Ma ijan, B anka
Wo king Pape
Th ough a glass, da kly: T anspa ency and mili a y AI
sys ems
CIGI Pape s, No. 315
P o ided in Coope a ion wi h:
Cen e o In e na ional Go e nance Inno a ion (CIGI), Wa e loo, On a io
Sugges ed Ci a ion: Ma ijan, B anka (2025) : Th ough a glass, da kly: T anspa ency and mili a y AI
sys ems, CIGI Pape s, No. 315, Cen e o In e na ional Go e nance Inno a ion (CIGI), Wa e loo
(On a io)
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/311807
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by/4.0/
CIGI Pape s No. 315 — Janua y 2025
Th ough a Glass, Da kly:
T anspa ency and Mili a y
AI Sys ems
B anka Ma ijan
CIGI Pape s No. 315 — Janua y 2025
Th ough a Glass, Da kly:
T anspa ency and Mili a y
AI Sys ems
B anka Ma ijan
Abou CIGI
The Cen e o In e na ional Go e nance Inno a ion (CIGI) is an independen ,
non-pa isan hink ank whose pee - e iewed esea ch and us ed analysis
in luence policy make s o inno a e. Ou global ne wo k o mul idisciplina y
esea che s and s a egic pa ne ships p o ide policy solu ions o he digi al
e a wi h one goal: o imp o e people’s li es e e ywhe e. Headqua e ed
in Wa e loo, Canada, CIGI has ecei ed suppo om he Go e nmen o
Canada, he Go e nmen o On a io and ounde Jim Balsillie.
À p opos du CIGI
Le Cen e pou l’inno a ion dans la gou e nance in e na ionale (CIGI) es un
g oupe de é lexion indépendan e non pa isan don les eche ches é aluées
pa des pai s e les analyses iables inci en les décideu s à inno e . G âce
à son éseau mondial de che cheu s plu idisciplinai es e de pa ena ia s
s a égiques, le CIGI o e des solu ions poli iques adap ées à l’è e numé ique
dans le seul bu d’amélio e la ie des gens du monde en ie . Le CIGI, don le
siège se ou e à Wa e loo, au Canada, béné icie du sou ien du gou e nemen
du Canada, du gou e nemen de l’On a io e de son onda eu , Jim Balsillie.
C edi s
Managing Di ec o and Gene al Counsel Aa on Shull
Di ec o , P og am Managemen Dianna English
P og am Manage and Resea ch Associa e Kailee Hil
Senio Publica ions Edi o Jenni e Goyde
Publica ions Edi o Ch is ine Robe son
G aphic Designe Sepideh Shomali
Copy igh © 2025 by he Cen e o In e na ional Go e nance Inno a ion
The opinions exp essed in his publica ion a e hose o he au ho and do no
necessa ily e lec he iews o he Cen e o In e na ional Go e nance Inno a ion
o i s Boa d o Di ec o s.
Fo publica ions enqui ies, please con ac [email p o ec ed].
The ex o his wo k is licensed unde CC BY 4.0. To iew a copy o his licence,
isi h p://c ea i ecommons.o g/licenses/by/4.0/.
Fo euse o dis ibu ion, please include his copy igh no ice. This wo k may
con ain con en (including bu no limi ed o g aphics, cha s and pho og aphs)
used o ep oduced unde licence o wi h pe mission om hi d pa ies.
Pe mission o ep oduce his con en mus be ob ained om hi d pa ies di ec ly.
Cen e o In e na ional Go e nance Inno a ion and CIGI a e egis e ed
adema ks.
67 E b S ee Wes
Wa e loo, ON, Canada N2L 6C2
www.cigionline.o g
Table o Con en s
i Abou he Au ho
1 Execu i e Summa y
1 In oduc ion
3 T anspa ency and Technical Unde s andings
4 T anspa ency and In e na ional Secu i y Go e nance
5 Mili a y AI Go e nance Discussions
6 T anspa en AI o he De ence Con ex
6 Explainabili y, Unde s andabili y
and P edic abili y
7 T anspa ency and Ope a ional Challenges
8 Towa d a Comp ehensi e T anspa ency App oach o Mili a y AI
10 Ex e nal T anspa ency and Go e nance F amewo ks
11 Conclusion
12 Wo ks Ci ed
i CIGI Pape s No. 315 — Janua y 2025 • B anka Ma ijan
Abou he Au ho
B anka Ma ijan is a CIGI senio ellow and a
senio esea che a P ojec Ploughsha es. She
is a lec u e in he Mas e o Global A ai s
p og am a he Munk School o Global A ai s
and Public Policy a he Uni e si y o To on o.
A Ploughsha es, B anka leads esea ch on
he mili a y and secu i y implica ions o
eme ging echnologies. He wo k examines
conce ns ega ding he de elopmen o
au onomous weapons sys ems and he
impac o a i icial in elligence and obo ics
on secu i y p o ision. He esea ch in e es s
include ends in wa a e, ci ilian p o ec ion,
use o d ones and ci il-mili a y ela ions.
She holds a Ph.D. om he Balsillie School o
In e na ional A ai s wi h a specializa ion in
con lic and secu i y. She has conduc ed esea ch
on pos -con lic socie ies and published academic
a icles and epo s on he impac s o con lic
on ci ilians and di e se issues o secu i y
go e nance, including secu i y sec o e o m.
