148
AI and image banks: A esea ch
me hodology
Pe e F eixa
Uni e si a Pompeu Fab a, Spain
h ps://o cid.o g/0000-0002-9199-1270
Ma Redondo-A olas
Uni e si a de Ba celona, Spain
h ps://o cid.o g/0000-0002-0000-8593
Lluís Codina
Uni e si a Pompeu Fab a, Spain
h ps://o cid.o g/0000-0001-7020-1631
Ca los Lopezosa
Uni e si a de Ba celona, Spain
h ps://o cid.o g/0000-0001-8619-2194
F eixa, P., Redondo-A olas, M., Codina, L., & Lopezosa, C. (2025). AI and image banks:
A esea ch me hodology. In J. Gualla , M. Vállez, & A. Ven u a-Cisquella (Coo ds). Digi al
communica ion. T ends and good p ac ices (pp. 148-160). Ediciones P o esionales de la
In o mación. h ps://doi.o g/10.3145/cu icom.11.eng
149
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
Abs ac
This chap e p esen s a me hodological amewo k o analysing gende bias and he p es-
ence o sociocul u al s e eo ypes in p o essional s ock image banks, wi h a speci ic ocus on
he isual esul s e u ned by pho og aphic and AI-gene a ed pla o ms. The s udy is based
on he hypo hesis ha neu al p omp s — hose lacking explici e e ences o gende , age,
o e hnici y — should, in he absence o cul u al o echnical bias, yield a balanced isual
ep esen a ion ac oss di e en social ca ego ies. Any signi ican de ia ion om such p opo -
ionali y may indica e he exis ence o implici biases o ecu en isual clichés. To explo e
his, he au ho s analysed images e ie ed om ou p o essional pla o ms — wo based
on con en ional pho og aphy and wo elying on AI image gene a ion. A sys em o coded
indica o s was de eloped o classi y he ep esen a ions in e ms o gende , age, e hnici y,
unc ional di e si y, beau y no ms, and depic ed ac ions. The me hodology excluded g oup
images and nea -iden ical a ian s o ensu e di e si y and analy ical igou . The indings e-
eal ha AI-based pla o ms mo e consis en ly align wi h use p omp s (60.36%) compa ed
o adi ional pho og aphic da abases (44.84%). Howe e , bo h ypes o pla o ms exhibi
s e eo ypical pa e ns, sugges ing a pe sis ence o isual opes and clichés. The p oposed
me hodology p o es e ec i e in de ec ing hese biases and o e s a ans e able analy ical
amewo k. The chap e aims o con ibu e o b oade e o s owa ds mo e inclusi e isual
cul u es, encou aging u he in e disciplina y esea ch on algo i hmic image gene a ion and
ep esen a ion in digi al media.
Keywo ds
Gende bias; S e eo ypes; S ock image pla o ms; A i icial in elligence; Visual ep esen a ion;
Image p omp s; Algo i hmic in e p e a ion; Iconog aphic analysis; Media ep esen a ion.
1. In oduc ion
Image banks a e among he main esou ces used by he media o isually complemen he
con en hey publish (Codina, 2011; Kamin, 2023). Acco ding o se e al au ho s (Gynnild,
2017; Mo ensen & Gade, 2023), he downsizing o p o essional pho og aphy s a in adi-
ional media ou le s has led o a g owing eliance on s ock images o illus a e jou nalis ic
con en , a he expense o commissioned pho og aphy (Mo ensen e al., 2023; Fe y, 2023;
Mo ensen e al., 2024; Hugues, 2024). In his pionee ing s udy, Tsang (1984) had al eady
obse ed ha by he la e 1970s and ea ly 1980s, only abou one qua e o he pho og aphs
published we e p oduced in-house. The emainde o igina ed om agencies, image banks,
and eelance pho og aphe s. I is he e o e no su p ising ha image banks, alongside news
agencies, ha e become he isual esou ce wi h he g ea es capaci y o in luence and es ab-
lish isual s anda ds (Machin, 2004; Machin & Polze , 2015; F osh, 2015, 2020).
