Papae angelou, Cha is; Vo a, Fabio
A icle
T ading nuance o scale? Pla o m obse abili y and
con en go e nance unde he DSA
In e ne Policy Re iew
P o ided in Coope a ion wi h:
Alexande on Humbold Ins i u e o In e ne and Socie y (HIIG), Be lin
Sugges ed Ci a ion: Papae angelou, Cha is; Vo a, Fabio (2025) : T ading nuance o scale? Pla o m
obse abili y and con en go e nance unde he DSA, In e ne Policy Re iew, ISSN 2197-6775,
Alexande on Humbold Ins i u e o In e ne and Socie y, Be lin, Vol. 14, Iss. 3, pp. 1-31,
h ps://doi.o g/10.14763/2025.3.2037
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/330355
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/3.0/de/deed.en
Volume 14 |
T ading nuance o scale? Pla o m
obse abili y and con en go e nance
unde he DSA
Cha is Papae angelou Uni e si y o Ams e dam c.papae[email p o ec ed]
Fabio Vo a Uni e si y o Ams e dam .a.[email p o ec ed]
DOI: h ps://doi.o g/10.14763/2025.3.2037
Published: 17 Sep embe 2025
Recei ed: 11 No embe 2024 Accep ed: 20 June 2025
Funding: This wo k was suppo ed by he Du ch Minis y o Educa ion, Cul u e and
Science unde G an 024.005.017 (G a i a ion Resea ch P og amme “Public Values in
he Algo i hmic Socie y”).
Compe ing In e es s: The au ho has decla ed ha no compe ing in e es s exis ha
ha e in luenced he ex .
Licence: This is an open-access a icle dis ibu ed unde he e ms o he C ea i e
Commons A ibu ion 3.0 License (Ge many) which pe mi s un es ic ed use,
dis ibu ion, and ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly
ci ed. h ps://c ea i ecommons.o g/licenses/by/3.0/de/deed.en
Copy igh emains wi h he au ho (s).
Ci a ion: Papae angelou, C., & Vo a, F. (2025). T ading nuance o scale? Pla o m
obse abili y and con en go e nance unde he DSA. In e ne Policy Re iew, 14(3).
h ps://doi.o g/10.14763/2025.3.2037
Keywo ds: Digi al Se ices Ac (DSA), Pla o m go e nance, Pla o m obse abili y,
T anspa ency, Con en mode a ion
Abs ac : The Digi al Se ices Ac (DSA) ma ks a pa adigma ic shi in pla o m go e nance,
in oducing mechanisms like he S a emen o Reasons (SoRs) da abase o os e anspa ency and
obse abili y o pla o ms’ con en mode a ion p ac ices. This s udy in es iga es he DSA
T anspa ency Da abase as a egula o y mechanism o enabling obse abili y, ocusing on he
au oma ion and e i o ial applica ion o con en mode a ion ac oss he EU/EEA. By analysing 439
million SoRs om eigh Ve y La ge Online Pla o ms (VLOPs), we ind ha he as majo i y o
con en mode a ion decisions a e en o ced au oma ically and uni o mly ac oss he EU/EEA. We also
iden i y signi ican disc epancies in con en mode a ion s a egies ac oss VLOPs, wi h TikTok,
YouTube and X exhibi ing he mos dis inc p ac ices, which a e u he analysed in he pape . Ou
indings e eal a s ong co ela ion be ween au oma ion and he speed o con en mode a ion,
au oma ion and he e i o ial scope o decisions. We also highligh se e al limi a ions o he
da abase, no ably he lack o language-speci ic da a and inconsis encies in how SoRs a e epo ed
by VLOPs. We conclude ha despi e such sho comings, he DSA and i s T anspa ency Da abase
may enable a wide cons ella ion o s akeholde s o pa icipa e in pla o m go e nance, pa ing he
way o mo e meaning ul pla o m obse abili y.
Issue 3
In oduc ion
The Digi al Se ices Ac (DSA; Regula ion (EU) 2022/2065, 2022) has been de-
sc ibed as a “ anspa ency machine” (Zo ne a, 2024) wi h he po en ial o c ea ing
a “global anspa ency egime” (Helbe ge & Samuelson, 2024). The DSA, among
o he s, equi es pla o ms1 o p o ide a jus i ica ion o hei con en decisions
(e.g., g ounds o con en emo al o es ic ion) in he o m o a S a emen o Rea-
sons (SoRs; DSA A . 17(1)) o shed ligh on he o en-opaque decision-making
p ocesses o hei con en mode a ion sys ems. These SoRs a e hen au oma ically
uploaded and agg ega ed o a publicly accessible online da abase ope a ed by he
Eu opean Commission (DSA A . 24(5)) pu ing in place a sys em o “au oma ed
anspa ency” (Kaushal e al., 2024).
The ea ly impac o he DSA is al eady appa en . Fo example, esea che s ha e
unco e ed inconsis encies in he sel - epo ed p ac ices o X, which claims o ely
exclusi ely on manual mode a ion in i s SoRs submissions bu acknowledges using
au oma ed ools in i s anspa ency epo s (De gache a e al., 2023; D olsbach &
P öllochs, 2023). Such cases highligh he la en egula o y po en ial o he DSA
and demons a e how anspa ency can enable pla o m obse abili y (Riede &
Ho mann, 2020), a dynamic and analy ical go e nance mechanism ha acili a es a
deepe unde s anding o pla o ms’ go e nance p ac ices (Lee ssen, 2024).
In his pape , we app oach he T anspa ency Da abase as an ins umen o pla -
o m obse abili y, which builds upon anspa ency, o conduc an empi ical case
s udy on he kinds o insigh s ha can be in e ed om he T anspa ency Da a-
base. We seek o unde s and he ex en o which he SoR da abase os e s obse -
abili y, which we ecognise as a go e nance mechanism ha is necessa y, in addi-
ion o anspa ency, o ensu e pla o m accoun abili y and deepen ou unde -
s anding o how pla o ms ope a e and exe cise hei powe o e ou digi al public
sphe es ( an Dijck e al., 2018). Ou s udy, speci ically, ocuses on explo ing he
use o au oma ion in con en mode a ion and he di e ences in he e i o ial
scope o con en mode a ion p ac ices, as well as he implica ions he eo o he
DSA and pla o m go e nance.