B anka closely ollows Uni ed Na ions disa mamen
e o s and a ends in e na ional and na ional
consul a ions and con e ences. She is a boa d
membe o he Peace and Con lic S udies
Associa ion o Canada and a esea ch ellow
a he Kind ed C edi Union Cen e o Peace
Ad ancemen a he Uni e si y o Wa e loo.
1Th ough a Glass, Da kly: T anspa ency and Mili a y AI Sys ems
Execu i e Summa y
In e na ional go e nance discussions on mili a y
a i icial in elligence (AI) sys ems o en emphasize
he need o anspa ency. Howe e , anspa ency
is a complex and mul i- ace ed concep , unde s ood
in a ious ways wi hin in e na ional deba es
and li e a u e on he esponsible use o AI. I
encompasses dimensions such as explainabili y,
in e p e abili y, unde s andabili y, p edic abili y
and eliabili y. The deg ee o which hese aspec s
a e e lec ed in s a e app oaches o ensu ing
anspa en and accoun able sys ems emains
unclea and equi es u he in es iga ion.
Addi ionally, achie ing anspa ency in mili a y
AI applica ions p esen s se e al challenges. Fi s ,
he inhe en opaci y o he echnology can make i
di icul o ace and unde s and decision-making
p ocesses. Second, mili a y ins i u ions a e mo e
likely o adop olun a y anspa ency measu es
ha ocus on ensu ing ope a o s ha e a gene al
unde s anding o sys em unc ionali y, wi hou
ully add essing he nuances o accoun abili y.
Fu he mo e, dispa i ies in echnological capabili ies
among s a es sugges une en es ing and aining
s anda ds, complica ing he e alua ion o human
decision making and accoun abili y. Las ly, gi en
he sensi i i y o na ional de ence and in e na ional
secu i y, mili a y AI sys ems a e expec ed o emain
highly classi ied, making ex e nal e alua ion
di icul . This pape p oposes pa hways o o e come
hese challenges and ou lines a amewo k o
comp ehensi e anspa ency, which is essen ial
o he esponsible use o AI in mili a y con ex s.
In oduc ion
In in e na ional discussions on esponsible mili a y
use o AI, anspa ency is equen ly emphasized.
T anspa ency is also a cen al conce n ac oss AI
e hical p inciples in ci ilian con ex s (Jobin, Ienca
and Vayena 2019). Howe e , he concep ualiza ion
o anspa ency a ies conside ably. Fo some
go e nmen s, anspa ency en ails some disclosu e
o in o ma ion ega ding he es ing, e alua ion
and unc ioning o a ious sys ems by s a es. Fo
o he s, i means ha mili a y AI sys ems mus be
su icien ly anspa en o hei own mili a ies
and ensu e ha commande s unde s and hei
ope a ions and can in e ene when hese sys ems
p oduce e o s o unp edic able ou pu s. In his
way, he unde s anding o anspa ency is gene ally
one o “ he unde s andabili y and p edic abili y
o sys ems” (Endsley, Bol e and Jones 2003, 146;
Na ional Academies o Sciences, Enginee ing,
and Medicine 2022). Howe e , he challenge
emains ha hese a ying in e p e a ions o
anspa ency will become e en mo e signi ican
as s a es begin ope a ionalizing esponsible AI
p inciples. These p inciples will be especially
impo an o ensu ing he esponsible use o AI
and au onomous sys ems by mili a y o ces.
Al eady in p ac ice in con empo a y con lic zones
such as Uk aine and Gaza, commi men s o ha ing
mili a y commande s unde s and AI sys ems
a e being challenged due o he na u e o he
echnology, he use o o - he-shel echnologies and
he lack o clea guidelines ega ding he ex en o
which such unde s anding is equi ed. The e is also
a b oade lack o disclosu e abou he ypes and
sophis ica ion o AI-enabled sys ems being used and
how hey unc ion. No ably, he AI a ge gene a ion
and decision suppo sys ems used by he Is ael
De ense Fo ces (IDF) in Gaza ha e aised conce ns as
in es iga i e epo s publicized hei use, leading o
mo e ques ions abou hei unc ion (Ab aham 2024;
Da ies, McKe nan and Sabbagh 2023). Howe e , li le
in o ma ion has been p o ided by Is ael on how he
sys ems unc ion and he coun y has a gued ha
i is no using AI sys ems o au onomously selec
a ge s wi hou human in ol emen (Va ella and
Acheson 2024, 5). These assu ances ha e no been
seen as su icien by hose ala med a he epo s
ega ding Is aeli sys ems. T anspa ency ega ding
AI and au onomous sys ems in he mili a y domain
also in ol es some abili y o access in o ma ion on
sys ems, ideally o ha e hese sys ems be e alua ed
2CIGI Pape s No. 315 — Janua y 2025 • B anka Ma ijan
o audi ed, p e e ably by a epu able hi d pa y.
Such an ex ensi e e alua ion and audi ing, while
likely o be done in e nally, is unlikely possible
o be pe o med ex e nally. As such, in o ma ion
sha ing and con idence-building measu es a he
global le el will need o be c ea i ely de eloped.