Companies specialising in he dis ibu ion o s ock pho og aphy ha e signi ican ly expanded
hei capaci y o gene a e all kinds o illus a ions and isual esou ces — bo h pho o ealis ic
and ic ional — h ough he inco po a ion o a i icial in elligence (AI) echnologies (Pe dices-
Cas illo & Pe ianes-Rod íguez, 2011; Codina & Lopezosa, 2020; V abič-Dežman, 2024). AI
150
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
has also enabled he eme gence o new playe s in he isual in o ma ion ma ke , who use
image-gene a ion ools o p oduce as amoun s o isual con en . Th ough sea ch in e aces,
media ou le s can access, loca e, and selec he images ha bes sui hei needs wi hin a
highly compe i i e ma ke (Alla d, 2023; B igh , 2023; Soji -Pejcha & C abapple, 2023; F eixa
& Redondo-A olas, 2023; 2024).
The g owing p ominence o s ock images in he media has elici ed mixed esponses. C i ics
o he use o bo h s ock images and AI-gene a ed con en in jou nalism a gue, among o he
poin s, ha such esou ces ep oduce clichés and s e eo ypes, hus u he pe pe ua ing biases
and unde mining media c edibili y (Mo ensen e al., 2023; Aiello e al., 2023). Mo eo e , hey
con end ha hese images lack edi o ial o documen a y alue, as hey a e no ied o speci ic
e en s (Mo ensen e al., 2024). As F osh (2003) desc ibes, “The mos s iking ea u e o s ock
images is hei close ela ionship wi h classi ica ion. Insc ibed wi hin ad e ising ideology, hese
images pa icipa e in a ep esen a ion o eali y shaped by selec i e ca ego isa ion” (p. 91).
None heless, while p o ide s o s ock image y a e o en c i icised o ep oducing s anda dised
and s e eo ypical codes and o mulas — ega dless o he sou ce o he pho og aphs — Kwak
and An (2016) demons a ed, using deep lea ning sys ems and la ge da ase s, how media
ou le s use published p ess images o con ey speci ic messages and o en ein o ce clichés
and biases, i espec i e o he images’ o igins.
Conce ns abou he ep oduc ion o biases —be hey gende , acial, cul u al, o ideological—
in p ess image y ha e been widely acknowledged and explo ed bo h in gene al e ms (Mille ,
1975; Luebke, 1989; Rodge s & Tho son, 2000; Rodge s e al., 2007; Thu low e al., 2020;
F eixa e al., 2025) and in he con ex o poli ical communica ion (Waldman & De i , 1998;
Goodnow, 2010; Rönnback e al., 2025).
These s udies ha e employed combined quali a i e and quan i a i e analy ical me hodolo-
gies, including s a is ical coun ing and bo h iconog aphic and iconological con en analysis.
O en, image analysis has been complemen ed by he examina ion o he accompanying
ex . Mille ’s pionee ing wo k, o example, in ol es coun ing he numbe o imes men and
women appea in he Los Angeles Times and The Washing on Pos o e he cou se o a yea ,
iden i ying he newspape sec ions in which hey a e published and he oles depic ed (Mille ,
1975). By con as , Goodnow (2010) ocuses his s udy on a smalle sample o images, sub-
jec ing hem o a de ailed semio ic analysis ha del es in o he eading and in e p e a i e
codes inhe en in jou nalis ic image y. As no ed, Kwak and An (2016) adop ed a quan i a i e
app oach, wo king wi h a much la ge sample comp ising wo million images. Thu low e al.
(2020) and F eixa e al. (2025), meanwhile, applied semio ic analysis o samples o 600 images
o assess he p esence o s e eo ypes on s ock image pla o ms.
This chap e p esen s a me hodological p oposal o conduc ing isual esea ch on images,
clea ly de ining he obse a ion pa ame e s and he sample size. The p oposed sys em has
p o en e ec i e in s udies examining gende bias and s e eo ypes in bo h pho og aphic s ock
images and AI-gene a ed isuals (F eixa e al., 2025).