We conduc ed ou analysis wi h wo leading ques ions in mind: how do con en
go e nance decisions a y among pla o ms and membe -s a es in he EU, and in
wha ways does he use o au oma ion in mode a ion di e ac oss pla o ms, he
1. The DSA co e s all in e media y pla o ms, including e-comme ce and ma ke places (e.g., Zalando,
App S o e), ide-sha ing apps (e.g., Ube ), and o he s.
2 In e ne Policy Re iew 14(3) | 2025
EU and i s membe -s a es. We concen a ed ou analysis on he mode a ion p ac-
ices o eigh digi al Ve y La ge Online Pla o ms (VLOPs), namely X, Facebook, In-
s ag am, YouTube, TikTok, Snapcha , Pin e es and LinkedIn, mos o which ha e
been ounda ional o ou mode n digi al public sphe es and he b oade digi al
pla o m ecosys em.
We used R, a p og amming language ha is widely used o da a analysis (R Co e
Team, 2021), o e ie e 439 million SoRs om a pe iod o ou mon hs (25 Sep-
embe 2023 o 25 Janua y 2024). We conclude ha he DSA enables a a ie y o
ac o s (Helbe ge e al., 2018), including esea che s and membe s o ci il socie y,
o make use o echnically-sophis ica ed egula o y mechanisms, such as he
T anspa ency Da abase, supplemen ing he mo e adi ional go e nance mecha-
nism o anspa ency. In ha sense, he T anspa ency Da abase, despi e i s sho -
comings, ma ks a c ucial u ning poin in go e ning digi al pla o ms as i pa es
he way o obse ing pla o ms’ beha iou and con en mode a ion p ac ices in a
way ha was una ainable be o e. Ou pape con ibu es o a g owing body o li -
e a u e on pla o m go e nance in egula ed digi al en i onmen s, especially in
he EU. In doing so, i also o e s a c i ical pe spec i e on he DSA’s egula o y and
go e nance ambi ions.
In summa y, we disce ned h ee key indings. Fi s , we ound ha 99% o all mod-
e a ion decisions we e applied uni o mly ac oss he EU/EEA, exhibi ing a endency
o uni o m applica ion o con en mode a ion s a egies. Second, we iden i ied
a ia ions in how VLOPs en o ce con en mode a ion, pa icula ly in e ms o e-
gional di e ences, wi h only h ee, namely TikTok, You ube and X, epo ing con-
duc ing e i o ially speci ic con en mode a ion. Thi d, we obse ed ha he use
o au oma ion in con en mode a ion co ela es wi h as e en o cemen imelines
and uni o m applica ion, whe eas manual mode a ion o en means longe delays
and mo e e i o ially speci ic applica ion. This las inding also aises ques ions
abou he in e ac ion o EU and na ional legal amewo ks and he VLOPs’ capaci y
o handle “low- esou ce languages” (Nicholas & Bha ia, 2023). Las , we conside
some echnical sho comings o he da abase and egula o y blindspo s o he DSA
such as he lack o an obliga ion o epo he language o he con en ha was
mode a ed. The nex sec ions a e s uc u ed as ollows: i s we si ua e ou pape
heo e ically h ough a li e a u e e iew o wo ks ele an o con en mode a ion,
pla o m go e nance and egula ion in o de o build ou case o pla o m obse -
abili y; second, we expound on ou me hodological app oach; hi d, we p esen ou
indings in de ail; inally, we discuss he implica ions o ou indings o he DSA
and pla o m go e nance.
3 Papae angelou, Vo a
Concep ual amewo k: Con en
mode a ion, pla o m go e nance and
he need o pla o m obse abili y
Ou concep ual amewo k is de eloped in h ee s eps: i s , we do a b ie li e a u e
e iew o schola ship on con en mode a ion as an ins umen al pa o pla o m
go e nance and, subsequen ly, pla o ms’ powe ; second, con inuing ou li e a u e
e iew, we ace he ise o au oma ion in con en mode a ion, along wi h i s s uc-
u al limi a ions; and hi d, we ad ance he need o adop pla o m obse abili y
as a c i ical lens h ough which o s udy pla o m go e nance and as a i ing egu-
la o y app oach, as we la e show h ough ou case s udy on he DSA.
Con en mode a ion and pla o m go e nance
In ecen yea s, a ich body o esea ch om c i ical media o legal schola ship has
ex ensi ely explo ed he poli ics and mechanisms o con en mode a ion and i s
implica ions o democ acy. Wi hou a emp ing an exhaus i e discussion o he
li e a u e, we d aw upon schola ship on pla o m go e nance which has illumina -
ed how pla o ms na iga e he balance be ween sel - egula ion and s a e o e -
sigh , highligh ing he in o mal and o mal mechanisms ha unde pin pla o m
go e nance (Go wa, 2019b, 2019a; Papae angelou, 2021). Go wa has also, c ucial-
ly, unde sco ed he ole o powe dynamics be ween pla o m go e nance s ake-
holde s in in luencing con en mode a ion, aking s ock o he poli ical eali y
wi hin which pla o m go e nance is insc ibed (Go wa, 2024). In a simila ein,
G i in (2023) has heo ised abou he poli ics behind con en mode a ion deci-
sions, demons a ing how pla o ms balance legal, public and comme cial in e es s
in hei go e nance s a egies. This s and o schola ship has, he e o e, showcased
he inhe en ly poli ical p ocess o mode a ing con en ha is o en obscu ed by
pla o ms’ sophis ica edly opaque sys ems, which include ex ac i e p ocesses o
in isible labou (Robe s, 2019) and which a e eiled behind discou ses o neu ali-
y (Gillespie, 2010).