Se e al ques ions a ise when seeking o es ablish
a deepe unde s anding o anspa ency ha
sa is ies bo h in e na ional go e nance bodies and
echnical and ope a ional equi emen s. Does he
mili a y commande need o unde s and how each
node o he AI sys em is connec ed? Would a deep
enough unde s anding be possible o equi ed,
and o which uses? Wha would be a su icien
le el o unde s anding by he human ope a o o
wa igh e o ensu e hei clea accoun abili y o
ac ions aided by o ca ied ou by an AI sys em?
Addi ionally, wha in o ma ion needs o be sha ed
among a ious go e nmen s o ensu e con idence in
he esponsible use o AI and au onomous sys ems?
These ques ions a e conside ably mo e ele an as
mili a ies a e inc easingly using AI sys ems ac oss a
a ie y o unc ions, including ec ui men , aining,
logis ics, equipmen main enance, su eillance and
a ge ing (G and-Clémen 2023). The di e en uses
will ha e a ying equi emen s o anspa ency ha
se e di e en unc ions and sa is y e hical and legal
equi emen s a a ious le els o go e nance. Fo
some uses, like hose desc ibed as “back-end” o ice
unc ions such as ec ui men , he equi emen s
will p ima ily ocus on ensu ing ai ness and
p i acy, as well as mee ing a ious domes ic laws
on employing indi iduals (Taddeo e al. 2019). On
he o he end o he spec um, and o mos conce n
o his pape , a e high- isk applica ions, such as
he use o AI sys ems in decision suppo ela ed
o deploymen o o ce o in weapon sys ems, wi h
a ying deg ees o au onomy. The equi emen s
will be mo e s ingen , needing o mee in e nal
and na ional s anda ds as well as in e na ional
legal equi emen s and go e nance mechanisms.
The la e issue, while pa icula ly c i ical o
in e na ional secu i y, emains he mos challenging
o add ess due o inhe en secu i y conside a ions.
T anspa ency in he mili a y use o AI and
au onomy a he global le el aces se e al key
obs acles. Fi s , he inhe en complexi y o he
echnology, especially as sys ems become mo e
ad anced, lea n and e ol e, makes ensu ing hei
unde s andabili y challenging in p ac ice. The e is an
ac i e deba e on he ex en o which sys ems need
o be explainable as well as in e p e able by humans
and wha deg ee o unde s anding is equi ed by
hose deploying sys ems. Addi ionally, he dual-use
na u e o AI and he use o comme cial o - he-shel
echnologies and ools, as u ilized in Uk aine, may
in oduce sys ems ha ha e no been adequa ely
es ed o de ence con ex s. Second, while mili a ies
a e mo e inclined o commi o anspa ency
measu es ha ensu e ope a o s unde s and he
sys ems, b oade anspa ency o allowing ex e nal
e alua ion o hese sys ems emains signi ican ly
mo e challenging. Thi d, and ela edly, mili a y AI
sys ems a e o en closely gua ded due o na ional
secu i y conce ns. This con iden iali y can hinde
he willingness o s a es o sha e in o ma ion
ega ding he capabili ies o a ious sys ems.
This endency is pa icula ly ue wi h mo e
ad e sa ial na ions, as anspa ency ega ding
he unc ioning o mili a y AI sys ems is unlikely
o be sha ed due o ea s o exposing con iden ial
echnologies ha may p o ide a echnical edge o
o he s a e ac o s. T anspa ency, he e o e, o en
collides wi h na ional secu i y (E zioni 2018).
This pape explo es he easibili y o achie ing
anspa ency in mili a y AI sys ems, iden i ies
he associa ed challenges and p oposes pa hways
o de elop e ec i e anspa ency mechanisms.
I begins by examining di e ing de ini ions o
anspa ency, om echnical unde s andings o
in e na ional secu i y go e nance. I hen discusses
how hese a ious app oaches ha e eme ged
in he discou se on mili a y AI go e nance.
D awing on hese di e se pe spec i es, he
pape p oposes elemen s o a comp ehensi e
anspa ency app oach o conside o in e na ional
go e nance mechanisms. Ul ima ely, anspa ency
mechanisms in he mos conce ning mili a y AI
applica ions, such as decision making ela ed o
he use o o ce, will also equi e a laye ed se
o go e nance commi men s and con idence-
building measu es. These should include clea
legally binding commi men s, olun a y measu es
and exchanges o in o ma ion. Finally, many
mili a y applica ions o AI a e likely o emain
sh ouded in sec ecy. Howe e , achie ing a
sa is ac o y le el o anslucency in applica ions
wi h he mos signi ican impac on global
secu i y will g ea ly enhance global s abili y.
9Th ough a Glass, Da kly: T anspa ency and Mili a y AI Sys ems
such, igo ous es ing o ci ilian AI echnologies
be o e hey a e deployed in mili a y se ings will
be c ucial. A gene ally acknowledged aspec o AI
sys ems is hei b i leness; while he sys ems can
pe o m well in na ow applica ions, hey s uggle
wi h mo e gene al applica ions (Maye 2023).
Thus, ci ilian sys ems migh se e as he basis
o mili a y applica ions, bu he ansi ion om
ci ilian o mili a y use is augh wi h po en ial
pi alls. Wi hou ho ough es ing, hese sys ems
may exhibi un o eseen ulne abili ies o biases
ha could ha e ca as ophic consequences in a
comba en i onmen . These consequences include
he a ge ing o ci ilians and ci ilian in as uc u e.