2. Image banks and sea ch pla o ms
Image banks ely on digi al pla o ms o make hei collec ions a ailable o clien s. Th ough
sea ch in e aces, use s can access, loca e, and selec hose images ha bes sui hei needs.
This is a highly compe i i e ma ke . Sea ches (conduc ed ei he h ough keywo ds o ex ual
151
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
p omp s) gene a e isual galle ies in esponse, p esen ed as se s o humbnail images which
use s can hen selec and download. Depending on he complexi y o he sea ch, hese sys-
ems may e u n any hing om a hand ul o an o e whelming numbe o images.
To de elop he p oposed me hodological model, we begin wi h he assump ion ha he
p e-selec ions o e ed by hese sys ems a e designed o align wi h use que ies. Howe e ,
hey may also e eal algo i hmic biases and e lec he p esence o s e eo ypes, such as hose
ela ed o gende o ace, by p io i ising ce ain isual cons uc s o e o he s.
The wo king p emise is ha hese pla o ms deploy ained algo i hms o p o ide use s
wi h he mos ele an possible esul s — hose ha bes ma ch hei expec a ions. Wi h
some a ia ions, nea ly all sea ch in e aces o e unc ionali ies o con ex ualise and e ine
sea ches. Some o hese ea u es a e echnical, such as image size and o ma ; o he s a e
comme cial, such as copy igh - ela ed op ions. In addi ion, many pla o ms allow use s o
speci y p e e ences ela ed o a ibu es such as age, e hnici y, o gende o he indi iduals
ep esen ed. All o hem pe mi he use o keywo ds o help ob ain mo e p ecise esul s.
Le us conside he simples —and mos common— scena io: wha happens when a sea ch is
conduc ed using an especially neu al p omp , wi hou any indica ion o a speci ic g oup o
people o be ep esen ed, and wi hou using he ools a ailable o il e o pa ame e ise he
sea ch? How does he sys em ope a e o ensu e ha i mee s he clien ’s expec a ions?
The hypo hesis is ha he less in o ma ion p o ided o he pla o m, he g ea e he deg ee
o in e p e a i e esponsibili y he sys em mus assume o sa is y he que y. When a p omp is
ague o ambiguous, pla o ms end o e u n nume ous esul s wi h a wide ange o nuances
and a ia ions, he eby inc easing he likelihood ha a leas some o hem will ma ch he
use ’s needs.
This me hodology is speci ically designed o p o oke such a si ua ion. We a gue ha his
scena io p o ides he mos sui able esea ch en i onmen in which o obse e he p esence
o clichés, biases, and s e eo ypes wi hin image banks. When a neu al p omp is submi ed
o a pla o m —wi hou any addi ional pa ame e s— he sys em is compelled o deli e a
la ge and a ied selec ion o esponses o accommoda e a b oad ange o po en ial use s.
The obse a ion, classi ica ion, and cha ac e isa ion o he a iables p oduced by he sys-
em in esponse o an explici ly neu al eques can e eal dominan isual cons uc s, highly
s e eo yped pa e ns o esponse, and ul ima ely expose he p esence o biased codes and
encoding sys ems.
3. Pla o m selec ion, p omp de ini ion, and sample size
The comme cialisa ion o s ock images akes place in an ex emely compe i i e ma ke , whe e
a wide a ay o pla o ms o e hei se ices o bo h p i a e use s and media o ganisa ions.
These may be ei he paid o ee o cha ge, and may specialise in ei he con en ional pho-
og aphy o AI-gene a ed image y. Based on he c i e ion o p o essional epu a ion (Pii onen,
2022; 2023; May, 2024), ou p o essional pla o ms we e selec ed o his esea ch — wo
specialising in adi ional pho og aphic images and wo o e ing AI-gene a ed isuals. I was
also ensu ed in he selec ion p ocess ha each pla o m is capable o e u ning a leas 50
images in esponse o a gi en que y, wi hin he isual galle y p o ided, wi hou equi ing he
use o e o mula e he sea ch o modi y he p omp .