Subsequen ly, con en mode a ion is closely ied o he pla o ms’ poli ical powe ,
ha is, he powe o shape he no ms, ules and condi ions unde which in o ma-
ion ci cula es online and, by ex ension, in luence he s uc u es o ou digi al pub-
lic sphe es and he capaci y o ci izens o o m poli ical opinions (Gillespie, 2018;
Helbe ge , 2020). Con en mode a ion, hus, eme ges no simply as a echnical
4 In e ne Policy Re iew 14(3) | 2025
unc ion bu a cen al mechanism o pla o m i ms (G immelmann, 2017), go e n-
ing he socioeconomic, cul u al and poli ical in e ac ions o end-use s, comple-
men o s and o he ele an ac o s ha con ene in he mul i-sided ma ke s ha
la ge pla o ms cons i u e (see Poell e al., 2021 Chap e 4). F om his pe spec i e,
social media pla o ms ely hea ily on con en mode a ion o main ain hei ad e -
ising-d i en business models. The abili y o egula e, il e and o ganise use -gen-
e a ed con en enables pla o ms o c ea e an en i onmen conduci e o hei e -
enue goals, whe eby he goal is o maximise p o i s h ough he inc ease o use
engagemen , while minimising ha m ul o con o e sial con en ha could alien-
a e end-use s o ad e ise s (G i in, 2022; Jimenez-Du an, 2022).
Au oma ion in con en mode a ion and i s s uc u al
limi a ions
A i icial In elligence (AI) and au oma ed decision-making p ocesses (Bloch-Wehba,
2020; Gillespie, 2020; Go wa e al., 2020) ha e been c ucial in dealing wi h he
massi e olume o use -gene a ed con en uploaded e e y ins an on social media
pla o ms, while educing he cos o con en mode a ion (e.g., ins ead o hi ing
human e iewe s). Typically, AI e e s o sys ems ha use machine lea ning models
o iden i y, classi y, o p edic pa e ns in da a. Combined wi h au oma ion and au-
oma ed-decision making, which ypically e e o p og ammed p ocesses ha op-
e a e wi hou human in e en ion (a leas no necessa ily isible labou ), hese
socio echnical sys ems ha e gi en ise o he model o “algo i hmic comme cial
con en mode a ion” (Go wa e al., 2020, p. 3). The la e e e s o “sys ems ha
classi y use -gene a ed con en based on ei he ma ching o p edic ion, leading o
a decision and go e nance ou come” (Go wa e al., 2020, p. 3).
The le el o au oma ion a ies depending on he ype o con en i is deployed
agains and he speci ic legal amewo k in place. Fo ins ance, he iden i ica ion o
e o is o copy igh ed pieces o con en is a p edominan ly au oma ed p ocess
based on hash-ma ching, whe eas oxic o ha e ul con en migh in ol e a combi-
na ion o au oma ion and human e iewing (Go wa e al., 2020, p. 7). Algo i hmic
mode a ion is also eiled wi h a neu ali y discou se, obscu ing he a o emen-
ioned poli ical-economic dimension o pla o m go e nance and eme ges as “a
sha ed imagina y o echnological solu ionism” (Udupa e al., 2023, p. 1). Au oma-
ion and echnology, in his con ex , a e p esen ed as p o ide s o e icien , accu-
a e and swi solu ions o issues ele an o ou democ acy (c . Poell e al., 2021
Chap e 7). As a esul , ele an schola ship has no ed an inc easing eliance on
and emb acing o au oma ion o “sani ise” online spaces (G i in, 2022), o en a
5 Papae angelou, Vo a
he expense o benign, ma ginalised, o non-con o ming discou ses (A e, 2022).
In ha sense, au oma ion in mode a ion, and he discu si e emb ace he eo , no
only ob usca es i s sho comings bu , mo e b oadly, seeks o depoli icise pla o m
go e nance and blu he messy eali y o i s poli ical economy (Go wa e al., 2020,
p. 13). Mo eo e , his dynamic ends o obscu e he c i ical-ye -unde alued human
labou ac i i ies like da a anno a ion and con en e iewing, which a e essen ial
o aining he models unde gi ding he au oma ed p ocesses o con en mode a-
ion and, la gely, o he poli ical economy o con en mode a ion (Posada, 2022;
Robe s, 2019). Such labou , o en ou sou ced in he majo i y wo ld, is a he co e o
pla o ms’ con en mode a ion indus ial app oach (Caplan, 2018), ha is a mas-
si e au oma ed, sys ema ised and s anda dised sys em o mode a ion designed o
scalabili y and c oss-ju isdic ional applica ion. Las , he inadequacy o au oma ion
in con en mode a ion becomes e en mo e p onounced when engaging wi h “low-
esou ce languages”, ha is, languages ha ha e a sca ce digi al oo p in and,
hus, ha e no been inco po a ed in he aining o hese models (Nicholas & Bha-
ia, 2023)
The limi s o anspa ency and he u n o pla o m
obse abili y
These s uc u al p oblems and limi a ions o indus ial au oma ed con en mode -
a ion highligh he inadequacy o cu en go e nance mechanisms ha a e oo ed
in echno-solu ionism. I is no su p ise, hen, ha he esponse, p ima ily by pla -
o ms bu also some policymake s and expe s, was o ask o (mo e) anspa ency
o un a el he “black box” (Pasquale, 2016). Indeed, aced wi h public ou c y, majo
social media pla o ms pu a ious anspa ency mechanisms in place, anging
om sel - epo ed con en mode a ion anspa ency epo s o es ablishing spe-
cialised eams– he indus y e m is ‘T us and Sa e y’–dealing wi h con en mode -
a ion (Go wa & Ash, 2019). These de elopmen s led pla o ms o “ne wo ked pla -
o m go e nance” s uc u es ha in ol e a combina ion o au oma ed means, dedi-
ca ed wo k o ce and he implica ion o ex e nal s akeholde s o go e n hei digi-
al spaces and, pu po edly, os e us in in ol ed s akeholde s (Caplan, 2023). In
ac , in his no el con igu a ion, companies o e -emphasised he ole o ans-
pa ency and o “openness and access […] o es ablish us wo hiness o ex e nal
ac o s (Caplan, 2023, p. 3462).