In ci ilian con ex s, AI sys ems a e subjec o
ex ensi e es ing and alida ion o ensu e ha
hey pe o m as in ended. This p ocess in ol es
e alua ing he sys em’s accu acy, eliabili y
and esilience unde a ious condi ions. When
hese sys ems a e adap ed o mili a y use,
hey mus unde go an e en mo e s ingen
es ing egime. The s akes a e highe in mili a y
applica ions, whe e he ma gin o e o is minimal
and he po en ial o ha m is signi ican .
P ope es ing should encompass no only
echnical pe o mance bu also e hical
conside a ions. Fo ins ance, he biases inhe en
in many AI sys ems can be ampli ied in a mili a y
con ex , leading o unjus ou comes. The e o e,
es ing p o ocols mus include assessmen s
o ai ness and bias mi iga ion s a egies.
Comp ehensi e T aining o Use s
T anspa ency in mili a y AI sys ems also
depends on he comp ehensi e aining o
indi iduals who ope a e hese echnologies.
I is no enough o ope a o s o me ely
unde s and how o use he sys ems; hey mus
also g asp he unde lying p inciples guiding
he AI’s decision-making p ocesses (Lyons e
al. 2017). The ope a o s hen need o sha e hei
knowledge ega ding how he sys ems unc ion
wi h mili a y commande s. This knowledge is
c ucial o ensu ing ha human commande s
emain in con ol, can make in o med decisions
when in e acing wi h AI and ha e a s ong
deg ee o unde s andabili y o he AI sys ems.
T aining p og ams should be designed o
p o ide ope a o s and commande s wi h a deep
unde s anding o he AI sys ems ha hey a e
using. This aining should include no only he
sys ems’ echnical aspec s bu also he ways
in which hey may be designed o in luence o
“nudge” decision making (Milla 2015). Fo ins ance,
AI sys ems o en p esen da a in ways ha can
sub ly guide ope a o s owa d speci ic conclusions
o ac ions. Recognizing hese nudges is essen ial
o main aining human o e sigh and p e en ing
he o e eliance on AI ecommenda ions.
Mo eo e , aining should emphasize he
impo ance o c i ical hinking and e hical
conside a ions. Ope a o s mus be able o
ques ion and e alua e he AI’s ou pu s, a he
han accep ing hem a ace alue. As such,
mili a y o ganiza ions — whe e he adi ional
hie a chical s uc u e can some imes discou age
ques ioning o au oma ed sys ems — mus unde go
a cul u al shi . By os e ing a cul u e o c i ical
engagemen , he mili a y can ensu e ha AI
sys ems a e used esponsibly and anspa en ly.
Con inuous Assignmen
o Accoun abili y
As mili a y AI sys ems e ol e and a e upda ed, he
assignmen o accoun abili y mus be a con inuous
and dynamic p ocess. One o he signi ican
challenges in AI go e nance is ensu ing ha
accoun abili y is main ained h oughou he li e
cycle o he sys em, including no only he ini ial
deploymen , bu also subsequen upda es and
modi ica ions.
When AI sys ems a e upda ed, new ea u es o
adjus men s can in oduce un o eseen isks o
al e he sys em’s beha iou in ways ha a e no
immedia ely appa en . The e o e, i is essen ial o
ha e mechanisms in place o e-e alua ing he
sys em’s pe o mance and e hical implica ions
a e each upda e. This p ocess should in ol e
a di e se g oup o s akeholde s, including
echnical expe s, e hicis s and end-use s.
Mo eo e , clea lines o accoun abili y mus be
es ablished o e e y s age o he AI sys em’s li e
cycle. This accoun abili y includes iden i ying who
is esponsible o designing, es ing, deploying
and main aining he sys em. In cases whe e he AI
sys em makes a c i ical e o o exhibi s unin ended
beha iou , he e should be a anspa en p ocess
o de e mining esponsibili y and implemen ing
co ec i e measu es. As no ed ea lie , sys ems
need o ha e a deg ee o explainabili y o ensu e
ha hose looking a ex pos accoun abili y can
show whe he a sys em mal unc ioned o whe he
10 CIGI Pape s No. 315 — Janua y 2025 • B anka Ma ijan
he human decision make ac ed in a manne
ha led o a pa icula c ime being commi ed.
The assignmen o accoun abili y should also
ex end o he p ocu emen p ocess. Mili a y
o ganiza ions mus ensu e ha con ac o s
and endo s adhe e o s ingen e hical and
anspa ency s anda ds. The p ocess should
include equi ing de ailed documen a ion o
he AI sys em’s de elopmen p ocess, es ing
p o ocols and any known limi a ions o biases.
By holding endo s accoun able, he mili a y
can enhance he o e all anspa ency and
eliabili y o i s AI sys ems. Mili a ies will also
need o de elop p ocedu es o e alua e he sel -
ce i ica ion ha endo s a e likely o p esen
and p opose, as hese can be manipula ed
o “gamed” by he de elope s (Say 2024).
Ex e nal T anspa ency
and Go e nance
F amewo ks
In e nal anspa ency, while necessa y, is no
su icien o add ess he legal and e hical challenges
posed by he deploymen o AI-enabled sys ems
by mili a ies. To ensu e esponsible use, global
s anda ds and oppo uni ies o con idence-building
measu es a e also needed. Ex e nal anspa ency
can in ol e a di e se g oup o s akeholde s.