152
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
Table 1
Selec ion o p o essional s oxk image pla o ms
Pho og aphic image banks AI-Gene a ed image banks
Shu e s ock h ps://www.shu e s ock.com/pho os Lexica (S able di usion) h ps://lexica.a /
Ge y Images h ps://www.ge yimages.com/ Adobe S ock (opción IA) h ps://s ock.adobe.com/es/
Sou ce: Own elabo a ion.
To o mula e he sea ch que ies, we op ed o use he na u al language p omp unc ionali y
o e ed by hese pla o ms, a he han elying on sea ch equa ions based on Boolean
ope a o s (Ba elle, 2006; Abadal & Codina, 2008; Codina, 2018). As p e iously men ioned,
ou hypo hesis is ha he use o neu al p omp s ac i a es he algo i hmic in e p e a ion
sys ems embedded in hese applica ions, o deli e wha he sys em deems o be he mos
app op ia e esponse o he eques .
This app oach is conside ed pa icula ly sui able o de ec ing po en ial biases and s e eo-
ypes: he esul s e u ned by he sys em a e likely o align wi h wha is conside ed common,
no ma i e, o s anda dised wi hin i s da abase o aining co pus.
The decision o use na u al language p omp s ins ead o sea ch equa ions is mo i a ed by
h ee main easons. Fi s , because o hei po en ial o igge he algo i hmic in e p e i e
logic we aim o s udy. Second, because we a e in e es ed in explo ing and es ing his mode
o in e ac ion wi h image banks, gi en i s eme ging and inno a i e na u e. And hi d, because
we wish o analyse wha occu s when use s, despi e some imes ha ing he op ion o ca y ou
Boolean sea ches, choose ins ead o in e ac h ough p omp s.
To cons uc he expe imen al amewo k o he s udy, a p elimina y es (p e- es ) was
conduc ed using wo pla o ms: Lexica and Shu e s ock. The objec i e was o assess he
e ec i eness o he p omp s selec ed o he s udy. In his ini ial phase, six p omp s we e
o mula ed combining one cons an elemen wi h wo a iables. The cons an elemen was
he ph ase “smiling pe son”, selec ed o he ollowing conno a ions:
– I s singula o m a ou s he ep esen a ion o a single indi idual.
– The e m does no imply a speci ic gende .
– The e m does no imply a speci ic age.
– The e m does no imply a speci ic e hnici y.
– The e m does no imply any pa icula ac i i y.
– I is expec ed ha he pe son will be depic ed on ally, despi e no aming o composi-
ional ins uc ions being p o ided.
As a iables, wo di e en ia ing elemen s we e in oduced. Fi s , a spa ial indica o was added
o p o ide geog aphical con ex o he que ies. Th ee al e na i e se ings we e selec ed: ci y,
landscape, and beach.
The choice o he e m landscape was based on a p elimina y compa ison be ween he use o
landscape and coun yside. As no signi ican di e ences we e obse ed in he esul s e u ned
by he pla o ms, landscape was selec ed as he mo e commonly used e m.
Second, we in oduced a ia ions ega ding he o mal ende ing o he images, wi h he aim
o analysing how AI sys ems in e p e he concep o pho og aphic ep esen a ion. Th ee o -
153
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
mula ions we e es ed: pho og aphy o , pic u e o , and image wi h pho o ealis ic look. This
compa ison seeks o examine no only how he sys ems in e p e and ep esen deg ees o
ealism o hype ealism, bu also he composi ional and aming elemen s ha eme ge in he
images gene a ed o selec ed in esponse o each p omp .
The p omp s used in he p e- es we e as ollows:
– Pic u e o a smiling pe son in a landscape.
– Pho og aphy o a smiling pe son in a landscape.
– Image o a smiling pe son in a landscape. Pho o ealis ic look.
Simila es s we e conduc ed o he wo o he se ings, ci y and beach.
