ox o d Handbook o
Algo i hmic Music. Ox o d Uni e si y P ess, 2018.
[14] E. Lemmon, “The impac o ins i u ional suppo on
a is ic esea ch and c ea ion: The columbia-p ince on
elec onic music cen e and he ca ma k-ii,” in In e na-
ional Compu e Music Con e ence P oceedings, 2019.
[15] E. R. Mi anda, “A he c oss oads o e olu iona y com-
pu a ion and music: Sel -p og amming syn hesize s,
swa m o ches as and he o igins o melody,” E o-
lu iona y Compu a ion, ol. 12, no. 2, pp. 137–158,
2004.
[16] R. L. De Man a as and J. L. A cos, “Ai and mu-
sic: F om composi ion o exp essi e pe o mance,” AI
magazine, ol. 23, no. 3, pp. 43–43, 2002.
[17] T. Sex on, “Musene ,” Music Re e ence Se ices Qua -
e ly, ol. 26, no. 3-4, pp. 151–153, 2023.
[18] A. Robe s, J. Engel, Y. Mann, J. Gillick, C. Kay-
acik, S. Nø ly, M. Dinculescu, C. Radebaugh,
C. Haw ho ne, and D. Eck, “Magen a s udio: Aug-
men ing c ea i i y wi h deep lea ning in able on li e,”
in P oceedings o he in e na ional wo kshop on musi-
cal me ac ea ion (mume), 2019, p. 11.
[19] R. Mi a and I. Zualke nan, “Music gene a ion using
deep lea ning and gene a i e ai: A sys ema ic e iew,”
IEEE Access, 2025.
[20] “MuseNe ,” Sep. 2022. [Online]. A ailable: h ps:
//openai.com/index/musene /
[21] R. Louie, A. Coenen, C. Z. Huang, M. Te y, and
C. J. Cai, “No ice-ai music co-c ea ion ia ai-s ee ing
ools o deep gene a i e models,” in P oceedings o
he 2020 CHI con e ence on human ac o s in compu -
ing sys ems, 2020, pp. 1–13.
[22] A. Meh a, S. Chauhan, and M. Choudhu y, “Missing
melodies: Ai music gene a ion and i s" nea ly" com-
ple e omission o he global sou h,” a Xi p ep in
a Xi :2412.04100, 2024.
[23] M. Bou eau, F. Mo eau, and P. Wiks öm, “Does dig-
i iza ion lead o he homogeniza ion o cul u al con-
en ?” Economic Inqui y, ol. 60, no. 1, pp. 427–453,
2022.
[24] H. Wang, J. Zou, M. Moze , A. Goyal, A. Lamb,
L. Zhang, W. J. Su, Z. Deng, M. Q. Xie, H. B own
e al., “Can ai be as c ea i e as humans?” a Xi
p ep in a Xi :2401.01623, 2024.
[25] D. C opley, “Is a i icial in elligence mo e c ea i e han
humans?: Cha gp and he di e gen associa ion ask,”
Lea ning Le e s, ol. 2, pp. 13–13, 2023.
[26] M. A. Boden, “C ea i i y and a i icial in elligence,”
A i icial in elligence, ol. 103, no. 1-2, pp. 347–356,
1998.
[27] R. Wings öm, J. Hau ala, and R. Lundman, “Rede in-
ing c ea i i y in he e a o ai? pe spec i es o compu e
scien is s and new media a is s,” C ea i i y Resea ch
Jou nal, ol. 36, no. 2, pp. 177–193, 2024.
[28] R. La ikka, J. Be gdahl, N. Sa ela, and A. Oksanen,
“Ai as an a is ? a wo-wa e su ey s udy on a i udes
owa d using a i icial in elligence in a ,” Poe ics, ol.
101, p. 101839, 2023.
[29] M. A. Runco, “Ai can only p oduce a i icial c ea i -
i y,” Jou nal o C ea i i y, ol. 33, no. 3, p. 100063,
2023.