When s a es disclose in o ma ion ega ding hei
policies and he echnical aspec s o AI sys ems,
he global scien i ic communi y can weigh in
wi h insigh s om o he sa e y-c i ical ields. This
collabo a i e app oach allows o a comp ehensi e
e alua ion o he issues, ensu ing ha mul iple
pe spec i es a e conside ed in add essing he
challenges o AI in he mili a y domain.
Regula ion mus guide anspa ency mechanisms,
wi h s a es deciding which sys ems and p ocesses
a e accep able o be enabled o ca ied ou by AI
echnologies. A he in e na ional le el, a legally
binding ins umen p o iding clea ed lines will
be necessa y. Such a p ocess could happen wi hin
he amewo k o he Uni ed Na ions, which
has o a ep esen a i e o he wides numbe
o coun ies. Addi ionally, he e a e es ablished
o a whe e s a es can exchange in o ma ion
and p o ide insigh s in o hei app oaches and
policies on he applica ions o AI in he mili a y
domain. The CCW is one such o um likely o
con inue dialogue on au onomous weapons.
Beyond he CCW, he US-led poli ical decla a ion
and he wo king g oup on accoun abili y and
anspa ency can con ibu e o de eloping bes
p ac ices and empla es o s a es o use in acking
how unde s andabili y and p edic abili y can be
main ained h ough a ious s ages o de elopmen .
Summi s and in e na ional mee ings can also
help es ablish no ms on he ypes o in o ma ion
ha can be sha ed. Fo example, he Summi on
Responsible AI in he Mili a y Domain (REAIM)
held in The Hague, Ne he lands in 2023 and in
Seoul, Sou h Ko ea in 2024, b ings oge he a
b oade g oup o s a es, including China and
mul iple s akeholde s. REAIM is hus well
posi ioned o p o ide a space o dialogue wi hou
p essu e o speci ic egula ions o mo e conc e e
commi men s, and i ea u es a mul i-s akeholde
en i onmen (Cse na oni 2024). Mo e exchanges
be ween go e nmen s and in mili a y- o-mili a y
dialogues can be help ul in unde s anding which
go e nmen s a e echnically capable, bu ha e
ecei ed less a en ion ega ding hei posi ions
in wha hey iew as he esponsible use o
mili a y AI, o example India. Al hough India has
spoken abou he esponsible use o AI, i also
signalled i s wish o deploy hese echnologies
mo e widely in CCW discussions. Indeed, g ea e
cla i y on mili a y AI policies can con ibu e o
knowledge and us building among s a es.
Con idence-building measu es a e he e o e c i ical
o ex e nal anspa ency. Michael C. Ho owi z,
Lau en Kahn and Casey Mahoney (2020) examine
he way in which con idence-building measu es
ha e his o ically con ibu ed o in e na ional
secu i y and a ms con ol ag eemen s. They
poin ou ha while old app oaches a e no
exac empla es o mili a y AI, hese app oaches
o e conside a ions necessa y o in e na ional
coope a ion. They no e ha s a es can exchange
a deg ee o echnical in o ma ion on a ious
sys ems and hei policies ha will os e a g ea e
willingness o engage in dialogue, e en among
mo e ad e sa ial s a es. Sha ing o in o ma ion
abou e hical conside a ions in a ious o a can
also hen p o ide insigh s on whe e a ious
s a es s and. Thus, con idence-building measu es
a e going o become mo e impo an as ex e nal
e alua ion o ce i ica ion by hi d-pa y
11Th ough a Glass, Da kly: T anspa ency and Mili a y AI Sys ems
ins i u ions does no appea o be possible in he
nea e m. Ioana Puscas (2022) poin s ou ha
isk-based app oaches can p o ide some ocus on
p io i izing con idence-building e o s in mili a y
applica ions o AI. Puscas p o ides he example
o s a es ag eeing o cons ain s on deploying AI
in a eas whe e he isk is excep ionally high, such
as nuclea weapons. Such an ag eemen could
pe haps lead o a isk-based go e nance amewo k
o au onomous weapons and mili a y AI, which
could ensu e ha sys ems posing unaccep able
isks a e ei he p ohibi ed o subjec ed o s ic
es ic ions on hei use (Ma ijan 2021).
Conclusion
The need o anspa ency in mili a y AI
sys ems ex ends beyond immedia e ope a ional
conce ns. I is also a c ucial ac o in main aining
public us and in e na ional s abili y. As AI
echnologies become e e mo e in eg a ed in o
mili a y ope a ions, he po en ial o misuse o
unin ended consequences inc eases. Due o he
hype su ounding he echnology and geopoli ical
eali ies, he e migh be p essu e o deploy
echnologies ha a e no app op ia ely es ed
o unde s ood. T anspa en p ac ices can help
mi iga e hese isks by ensu ing ha AI sys ems
a e used e hically and esponsibly. T anspa ency
os e s con idence ha hese echnologies a e being
used in acco dance wi h e hical s anda ds and ha
he e a e obus mechanisms in place o add ess
any issues ha a ise. This us is pa icula ly
impo an in democ a ic socie ies whe e public
opinion can in luence de ence policies.