Howe e , as c i ical pla o m schola s ha e no ed, anspa ency can only go so a
as a pla o m go e nance mechanism due o he inhe en socio echnical complexi-
ies o digi al pla o ms (Ananny & C aw o d, 2016), pushing o he adop ion o
6 In e ne Policy Re iew 14(3) | 2025
pla o m obse abili y as a mo e sui able go e nance mechanism (Riede & Ho -
mann, 2020; Lee ssen, 2024). Indeed, anspa ency is o en conside ed and c i-
iqued as a s a ic (Lee ssen, 2024) o m o “ isibili y managemen ”, ha is a highly
cu a ed and media ed p ocess o disclosu e o da a, in o ma ion and insigh s (Fly-
e bom, 2019, p. 18) by pla o ms o ex e nal s akeholde s. Bu he unde lying
p emise o anspa ency as a p e equisi e componen o “c ea e he knowledge e-
qui ed o go e n and hold sys ems accoun able” (Ananny & C aw o d, 2016, p. 975)
holds. Hence, schola s s udying pla o m go e nance do no a gue o abandoning
anspa ency2bu a he complemen ing i wi h obse abili y and going e en u -
he .
D awing on ecen wo ks on pla o m obse abili y (Riede & Ho mann, 2020; Fe -
a i e al., 2023, Lee ssen, 2024), we concep ualise i as a con inuous and p ocess-
o ien ed app oach ha acknowledges he complexi ies o obse ing and unde -
s anding digi al pla o ms. Unlike anspa ency, obse abili y is an ac i e, p agma -
ic and explici ly subjec i e p ac ice, decen e ing ou ocus on “ he algo i hm” o
encompass mo e ace s o he p ocesses ha cons i u e pla o m in as uc u es
and hei au oma ed-decision making sys ems (Lee ssen, 2024, pp. 7-8). Riede and
Ho mann, speci ically, se h ee p inciples as ounda ional o obse abili y: ob-
se abili y in ela ion o, o in a ou o he public in e es (no ma i e p inciple);
obse abili y as a con inuous and dynamic p ocess (socio echnical p inciple); ob-
se abili y as a ca alys o collabo a i e o ms o pla o m go e nance (go e -
nance p inciple). Addi ionally, he au ho s connec he concep di ec ly o egula-
ion, a guing o a dual app oach in ol ing “ egula ing o obse abili y” and mak-
ing obse abili y pa o ou egula o y s a egies o pla o m i ms (Riede and
Ho mann, 2020, p. 22).
We u he unde s and obse abili y as an a emp o e e se neolibe al amings
o anspa ency as objec i e, which sough o depoli icise he no ion and disen an-
gle i om i s poli ical-economic, social and cul u al powe s uc u es (Bi chall,
2021). The e o e, we concep ualise obse abili y no simply as cons ained o he
disclosu e o in o ma ion bu as he egula o y capaci y o con inuously moni o and
sc u inise pla o m ope a ions, while conside ing he eme gen na u e o pla o m
beha io , in luenced by use in e ac ions, sociopoli ical and economic s a egies o
pla o m owne s and pla o ms’ au oma ed-decision making algo i hms. Conse-
quen ly, obse abili y does no simply enable a echnical unde s anding o pla -
o m echnologies bu se s in place he a chi ec u e o a ne wo k o ac o s (Ca-
2. Fo a comp ehensi e compa ison o anspa ency and obse abili y, including o hei me apho s,
see Lee ssen, 2024.
7 Papae angelou, Vo a
plan, 2023), beyond adi ional and o en-opaque egula o y p ocesses like audi -
ing (Te zis e al., 2024) o ac i ely and coope a i ely pa icipa e in digi al go e -
nance (Helbe ge e al., 2018; Kelle & Lee ssen, 2020). In ha sense, obse abili y
complemen s and goes u he han anspa ency o empowe go e nance ac o s
o ac in an in o med manne , educe in o ma ion asymme ies and inc ease us
among s akeholde s and, gene ally, imp o e he condi ions necessa y o accoun -
abili y (Bi chall, 2021, p. 6; Go wa e al., 2020, p. 11). Taking he socio echnical as-
pec o pla o m obse abili y “as a egula o y p og am” (Lee ssen, 2024, p. 9) a
s ep u he , we hold ha i should also include mo e opaque ac i i ies ha exis
in pla o ms’ alue chains and which enable hei ope a ions, anging om labou
ela ions (e.g., eal- ime in o ma ion abou human mode a o s employed o con-
ac ed) o en i onmen al implica ions (e.g., eal- ime insigh s in o compu ing
powe used along wi h i s en i onmen al oo p in ). I we a e o lea e he “snap-
sho logic” (Riede & Ho mann, 2020, p. 7) behind o a mo e comp ehensi e and
holis ic egula o y amewo k, hen we mus conside he laby in hine, in e ela ed
complexi ies o pla o ms’ poli ical economy.
Building on Riede and Ho mann’s concep ual ounda ion, Fe a i and colleagues
(2023) p opose a complemen a y amewo k o he go e nance o gene a i e AI
sys ems, which in oduces h ee “o e sigh condi ions” (p. 2) o obse abili y: in-
dus ial obse abili y ( he capaci y o sc u inise he in e linked, ma e ial and poli -
ical-economic laye s o gene a i e AI models and sys ems om a mac o-le el),
public inspec abili y ( he need o deep egula o y inspec ion o all laye s cons i-
u ing such sys ems) and echnical modi iabili y ( he capaci y o make undamen-
al changes o hese sys ems o ensu e compliance wi h o e sigh mechanisms).
The simila i ies a e e iden in he p ominence gi en by bo h amewo ks o he
public in e es and he ole o public egula ion, bu wha s ands ou is he “ ech-
nical modi iabili y” condi ion as i seems o indica e and ad oca e o u he ac-
ion–beyond egula o y– o be aken as a esponse o he insigh s we may de i e
om obse ing pla o m and AI sys ems. A guably, his las condi ion shows ha
obse abili y oo mus be u he expanded when dealing wi h ech companies
and hei complex alue chains.
The Digi al Se ices Ac and egula ed
pla o m obse abili y
The EU’s DSA (Regula ion (EU) 2022/2065, 2022), o ed in 2022 and enac ed in ull
o ce in 2024, imposes due diligence obliga ions on pla o ms o ensu e a sa e
8 In e ne Policy Re iew 14(3) | 2025
a e on ou indings ela ing o hese wo cen al dimensions o pla o m go e -
nance: he e i o ial scope o con en mode a ion decisions and he au oma ion o
de ec ion and en o cemen p ocesses.