In all cases, p omp s con aining he e m pho og aphy yielded mo e hype ealis ic esul s
han hose using pic u e o image o (…) pho o ealis ic look. The la e op ion occasionally
p oduced ca ica u e-like images wi h clea ly dis o ed ea u es, cha ac e is ic o g aphic illus-
a ion.
I was obse ed ha inco po a ing he e m pho og aphy in he p omp —added, as ex-
plained, o delimi he o mal na u e o he image in con as o illus a ion— may in oduce
a ia ions in he esul s, pa icula ly when analysing he “ac i i ies” depic ed in he scenes.
To assess he impac o his a iable, addi ional es s we e ca ied ou using wo compa able
p omp s: Pho og aphy o a smiling pe son in he ci y and Smiling pe son in he ci y, applied
ac oss he ou pla o ms unde analysis.
In he case o Shu e s ock, including he e m pho og aphy did indeed esul in changes:
se e al images ea u ed people aking pho og aphs, whe eas such scenes we e sca cely
p esen when he e m was omi ed. By con as , Ge y Images e u ned i ually iden ical
esul s in bo h cases, wi h no ele an a ia ions a ibu able o he use o pho og aphy.
On he Lexica pla o m, as p e iously no ed, he e m pho og aphy unc ioned as an e ec i e
il e o exclude unwan ed illus a ions om he esul s. Howe e , i s inclusion o omission
did no signi ican ly al e any o he analy ical pa ame e s unde conside a ion. Simila ly, on
Adobe S ock, no subs an ial changes in isual ou comes we e obse ed in ela ion o he use
o he e m. As wi h Lexica, he absence o pho og aphy o en led o he appea ance o illus-
a ions among he e ie ed images.
In summa y, i was ound ha he inclusion o exclusion o he e m pho og aphy had no
impac on he ep esen a ion o biases linked o gende , ace, age, o beau y- ela ed s e eo-
ypes. Only in he case o Shu e s ock was a sligh de ia ion de ec ed conce ning he ac ions
depic ed, a ibu able o p omp s explici ly including he e m pho og aphy.
Consequen ly, he ollowing h ee p omp s we e ul ima ely submi ed o all ou pla o ms:
– Pho og aphy o a smiling pe son in a landscape.
– Pho og aphy o a smiling pe son in he ci y.
– Pho og aphy o a smiling pe son on he beach.
These p omp combina ions yielded he analy ical sample, composed o a o al o 600 images.
The sample was gene a ed by collec ing he i s 50 images e u ned by each o he ou
pla o ms, p o ided hey me he es ablished c i e ia and cons i u ed a cohe en esponse
154
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
o he h ee p oposed que ies. The size o his sample —600 images— has p e iously been
employed in simila s udies, such as ha o Thu low e al. (2020), which we e likewise aimed
a iden i ying biases and s e eo ypes in digi al image banks.
4. Wha can be obse ed and how
This s udy is based on he hypo hesis ha p omp s o mula ed wi hou any explici e e ence
o he gende , age, o e hnici y o he indi iduals depic ed should, in he absence o cul u al
o echnical bias, yield a p opo ional dis ibu ion ac oss he a ious obse able ca ego ies.
Any de ia ion om his heo e ical balance is in e p e ed in he specialis li e a u e as a po-
en ial indica ion o gende dispa i ies and he p esence o clichés o s e eo ypes (Cook &
Cusack, 2010; Cas illo-Mayén & Mon es-Be ges, 2014).
The o e ep esen a ion o pa icula g oups —whe he in e ms o gende , pheno ype, o age
ange— may e eal isual epe i ions: images ha a e simila in na u e, ea u ing ecu ing
composi ional s uc u es o iconic elemen s ha e lec cul u ally accep ed o ms o coding.
These, he e o e, can be in e p e ed as s e eo ypes (Ángeles-Galiano, 2023).