[30] ——, “Upda ing he s anda d de ini ion o c ea i i y
o accoun o he a i icial c ea i i y o ai,” C ea i i y
Resea ch Jou nal, ol. 37, no. 1, pp. 1–5, 2025.
[31] R. K. Sawye and D. Hen iksen, Explaining c ea i i y:
The science o human inno a ion. Ox o d uni e si y
p ess, 2024.
[32] S. Canyakan, “The ole o ai in c ea i e p ocesses: e h-
ical and legal pe spec i es in he music indus y,” Jou -
nal o Music Theo y and T anscul u al Music S udies,
pp. 143–158, 2024.
[33] F. Mo eale, “Whe e does he buck s op? e hical and
poli ical issues wi h ai in music c ea ion,” T ansac ions
o he In e na ional Socie y o Music In o ma ion Re-
ie al, ol. 4, no. 1, 2021.
[34] H.-K. Lee, “Re hinking c ea i i y: c ea i e indus ies,
ai and e e yday c ea i i y,” Media, Cul u e & Socie y,
ol. 44, no. 3, pp. 601–612, 2022.
[35] H. H. Jiang, L. B own, J. Cheng, M. Khan, A. Gup a,
D. Wo kman, A. Hanna, J. Flowe s, and T. Geb u, “Ai
a and i s impac on a is s,” in P oceedings o he
2023 AAAI/ACM Con e ence on AI, E hics, and Soci-
e y, 2023, pp. 363–374.
[36] “Deed - A ibu ion-NonComme cial 4.0 In e na ional
- C ea i e Commons.” [Online]. A ailable: h ps:
//c ea i ecommons.o g/licenses/by-nc/4.0/deed.en
[37] C. E. Hill, S. Knox, B. J. Thompson, E. N. Williams,
S. A. Hess, and N. Ladany, “Consensual quali a i e e-
sea ch: An upda e,” Jou nal o Counseling Psychol-
ogy, ol. 52, no. 2, pp. 196–205, 2005.
[38] A. Song, “Ai song con es ,” 2024. [Online]. A ailable:
h ps://www.aisongcon es .com/join
[39] “F ee Tex o Speech & AI Voice Gene a o ,” Ma .
2025. [Online]. A ailable: h ps://ele enlabs.io
[40] Shu e s ock, “Why Ele enLabs Le e aged Shu -
e s ock’s Rich Da a o Build I s G oundb eaking
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
34
AI Sound E ec Gene a o ,” No . 2024. [On-
line]. A ailable: h ps://www.shu e s ock.com/blog/
ele enlabs-shu e s ock-ai-sound-e ec s-gene a o
[41] Y. Han, Z. Qiu, J. Cheng, and R. Lc, “When eams em-
b ace ai: human collabo a ion s a egies in gene a i e
p omp ing in a c ea i e design ask,” in P oceedings o
he 2024 CHI Con e ence on Human Fac o s in Com-
pu ing Sys ems, 2024, pp. 1–14.
[42] Y. Fu, M. Newman, L. Going, Q. Feng, and J. H. Lee,
“Explo ing he collabo a i e co-c ea ion p ocess wi h
ai: A case s udy in no ice music p oduc ion,” a Xi
p ep in a Xi :2501.15276, 2025.
[43] X. Hao, E. Demi , and D. Eye s, “Explo ing collab-
o a i e decision-making: A quasi-expe imen al s udy
o human and gene a i e ai in e ac ion,” Technology in
Socie y, ol. 78, p. 102662, 2024.
[44] N. B yan-Kinns, B. Zhang, S. Zhao, and B. Bana , “Ex-
plo ing a ia ional au o-encode a chi ec u es, con ig-
u a ions, and da ase s o gene a i e music explainable
ai,” Machine In elligence Resea ch, ol. 21, no. 1, pp.
29–45, 2024.
[45] J. B usila, M. Cloonan, and K. Rams ed , “Music, digi-
aliza ion, and democ acy,” Popula music and socie y,
ol. 45, no. 1, pp. 1–12, 2022.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
35