Te i o ial aspec o con en mode a ion
One o he mos s iking insigh s om ou analysis is he e i o ial dis ibu ion o
SoRs ac oss he Eu opean Economic A ea (EEA), EU and indi idual membe -s a es.
The op wo e i o ial scopes conce n he EEA (wi h and wi hou Iceland and No -
way; Table 2), accoun ing o 50.40% (o 221M SoRs) and 48.41% (o 213M SoRs)
o all SoRs, while he hi d one conce ns he EU, accoun ing o 0.32% o ou co -
pus (o 1M o SoRs). Pu simply, almos all mode a ion decisions (99.13% o 435M
ou o he 439M SoRs o ou co pus) made by VLOPs in ou sample we e applied
uni o mly ac oss he en i e EEA and EU. As ega ds speci ic membe s a es, we
ound ha Ge many leads in e ms o coun y-speci ic SoRs, wi h 588K SoRs
(0.13%), ollowed by F ance (432K, 0.10%) and I aly (323K, 0.07%); his is o be ex-
pec ed gi en ha hese coun ies hos he la ges popula ions in he EU.
To in e be e coun y- and language-speci ic insigh s, we decided o elimina e
pla o ms whose con en mode a ion decisions we e applied uni o mly ac oss he
EU o EEA. This app oach na owed ou ocus o he emaining 0.87% (o app oxi-
ma ely 4M SoRs), whe e e i o ial a ia ion was e iden . To illus a e, Facebook’s
(in blue colou ; Figu e 1) con en mode a ion decisions a e almos exclusi ely ap-
plied on an EU and EEA le el, p o iding no meaning ul di e gence o analysis a
he membe -s a e le el.
15 Papae angelou, Vo a
TABLE 2: Top 10 e i o ial scope o con en mode a ion decisions in he EU/EEA
RANK TERRITORIAL SCOPE SORS %
SORS
1 EEA (NO ICELAND)5 221M 50.40%
2 EEA 213M 48.41%
3 EUROPEAN UNION 1M 0.32%
4 GERMANY 588K 0.13%
5 FRANCE 432K 0.10%
6 ITALY 323K 0.07%
7 EEA (NO ICELAND OR NORWAY) 197K 0.04%
8 POLAND 185K 0.04%
9 SPAIN 169K 0.04%
10 FINLAND, HUNGARY, LIECHTENSTEIN, LITHUANIA, NORWAY, POLAND,
SLOVENIA 129K 0.03%
The e o e, we ocused on YouTube, X and TikTok, as hey we e he only pla o ms
whose da a showed some a ia ion in con en mode a ion decisions ac oss di e -
en membe -s a es (Table 4). TikTok applied mos o i s con en mode a ion deci-
sions uni o mly ac oss he EEA le el. Simul aneously, he pla o m domina ed ou
co pus showcasing he dominance o he ideo-sha ing pla o m in he EU ma ke
(235M SoRs, 53,54%). As such, TikTok demons a es a p e e ence o an EEA-wide
con en mode a ion, wi h minimal a ia ion be ween indi idual coun ies. When i
engages in coun y-speci ic mode a ion (e.g., in he Ne he lands o F ance), he
olume is signi ican ly lowe han i s egional en o cemen , which is in line wi h a
mo e indus ialised and s eamlined con en mode a ion s a egy. YouTube’s da a
indica es a mo e nuanced app oach o e i o ial con en mode a ion han TikTok,
which is also demons a ed by he geog aphic dispe sion o i s decisions ac oss a -
eas in ou sca e plo (Figu e 1). YouTube is, hus, mo e likely o apply a e i o i-
alised app oach o mode a ing con en in indi idual membe -s a es, he Ne he -
lands, F ance and I aly in pa icula , e lec ing a g ea e a ie y in i s o e all s a e-
gy, in spi e o applying mos o i s decisions on an EEA le el. Las ly, X epo s o be
5. THE TERRITORIAL SCOPES IN TABLE 2 REFLECT SELF-REPORTED LABELS IN THE DSA
TRANSPARENCY DATABASE. DUE TO INCONSISTENT LABELLING ACROSS PLATFORMS AND LACK OF
PROOF-CHECKING MECHANISMS BY THE COMMISSION, OVERLAPPING CATEGORIES SUCH AS ‘EEA’
AND ‘EEA (NO ICELAND)’ MAY REFER TO THE SAME GEOGRAPHIC SCOPES, AFFECTING THE
ACCURACY OF DISTINCT TERRITORIAL CATEGORIES’ NUMBERS.
16 In e ne Policy Re iew 14(3) | 2025
implemen ing i s con en mode a ion decisions in a highly e i o ialised manne ,
wi h he Ne he lands anking again he highes among indi idual membe -s a es.
Au oma ion in con en mode a ion
We hen p oceeded in isualising he da a in wo sca e plo s o ep esen he de-
g ee o au oma ion and i s impac on he speed o en o cemen o con en mode -
a ion decisions ac oss di e en e i o ies (Figu es 1 & 2). Fo cla i y, in he i s
sca e plo (Figu e 1), he ho izon al axis ep esen s he deg ee o au oma ed de-
ec ion, anging om manual de ec ion (on he le ) o ully au oma ed de ec ion
(on he igh ), while he e ical axis ep esen s he olume o SoRs pe 1,000
AMARs, wi h he sizes o he ci cles ep esen ing he olume o SoRs in a egion. In
he second sca e plo (Figu e 2), he ho izon al axis ep esen s he same as he
i s sca e plo , while he e ical axis ep esen s he median day o he applica-
ion o con en mode a ion decisions.
Figu e 1 - S a emen o Reasons conce ning au oma ically o manually de ec ed con-
en pe 1,000 Mon hly Ac i e Use s
17 Papae angelou, Vo a
Figu e 2 - Sca e plo showing he ela ion be ween ime and au oma ion o de ec ion
wi h he e i o ial scope o con en mode a ion decisions o TikTok (le -hand side)
and YouTube ( igh -hand side)
Ou indings he e also e eal signi ican a ia ions. TikTok shows a high eliance
on au oma ed de ec ion, pa icula ly when mode a ing con en on an EEA-le el
bu also in some coun ies like Spain and he Ne he lands (Figu e 1). No ably, as
he case o F ance shows in bo h sca e plo s, when con en is de ec ed manually
i is also mo e likely o be deal manually, aking signi ican ime o a decision o
be applied. Fo example, he median day o en o cemen o TikTok in F ance in
ou sample was mo e han 30 days (Figu e 2). Pu simply, he sho e en o cemen
delays in hese coun ies sugges ha TikTok uses au oma ion o add ess mos o
he con en .