In he ield o media, such codi ied isual o mulas, opes, and s e eo ypes ha e p o en o be
highly e ec i e communica i e ools o illus a ing e en s. They a e pa o p o essional ou-
ines in isual p oduc ion and a e especially p e alen in pho og aphic image y (Baeza, 2001;
F eixa y Redondo-A olas, 2022). Media ou le s no only ep oduce hese isual cons uc s, bu
also con inually upda e hem h ough sub le a ia ions, adap ing hem o he p e e ences and
expec a ions o hei audiences. In doing so, hey con ibu e o he ongoing main enance and
symbolic enewal o hese codes (Quin & McMahon, 1997; No aes-Ci janic, 2017).
The obse a ion o o mal and hema ic epe i ions, as well as he iden i ica ion o common
elemen s ac oss he images analysed, acili a es he de ec ion and desc ip ion o he mos
ecu en s e eo ypes wi hin he isual co pus. Fo his pu pose, iconological me hods a e
employed, and mo e speci ically, iconog aphic desc ip ion is used o iden i y hose a ibu es
po en ially associa ed wi h isual biases s emming om socio-cul u al clichés o s e eo ypes
(D ain ille, 2018; Ha iman & Lucai es, 2007, 2016; Pano sky, 1979).
The applica ion o a sys ema ic analysis shee is p oposed as a me hodological ool o as-
sessing he p esence o gende bias and s e eo ypes, s uc u ed a ound he pa ame e s and
indica o s de ailed in he ollowing sec ion.
Any image ea u ing mo e han one pe son, clea ly depic ing a g oup scene, o lacking a
human igu e al oge he , was excluded om he analysis in o de o ensu e consis ency and
compa abili y o he esul s. Addi ionally, he ex ual in o ma ion associa ed wi h each image
was used as suppo o aid in i s classi ica ion.
4.1. Pa ame e s and indica o s: Biases and s e eo ypes
Below is a desc ip ion o he indica o s p oposed o obse a ion, including he de ini ion o
each indica o , he obse a ion p ocedu e, and he alues used.
155
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
4.1.1. Indica o : Gende
– De ini ion: The p omp s a e designed o e ie e images ha may depic one o mo e indi-
iduals. The e is no explici ins uc ion speci ying he gende o he pe son(s) ep esen ed.
– P ocedu e: I he image depic s a single indi idual, classi y hem as male, emale, o unde-
ined. Reco d he numbe o i ems ep esen ing each ca ego y.
– Values: Female / Male / No de ined.
4.1.2. Indica o : Age
– De ini ion: The p omp s a e designed o e ie e images ha may depic one o mo e in-
di iduals. The e is no explici ins uc ion ega ding he age o he pe son(s) ep esen ed.
– P ocedu e: I he image depic s a single indi idual, classi y he age g oup ep esen ed.
Reco d he numbe o i ems in each ca ego y.
– Values: Child / Young pe son / Adul / Olde adul .
4.1.3. Indica o : E hnici y
– De ini ion: The p omp s a e designed o e ie e images ha may depic one o mo e
indi iduals. The e is no explici ins uc ion ega ding he e hnici y o e hnici ies o he pe -
son(s) ep esen ed.
– P ocedu e: I he image depic s a single indi idual, classi y he e hnici y ep esen ed. Re-
co d he numbe o i ems in each ca ego y.
– Values: Caucasian / Asian / A ican Ame ican / La ino-Medi e anean / Indigenous Ame i-
can / Middle Eas e n.
4.1.4. Indica o : Func ional di e si y
– De ini ion: The p omp s a e designed o e ie e images ha may depic one o mo e in-
di iduals. The e is no explici ins uc ion as o whe he he indi iduals shown may o may
no p esen any o m o unc ional di e si y.
– P ocedu e: I he image depic s a single indi idual, classi y hem acco ding o di e en
ca ego ies o unc ional di e si y. Reco d he numbe o i ems in each ca ego y.
– Values: Mo o disabili y / Senso y disabili y / In ellec ual disabili y / No disabili y.
4.1.5. Indica o : Beau y s anda ds
– De ini ion: The p omp s a e designed o e ie e images ha may depic one o mo e
indi iduals. The e is no explici ins uc ion as o whe he he indi iduals shown con o m o
pa icula beau y s anda ds o codes o isual ep esen a ion.