On he in e na ional s age, anspa ency in
mili a y AI sys ems can help p e en escala ion
and educe he isk o con lic . When s a es a e
open abou hei AI capabili ies and he measu es
ha hey ha e in place o ensu e e hical use, i
can build mu ual us and acili a e coope a ion.
Con e sely, a lack o anspa ency can lead o
suspicion and a ms aces, as coun ies may
eel compelled o de elop hei AI capabili ies
in sec e o main ain a s a egic ad an age.
Finally, anspa ency e o s can only supplemen ,
no eplace, egula ion. The use o AI-enabled
sys ems in he deploymen o o ce is oo
consequen ial o global secu i y o o go legally
binding ins umen s and o he measu es a he
in e na ional le el. Technical solu ions ha e hei
limi a ions, and AI-enabled sys ems a e likely o
encoun e signi ican ope a ional issues ou side
o labo a o y se ings. S a es mus conside he
b oade impac s o deploying hese sys ems. As
e i ed Gene al Ma k Milley, o me chai man
o he US Join Chie s o S a , poin s ou , “The
idea ha wa is an isep ic and he e a e wonde
weapons ou he e, ha we can somehow make
i painless… o hink ha echnology’s going o
esol e he ho o s o wa . I ’s no ” (quo ed in
F eedbe g 2024). Hence, anspa ency e o s a e
one c i ical pa o he b oade go e nance and
egula o y amewo k ha needs o be de eloped
o mili a y applica ions o AI and au onomy.
Wi hou a comp ehensi e anspa ency amewo k,
he po en ial o deploymen o echnologies
ha a e no sui able o a ious en i onmen s o
ha inc ease isk o con lic escala ion only ises.
Con empo a y con lic s al eady se e as a wa ning
sign ha assump ions abou wha is accep able
unde exis ing no ms and in e na ional laws a e
being e oded.
12 CIGI Pape s No. 315 — Janua y 2025 • B anka Ma ijan
Wo ks Ci ed
Ab aham, Yu al. 2024. “‘La ende ’: The AI machine di ec ing
Is ael’s bombing sp ee in Gaza.” +972Magazine, Ap il 3.
www.972mag.com/la ende -ai-is aeli-a my-gaza/.
Acheson, Ray and Lau a Va ella. 2024. “Topic 3: Risk Mi iga ion
and Con idence Building.” CCW Repo 12 (1): 23–30.
h ps:// eachingc i icalwill.o g/images/documen s/
Disa mamen - o a/ccw/2024/gge/ epo s/CCWR12.1.pd .
Ananny, Mike and Ka e C aw o d. 2018. “Seeing wi hou knowing:
Limi a ions o he anspa ency ideal and i s applica ion o
algo i hmic accoun abili y.” New Media and Socie y 20 (3):
973–89. h ps://doi.o g/10.1177/1461444816676645.
Babic, Bo is, Sa a Ge ke, Theodo os E geniou and I. Glenn Cohen.
2021. “Bewa e explana ions om AI in heal h ca e.”
Science 373 (6552): 284–6. www.science.o g/
doi/10.1126/science.abg1834.
Ball, Ca olyn. 2009. “Wha Is T anspa ency?” Public In eg i y 11 (4):
293–308. h ps://doi.o g/10.2753/PIN1099-9922110400.
Cse na oni, Raluca. 2024. “Go e ning Mili a y AI Amid
a Geopoli ical Mine ield.” Ca negie Endowmen
Eu ope. July 17. h ps://ca negieendowmen .o g/
esea ch/2024/07/go e ning-mili a y-ai-amid-a-
geopoli ical-mine ield?lang=en¢e =eu ope.
Cummings, Maisy. 2024. “How o ensu e esponsible use o
AI in mili a y decision-making – d awing lessons om
he au onomous weapons sys ems (AWS) deba e.”
REAIM Summi p esen a ion. Seoul, Sou h Ko ea.
Da ies, Ha y, Be han McKe nan and Dan Sabbagh. 2023.
“‘The Gospel’: how Is ael uses AI o selec bombing a ge s
in Gaza.” The Gua dian, Decembe 1.
www. hegua dian.com/wo ld/2023/dec/01/ he-
gospel-how-is ael-uses-ai- o-selec -bombing- a ge s.
Endsley, R. Mica. 2023. “Suppo ing Human-AI Teams:
T anspa ency, explainabili y, and si ua ion awa eness.”
Compu e s in Human Beha io 140, 107574.
h ps://doi.o g/10.1016/j.chb.2022.107574.
Endsley, R. Mica, Be y Bol e, and Deb a G. Jones. 2003.
Designing o Si ua ion Awa eness: An App oach o Human-
Cen e ed Design. London, UK: Taylo and F ancis.
E zioni, Ami ai. 2018. “The Limi s o T anspa ency.” In T anspa ency,
Socie y and Subjec i i y, edi ed by Emmanual Alloa. and Die e
Thomä, 179–201. Cham, Swi ze land: Palg a e Macmillan.
Finel, Be na d I. and K is in M. Lo d. 2000. Powe and Con lic in he
Age o T anspa ency. New Yo k, NY: Palg a e Macmillan.
F anzoni, Valen ina. 2023. “F om Black Box o Glass Box:
Ad ancing T anspa ency in A i icial In elligence Sys ems
o E hical and T us wo hy AI.” In Compu a ional
Science and I s Applica ions– ICCSA 2023 Wo kshops,
edi ed by Os aldo Ge asi, Beniamino Mu gan e,
Anna Ma ia A. C. Rocha, Chia a Ga au, F ancesco
Sco za, Yeliz Ka aca and Ca melo M. To e. Lec u e
No es in Compu e Science se ies, 14107: 118–30.