YouTube seems o employ mo e o a hyb id app oach, p edominan ly elying on
au oma ed de ec ion me hods supplemen ed wi h manual e iew. Gene ally, he
same pa e n as TikTok was also obse ed in his case, wi h manual de ec ion
demons a ing a co ela ion wi h manual en o cemen . Wha is mo e, wi h a close
look in ou da a, we disce ned ha YouTube’s median eac ion ime, including au-
oma ically de ec ed con en , can ex end in occasions o e 100 days (Figu e 2),
wi h delays inc easing d ama ically when manual e iew is in ol ed and when de-
cisions conce n speci ic e i o ial scope (e.g., F ance).
18 In e ne Policy Re iew 14(3) | 2025
In con as , X s ands ou as an ou lie , epo edly elying exclusi ely on manual
mode a ion. The eason why X’s da a was no isualised as a ime sca e plo in
ou analysis is ha he pla o m indica es mode a ing all con en manually and on
he same day as de ec ed (Figu e 1). As men ioned ea lie , his obse a ion con a-
dic s i s public anspa ency epo s which men ion he use o au oma ed means6,
making X’s da a un eliable. This disc epancy has been also co obo a ed by o he
s udies (De gache a e al., 2023; D olsbach & P öllochs, 2023; Kaushal e al.,
2024).
Finally, ou ables (3 & 6) con aining da a on human e iewe s indica e ha lan-
guages such as F ench and Du ch ecei e mo e a en ion om human mode a o s
ac oss pla o ms, especially o YouTube and TikTok. YouTube, o ins ance, allo-
ca es 176 F ench-speaking mode a o s and 24 Du ch-speaking mode a o s, while
TikTok alloca es 687 F ench-speaking mode a o s and 167 Du ch-speaking mode -
a o s, he la e being no ably high compa ed o o he pla o ms. Again, gi en ha
hese a e sel - epo ed igu es using non-s anda dised me hodologies, any c oss-
pla o m compa ison emains inhe en ly limi ed. This could sugges ha be e
language co e age allows pla o ms o manually mode a e mo e con en and, hus,
p o ide–a guably–mo e accu a e decisions ha , as shown in Figu e 2, ake mo e
ime han au oma ed decisions.
Ha ing said ha , he numbe o human mode a o s is no co ela ed o he use o
au oma ion in con en mode a ion. As Klonick’s (2018) analysis o Facebook’s con-
en mode a ion sys em has shown, hese pla o ms ha e di e en ie s o add ess
con en mode a ion issues acco ding o hei pe cei ed se e i y and impo ance
(see also Caplan & Gillespie, 2020). In o he wo ds, i migh e y well be ha Tik-
Tok s ill p edominan ly elies on au oma ed means o mode a e con en bu em-
ploys– he mos –human e iewe s o deal wi h mo e sophis ica ed issues (Table 3).
The e o e, a key limi a ion in unde s anding he ope a ional capaci ies o pla -
o ms ac oss EU membe -s a es lies in he he e ogenei y o epo ing me ics e-
ga ding human mode a ion s a .
6. A ailable he e: h ps:// anspa ency.x.com/dsa- anspa ency- epo .h ml.
19 Papae angelou, Vo a
TABLE 3: Human mode a o s employed by pla o ms based on hei anspa ency epo s
PLATFORM PERIOD AMARS MODERATORS AMARS/MODERATOR
FACEBOOK 01/04/2023-30/09/2023 259.000.000 1.362 190.161
INSTAGRAM 01/04/2023-30/09/2023 259.000.000 1.362 190.161
YOUTUBE 01/01/2023-30/06/2023 416.600.000 1.974 211.043
LINKEDIN 01/01/2023-30/06/2023 45.200.000 146 309.589
SNAPCHAT 01/02/2023-30/07/2023 101.973.520 1.545 66.002
PINTEREST 01/03/2023-30/06/2023 124.000.000 1.963 63.168
X (TWITTER) 01/04/2023-30/10/2023 126.120.951 2.496 50.529
TIKTOK 01/04/2023-30/09/2023 125.000.000 5.827 21.451
Discussion
Con en and pla o m go e nance in he EU
Ou analysis showcases how he DSA and, he T anspa ency Da abase in pa icula ,
is a s ep owa d os e ing a dynamic way o s udying pla o m go e nance and,
hus, in enabling pla o m obse abili y. Ou indings o eg ound he inhe en en-
sion in he pu sui o as e and mo e accu a e con en mode a ion decisions, espe-
cially ia he use o au oma ion. They also demons a e how calls o as e eac-
ion imes inhe en ly in i e mo e au oma ion and, hus, an ele a ed isk o igno e
i al con ex in con en mode a ion, in addi ion o he embedded limi a ions o au-
oma ion.
While mos con en mode a ion decisions in ou sample we e en o ced uni o mly
ac oss he EEA and EU–accoun ing o 99.13% o all SoRs–ou indings also poin
o a e i o ialised aspec o pla o m mode a ion, pa icula ly in membe s a es
like he Ne he lands, F ance, I aly and Spain. These coun ies consis en ly ap-
pea ed in ou da a as egions whe e pla o ms applied e i o ialised con en mod-
e a ion decisions and ha we e mo e likely o include manual means o mode a-
ion. Howe e , his e i o ial a ia ion does no necessa ily imply a legal agmen-
a ion. Ra he , i sugges s ha pla o ms a e mo e selec i e abou he localisa ion
o hei con en mode a ion s a egies wi hin he o e a ching ha monised egula-
o y en i onmen es ablished by he DSA. Addi ionally, his selec i e e i o ialisa-
ion o con en mode a ion–namely conce ning ce ain a o emen ioned membe -
20 In e ne Policy Re iew 14(3) | 2025
s a es–shown by ou indings indica e a ie ed app oach o con en mode a ion
ha no only has o do wi h na ional legal amewo ks and au ho i ies bu , po en-
ially, also wi h how la ge a ma ke is and, subsequen ly, how many esou ces a e
in es ed in he o m o human e iewe s speaking he local language.