– P ocedu e: I he image depic s a single indi idual, classi y a ious isual elemen s ha
may be associa ed wi h beau y s anda ds. The ecu ence o such elemen s will allow o a
sha ed labelling sys em. Reco d he numbe o i ems in each ca ego y.
– A p io i alues: None.
– Possible alues: Make-up / Sunglasses / Glasses / Ha s, caps, helme s / Jewelle y / Wa ches
/ Summe clo hing / Win e clo hing / Backpacks / Handbags / O he .
156
AI and image banks: A esea ch me hodology
Pe e F eixa; Ma Redondo-A olas; Lluís Codina; Ca los Lopezosa
Digi al communica ion. T ends and good p ac ices
4.1.6. Indica o : In en o y o ac ions
– De ini ion: The p omp s a e designed o e ie e images ha may depic one o mo e
indi iduals in unde ined se ings. The e is no explici ins uc ion as o whe he he
indi iduals should be pe o ming any pa icula ac ion. Ac ions may se e as indica o s o
p e-es ablished and s e eo yped beha iou al codes.
– P ocedu e: I he image depic s a single indi idual, classi y he ac ions being pe o med.
The ecu ence o elemen s will enable he es ablishmen o a sha ed labelling sys em.
Reco d he numbe o i ems in each ca ego y.
– A p io i alues: None, al hough a highe incidence o ep esen a ions in ol ing pho o-
g aphic de ices is expec ed due o he na u e o he p omp .
– Possible alues: Taking a sel ie / Looking a a map / Reading a book / Using a sma phone
o able / Ges u ing / Wo king / Ea ing o d inking / Using a came a o echnology /
Playing / D i ing o being in a ca / Dancing / Using public anspo / Doing spo /
Walking / Running / O he .
Once he esul s a e ob ained, compa isons a e p oposed o each pa ame e and indica o
ac oss he di e en pla o ms o con i m o ule ou he p esence o biases and s e eo ypes.
5. Expec ed esul s and limi a ions
The p ocess o ob aining he sample is subjec o pa icula i ies s emming om he algo i h-
mic unc ioning o he pla o ms used. Despi e he p ecision ha p omp s may con ey, he
sys ems do no always e u n images ha s ic ly ma ch he eques . In his s udy, i was ound
ha he esul galle ies equen ly include g oup pho og aphs, po ai s o couples, o images
in which no human igu e appea s a all.
Fu he mo e, adi ional pho og aphic s ock lib a ies end o o e mul iple e sions o he
same scene. These a ia ions ypically p esen minimal di e ences, usually ela ed o aming
o composi ion. Fo ins ance, in esponse o a p omp eques ing a po ai o a woman in an
u ban s ee , he sys em may e u n i e o six i ually iden ical images, di e ing only sligh ly
in aming o angle.
To cons uc he inal sample, wo ypes o images we e emo ed: so-called alse posi i es —
ha is, images no co esponding o he eques ed con en — and edundan a ia ions o he
same sho , o a oid dis o ion in he analysis and o ensu e g ea e isual di e si y.
I was obse ed ha image banks based on a i icial in elligence end o e u n mo e accu a e
esul s in esponse o use p omp s. By way o illus a ion, he esul s showed ha AI-based
pla o ms achie ed a 60.36% a e o p omp alignmen , compa ed o 44.84% o con en-
ional pho og aphic s ock banks. This di e ence can la gely be a ibu ed o he di e en
business models unde pinning each ype o pla o m. While adi ional banks a e based on
he exploi a ion o p e-exis ing a chi es —which leads hem o o e a ailable images e en
when hey do no ully ma ch he que y— AI-d i en pla o ms a e designed o gene a e new
con en ha mo e closely aligns wi h he use ’s eques . As a esul , hey p io i ise p omp
accu acy o e di e si y o esul s.
The p oposed me hodological sys em has p o en e ec i e o obse ing he de ined indi-
ca o s. The esul s ob ained (F eixa e al., 2025) e eal small bu signi ican a ia ions in he