F eedbe g, Sydney, J . 2024. “Can ech educe ci ilian dea hs
in con lic ? Ma k Milley isn’ so su e.” B eaking De ense,
May 8. h ps://b eakingde ense.com/2024/05/ma k-
milley-ci ilian-dea h-ai- echnology-alex-ka p-palan i /.
G and, Camille. 2003. “Nuclea Weapon S a es and he
Secu i y Dilemma.” In T anspa ency in Nuclea Wa heads
and Ma e ials, edi ed by Nicholas Za impas, 32–49.
Ox o d, UK: Ox o d Uni e si y P ess and SIPRI.
G and-Clémen , Sa ah. 2023. A i icial In elligence Beyond
Weapons: Applica ion and Impac o AI in he Mili a y
Domain. Gene a, Swi ze land: Uni ed Na ions Ins i u e o
Disa mamen Resea ch. h ps://unidi .o g/wp-con en /
uploads/2023/10/UNIDIR_AI_Beyond_Weapons_
Applica ion_Impac _AI_in_ he_Mili a y_Domain.pd .
Hagelin, Bjö n, Ma k B omley, John Ha , Shannon N. Kile, Zdzislaw
Lachowski, Wuyi Omi oogun, Ca alina Pe domo, Eamon Su y
and Siemon T. Wezeman. 2006. “T anspa ency in he a ms
li e cycle.” In SIPRI Yea book 2006: A mamen s, Disa mamen
and In e na ional Secu i y, 245–67. S ockholm, Sweden: SIPRI.
www.sip i.o g/si es/de aul / iles/YB06%20245%2006.pd .
Hansen, Hans K ause, La s Thøge Ch is ensen and Mikkel
Fly e bom. 2015. “In oduc ion: Logics o anspa ency in
la e mode ni y: Pa adoxes, media ion and go e nance.”
Eu opean Jou nal o Social Theo y 18 (2): 117–31.
h ps://doi.o g/10.1177/136843101455524.
Hassija, Vikas, Vinay Chamola, A mesh Mahapa a, Abhinandan
Singal, Di yansh Goel, Kaizhu Huang, Simone Sca dapane,
Ind o Spinelli, Mu i Mahmud and Ami Hussain. 2024.
“In e p e ing Black-Box Models: A Re iew on Explainable
A i icial In elligence.” Cogni i e Compu a ion 16: 45–74.
h ps://doi.o g/10.1007/s12559-023-10179-8.
Holland Michel, A hu . 2020. The Black Box, Unlocked: P edic abili y
and Unde s andabili y in Mili a y AI. Gene a, Swi ze land:
Uni ed Na ions Ins i u e o Disa mamen Resea ch.
h ps://unidi .o g/publica ion/ he-black-box-unlocked/.
13Th ough a Glass, Da kly: T anspa ency and Mili a y AI Sys ems
Hol om, Paul, Anna Mensah and Ruben Nicolin. 2022. “The
Case o S eng hening T anspa ency in Con en ional
A ms T ans e s.” A ms Con ol Today. No embe .
www.a mscon ol.o g/ac /2022-11/ ea u es/case-
s eng hening- anspa ency-con en ional-a ms- ans e s.
Ho owi z, Michael C., Lau en Kahn and Casey Mahoney. 2020.
“The Fu u e o Mili a y Applica ions o A i icial In elligence:
A Role o Con idence-Building Measu es?” O bis 64 (4):
528–43. h ps://doi.o g/10.1016/j.o bis.2020.08.003.
Ho owi z, Michael and Paul Scha e. 2021. AI and In e na ional
S abili y: Risks and Con idence-Building Measu es.
Cen e o a New Ame ican Secu i y. Janua y.
www.cnas.o g/publica ions/ epo s/ai-and-in e na ional-
s abili y- isks-and-con idence-building-measu es.
Jobin, Anna, Ma cello Ienca and E y Vayena. 2019. “The global
landscape o AI e hics guidelines.” Na u e Machine In elligence
1: 389–99. h ps://doi.o g/10.1038/s42256-019-0088-2.
La sson, S e an and F ed ik Hein z. 2020. “T anspa ency in
a i icial in elligence.” In e ne Policy Re iew 9 (2).
h ps://policy e iew.in o/concep s/
anspa ency-a i icial-in elligence.
Lindsay, Jon R. 2023/2024. “Wa Is om Ma s, AI Is om
Venus: Redisco e ing he Ins i u ional Con ex o Mili a y
Au oma ion.” Texas Na ional Secu i y Re iew 7 (1):
29–47. h ps://doi.o g/10.26153/ sw/50674.
Lyons, Joseph B., Ma hew A. Cla k, Alan R. Wagne and Ma hew
J. Schuelke. 2017. “Ce i iable T us in Au onomous Sys ems:
Making he In ac able Tangible.” AI Magazine 38 (3):
37–49. h ps://doi.o g/10.1609/aimag. 38i3.2717.
Ma ijan, B anka. 2021. “Unaccep able isk and au onomous
weapons.” The Ploughsha es Moni o 42 (4).
www.ploughsha es.ca/publica ions/unaccep able-
isk-and-au onomous-weapons.