Ou analysis e ealed a no able co ela ion be ween e i o ialised con en mode -
a ion and he manual de ec ion o lagged con en , pa icula ly h ough hi d-pa y
no i ica ions. Pu simply, when con en is mode a ed a he le el o speci ic mem-
be -s a es a he han uni o mly ac oss he EU o EEA, i is mo e likely o ha e
been lagged manually (e.g., h ough epo s by use s o o he hi d-pa ies) a he
han de ec ed h ough au oma ed sys ems. Unde A icle 16 o he DSA, pla o ms
a e equi ed o implemen use - iendly mechanisms enabling indi iduals o epo
po en ially illegal con en . Howe e , in ou da ase o TikTok, X, and YouTube, he
majo i y o lagged con en ell unde he ca ego y o “o he ypes o no i ica ions.”
This ague classi ica ion does no p o ide de ails abou he o igin o hese no i i-
ca ions, whe he om indi iduals, o ganisa ions, o o he en i ies.
Also, pla o ms a e no obliga ed unde DSA A icle 9 o issue S a emen s o Rea-
sons (SoRs) o con en emo ed a he eques o judicial o adminis a i e au ho -
i ies. This omission es ic s ou abili y o di e en ia e be ween no i ica ions s em-
ming om public au ho i ies e sus o he hi d pa ies in he con ex o he T ans-
pa ency Da abase. Howe e , a close examina ion o he ypes o con en lagged
by hi d pa ies (Table 5) o e s some insigh s. Fo ins ance, in ellec ual p ope y
iola ions o e whelmingly domina e YouTube’s mode a ed con en , unde sco ing
he in luence o igh sholde s on pla o m go e nance. On he o he hand, TikTok
and X show b oade di e si y in lagged ca ego ies, including ha m ul speech,
aud and p i acy iola ions. This sugges s ha di e en pla o ms ca e o a ying
ypes o s akeholde s’ demands and obse abili y can help us o eg ound hese la-
en powe dynamics.
A any a e, ou s udy shows ha we a e mo ing owa ds a ha monised amewo k
o pla o m go e nance in he EU. This ha monisa ion, howe e , is la gely en o ced
by pla o ms as he “delega ed en o ce s” unde he DSA (Huso ec, 2024). As
Kaushal e al. (2024) showed in hei s udy o he T anspa ency Da abase, mos o
he con en mode a ed is ound o be incompa ible wi h he pla o ms’ e ms o
se ices and/o communi y guidelines a he han he basis o na ional o EU law.
As a esul , his delega ion and in e p e a ion o he DSA, aises b oade ques ions
abou he powe o pla o ms b oadly and, pa icula ly, he di e si y and ib ancy
o he EU’s digi al public sphe es. Fo ins ance, how can we ensu e ha pla o ms
do no disp opo iona ely silence ma ginalised oices when hey ha e such disc e-
21 Papae angelou, Vo a
ion o e he ope a ionalisa ion o EU’s egula ions?
This delega ion, mo eo e , o en o cemen esponsibili ies o pla o ms, while
p ac ical om a egula o y pe spec i e o ensu e a mo e-o -less ha monised and
“p edic able” egula o y en i onmen (Huso ec & Roche Laguna, 2022), aises im-
po an conce ns abou he p io i isa ion o monocul u al and complian speech
(Douek, 2020; Kelle , 2024) and he po en ial cul i a ion o a sani ised digi al pub-
lic sphe e (G i in, 2022). Impo an ly, as Kelle has epea edly no ed (2022, 2024),
he DSA migh end up a ou ing incumben pla o m i ms ha ha e pu in place
an indus ialised con en mode a ion s uc u e, while disp opo iona ely a ec ing
smalle pla o ms o discou age he expe imen a ion wi h o he sys ems o mode -
a ion, like mo e communi y-o ien ed o a isanal app oaches (Caplan, 2018).
Au oma ion, o be su e, plays a pi o al ole in acili a ing pla o ms’ compliance
wi h he DSA. In ha sense, as demons a ed in ou analysis o TikTok, indus ial-
scale au oma ion is no only necessa y o he indus ial-scale o pla o ms’ con-
en ci cula ion (Gillespie, 2020) bu also s eamlines p ocesses o “au oma ed
anspa ency” (Kaushal e al., 2024) like he T anspa ency Da abase. This dynamic
e lec s a b oade end owa d he indus ialisa ion o con en mode a ion, whe e
pla o ms adop algo i hmic sys ems (Go wa e al., 2020), c ea e specialised T us
and Sa e y eams (Caplan, 2023) and implemen anspa ency mechanisms o com-
ply wi h egula ions.
All hese elemen s gi e way o a “ ac o y-like” app oach o con en mode a ion
(Kelle , 2024), econ igu ing pla o ms in o compliance-d i en en i ies ha p io i-
ise ope a ional e iciency o e nuanced con en mode a ion, p ecisely due o he
sys emic sho comings and p oblems plaguing au oma ion. We also ind ha he
way ha he DSA ope a ionalises pla o m obse abili y does no lea e much
space ei he o he kind o echnical modi iabili y ha Fe a i e al. (2023) consid-
e o be c ucial o digi al go e nance o o expe imen a ions wi h al e na i e
con en mode a ion sys ems. To be clea , hese obliga ions o mally apply only o
designa ed VLOPs. Howe e , he no ma i e g a i as o he DSA isks ein o cing in-
dus ialised, cen alised mode a ion app oaches ac oss he digi al ecosys em. As
such, he DSA isks u he en enching la ge pla o ms’ powe by es ic ing he
po en ial o ha ing a mo e ple ho ic and di e se con en mode a ion ecosys em
which would be mo e likely o be cul i a ed by smalle o decen alised pla o ms,
which may now be p essu ed o align wi h indus ial app oaches o con en go e -
nance (e.g., BlueSky’s app oach o a ‘s ackable’ con en mode a ion sys em; The
BlueSky Team, 2024).
22 In e ne Policy Re iew 14(3) | 2025
I may, mo eo e , depoli icise con en go e nance, alongside he p ocess o making
i legible and accoun able, allowing pla o ms o hide behind a discu si e aming
o con en mode a ion as he p oduc o neu al, au oma ed sys ems ha a e he
only op ion o comply wi h he DSA, a he han as decisions embedded in com-
plex sociopoli ical and cul u al con ex s (Ananny & C aw o d, 2016). The e o e,
while he indus ialisa ion o con en mode a ion has become una oidable due o
he scale and complexi y o digi al pla o ms, i s implica ions o pla o m go e -
nance demand c i ical sc u iny, ha is ou capaci y o obse e and p obe in o hese
sys ems.
Implica ions o pla o m obse abili y
In he con ex o his pape , whe e we ocus on he DSA, obse abili y ansla es o
he sys ema ic and –whe e e possible– eal- ime access o pla o ms’ da a and
he socio echnical sys ems ha allow o he p oduc ion o ha da a and he o e -
a ching ecosys ems. In o he wo ds, i e e s o c ea ion o he ins i u ional and
echnical condi ions necessa y o he sus ained, collabo a i e and si ua ed obse -
a ion o pla o ms. The T anspa ency Da abase in oduced by he DSA pa ially e-
sponds o hese calls o pla o m obse abili y. Re u ning o Riede and Ho mann’s
concep ualisa ion, we show ha he DSA, he e h ough he T anspa ency Da abase,
con ibu es o he no ma i e condi ion by manda ing anspa ency p ac ices ha
aim o se e public sc u iny. I also aligns wi h he socio echnical condi ion by en-
abling public access o s uc u ed and eal- ime da a, hough wi h well-documen -
ed limi a ions. Las , he go e nance condi ion is also a guably me p o ided ha
he Da abase can be accessed and used by a wide a ie y o s akeholde s who may
collec i ely pa icipa e in go e nance p ocesses.
Howe e , once we conside Fe a i and colleagues’ go e nance condi ions (2023),
we unde s and ha he T anspa ency Da abase pa ially mee s he condi ion o in-
dus ial obse abili y as i p o ides some insigh s in o he labou ela ions sup-
po ing pla o ms’ con en go e nance bu he e a e signi ican inconsis encies and
sho comings. I also pa ially mee s public inspec abili y as i makes a as o e
o da a and insigh s a ailable bu p ima ily ocuses on decisions o pla o ms’ au-
oma ed decision-making sys ems and no so much on mo e opaque p ocesses
(e.g., he models unde gi ding con en go e nance). Las , echnical modi iabili y is
absen as he e is no indica ion o how mode a ion sys ems can be econ igu ed in
esponse o obse ed sho comings beyond egula o y p oceedings and po en ial
inancial penal ies. In ha sense, hese SoRs p o ide us wi h jus i ica ions o he
pla o ms’ decisions a he han explain how hese decisions we e made and ca -
23 Papae angelou, Vo a
ied ou o empowe he in ol ed pa ies o ake u he ac ion (Lee ssen, 2024, p.
25).
Rega dless, we hold ha his new ound le el o access o impo an da a migh
imp o e ou chances o eigning in pla o m i ms’ unchecked powe , p ima ily
h ough ex-pos accoun abili y (e.g., ines). Bu gi en he unp eceden ed le el o
en anglemen o pla o m in as uc u es in ou li es and poli ical-economic sys-
ems his migh no longe be–i i e e was–enough. Fo example, as men ioned
ea lie , obse abili y should expand om jus he consume - acing (i.e., down-
s eam) pa o pla o ms’ alue chains o include mo e hidden laye s (i.e., up-
s eam) like he labou condi ions o hei ou sou ced con ac o s and en i onmen-
al impac o hei ma e ial in as uc u es (Te zis, 2023).
Recen s o ies like he layo o 300 human mode a o s om he Ams e dam o ice
o TikTok (Nijssen, 2024) hin a a doubling down on he echnical aspec o mod-
e a ion, namely au oma ion, po en ially exace ba ing he sho comings and p ob-
lems o au oma ion discussed ea lie in his pape . We mus no o ge ha pla -
o m i ms, and hei con en mode a ion sys ems, a e also la ge employe s ope a -
ing on human labou . In ha sense, condi ions o obse abili y should also ex end
o include he wo king spaces and he ela ions o p oduc ion o pla o m capi al-
ism (S nicek, 2017). Doing so should no only allow us o obse e and, ideally, en-
su e ha hese mode a o s wo k unde decen condi ions bu also o be e unde -
s and he alue chains o con en mode a ion and how pla o ms coo dina e hei
wo k o ce globally and, in so doing, c ea e new oppo uni ies o in e en ions. We
unde s and ha such a condi ion canno be solely add essed by pla o m egula-
ion bu may also need mo e adi ional egula o y amewo ks moo ed in labou
law (e.g., con en mode a o s migh need a simila ins umen o he EU’s pla o m
wo k Di ec i e).
Conclusion
This pape se ou o explo e how he DSA’s T anspa ency Da abase ope a ionalises
he concep o pla o m obse abili y. Building on a li e a u e e iew o pla o m
go e nance schola ship, we con ex ualised ou pape wi h a concep ualisa ion o
pla o m obse abili y, p ima ily ollowing Riede and Ho man (2023)’s amewo k
and complemen ing i wi h Fe a i e al. (2023)’s go e nance condi ions. Empi ical-
ly, we analysed nea ly 439 million SoRs om eigh digi al VLOPs, wi h a deepe
di e in o h ee (X, YouTube and TikTok), aiming o unde s and wha kind o insigh s
ela ed o obse abili y we could gain om his no el egula o y ins umen . We
24 In e ne Policy Re iew 14(3) | 2025
EU OFFICIAL LANGUAGE YOUTUBE X TIKTOK
Ge man 231 81 869
G eek 28 0 96
Hunga ian 25 0 63
I ish 0 0 439
I alian 91 2 0
La ian 11 1 9
Li huanian 11 0 6
Mal ese 0 0 0
Polish 99 1 208
Po uguese 464 41 75
Romanian 34 0 167
Slo ak 5 0 44
Slo enian 15 0 45
Spanish 507 20 468
Swedish 16 0 108
To al 1974 2496 5827
31 Papae angelou, Vo a