Ma qua d , James J. 2011. T anspa ency and Ame ican
P imacy in Wo ld Poli ics. Abingdon, UK: Ashga e.
Maye , Michael. 2023. “T us ing machine in elligence: a i icial
in elligence and human-au onomy eaming in mili a y
ope a ions.” De ense & Secu i y Analysis 39 (4): 521–38.
h ps://doi.o g/10.1080/14751798.2023.2264070.
Milla , Jason. 2015. “Expe es imony p o ided by Jason Milla
o he In o mal Mee ing o Expe s on Le hal Au onomous
Weapons Sys ems (LAWs), held wi hin he amewo k o he
UN Con en ion on Ce ain Con en ional Weapons (CCW),
Gene a, Swi ze land, Ap il 15, 2015.”
h ps://unoda-documen s-lib a y.s3.amazonaws.com/
Con en ion_on_Ce ain_Con en ional_
Weapons_-_In o mal_Mee ing_o _Expe s_(2015)/
Jason%2BMilla %2B-%2BMeaning ul%2BHuman%2B
Con ol%2Band%2BDual-Use%2BTechnolo.gy.pd .
Na ional Academies o Sciences, Enginee ing, and Medicine.
2022. Human-AI Teaming: S a e-o - he-A and Resea ch
Needs. Washing on, DC: The Na ional Academies P ess.
Newman, Ma issa. 2023. “Is ael’s new mili a y AI sys ems selec
a ge s and plan missions ‘in minu es.’” The Japan Times,
July 17. www.japan imes.co.jp/news/2023/07/17/
wo ld/is ael-quie ly-embeds-ai-mili a y-sys ems/.
Pe ch, Je emy, Shuang Di and Wal e Nelson. 2022. “Opening he
Black Box: The P omise and Limi a ions o Explainable Machine
Lea ning in Ca diology.” Canadian Jou nal o Ca diology 38
(2): 204–13. h ps://doi.o g/10.1016/j.cjca.2021.09.004.
Puscas, Ioana. 2022. “Con idence-Building Measu es o A i icial
In elligence: A F aming Pape .” Gene a, Swi ze land: Uni ed
Na ions Ins i u e o Disa mamen Resea ch. h ps://unidi .o g/
wp-con en /uploads/2023/05/Con idence-Building_Final.pd .
Say, Ma k. 2024. “Ada Lo elace Ins i u e u ges e alua ions
o ad anced AI.” UKAu ho i y. July 29.
www.ukau ho i y.com/a icles/ada-lo elace-
ins i u e-u ges-e alua ions-o -ad anced-ai/.
Scha e, Paul. 2023a. Fou Ba leg ounds: Powe in he Age o A i icial
In elligence. New Yo k, NY: W. W. No on & Company.
———. 2023b. “Ame ica Can Win he AI Race.”
Fo eign A ai s, Ap il 4. www. o eigna ai s.com/
uni ed-s a es/ai-ame ica-can-win- ace.
Sch aagen, Jan Maa en. 2023. “Responsible use o AI in mili a y
sys ems: p ospec s and challenges.” E gonomics 66 (11):
1719–29. h ps://doi.o g/10.1080/00140139.2023.2278394.
SIPRI. 2015. “LAWS a he CCW: T anspa ency and in o ma ion sha ing
measu es.” Ap il 17. h ps://docs-lib a y.unoda.o g/
Con en ion_on_Ce ain_Con en ional_
Weapons_-_In o mal_Mee ing_o _Expe s_
(2015)/20150416_CCW_TRANSPARENCY.pd .
Taddeo, Ma ia osa io, Da id McNeish, Alexande Blancha d
and Elizabe h Edga . 2019. “E hical P inciples o A i icial
In elligence in Na ional De ence.” Philosophy & Technology 34:
1707–29. h ps://doi.o g/10.1007/s13347-021-00482-3.
14 CIGI Pape s No. 315 — Janua y 2025 • B anka Ma ijan
Va ella, Lau a and Ray Acheson. 2024. “Gene al S a emen s.” CCW
Repo 12 (1): 3–8. h ps:// eachingc i icalwill.o g/
images/documen s/Disa mamen - o a/
ccw/2024/gge/ epo s/CCWR12.1.pd .
Ve gun, Da id. 2023. “U.S. Endo ses Responsible AI
Measu es o Global Mili a ies.” US Depa men o
De ense. No embe 22. www.de ense.go /News/
News-S o ies/A icle/A icle/3597093/us-endo ses-
esponsible-ai-measu es- o -global-mili a ies/.
Wood, Na han Gab iel. 2024. “Explainable AI in he mili a y
domain.” E hics and In o ma ion Technology 26, 29.
h ps://doi.o g/10.1007/s10676-024-09762-w.
Yo dano a, Ts e elina. 2015. “The T anspa ency–Secu i y Dilemma
in Na ional and In e na ional Con ex (A Compa a i e Analysis
o he UN and NATO’s T anspa ency/Sec ecy Policies).”
A pape p esen ed a he Fou h Global Con e ence on
T anspa ency Resea ch in Lugano, Swi ze land, June 4–6.
67 E b S ee Wes
Wa e loo, ON, Canada N2L 6C2
www.cigionline.o g
