ArticlePDF Available

Using Social Media to Monitor Conflict-Related Migration: A Review of Implications for A.I. Forecasting

Authors:

Abstract

Following the large-scale 2015-2016 migration crisis that shook Europe, deploying big data and social media harvesting methods became gradually popular in mass forced migration monitoring. These methods have focused on producing 'real-time' inferences and predictions on individual and social behavioral, preferential, and cognitive patterns of human mobility. Although the volume of such data has improved rapidly due to social media and remote sensing technologies, they have also produced biased, flawed, or otherwise invasive results that made migrants' lives more difficult in transit. This review article explores the recent debate on the use of social media data to train machine learning classifiers and modify thresholds to help algorithmic systems monitor and predict violence and forced migration. Ultimately, it identifies and dissects five prevalent explanations in the literature on limitations for the use of such data for A.I. forecasting, namely 'policy-engineering mismatch', 'accessibility/comprehensibility', 'legal/legislative legitimacy', 'poor data cleaning', and 'difficulty of troubleshooting'. From this review, the article suggests anonymization, distributed responsibility, and 'right to reasonable inferences' debates as potential solutions and next research steps to remedy these problems.
Soc.Sci.2022,11,395.https://doi.org/10.3390/socsci11090395www.mdpi.com/journal/socsci
Article
UsingSocialMediatoMonitorConflictRelatedMigration:
AReviewofImplicationsforA.I.Forecasting
HamidAkinUnver
DepartmentofInternationalRelations,ÖzyeğinUniversity,Istanbul34337,Turkey
Abstract:Followingthelargescale2015–2016migrationcrisisthatshookEurope,deployingbig
dataandsocialmediaharvestingmethodsbecamegraduallypopularinmassforcedmigration
monitoring.Thesemethodshavefocusedonproducing‘realtime’inferencesandpredictionson
individualandsocialbehavioral,preferential,andcognitivepatternsofhumanmobility.Although
thevolumeofsuchdatahasimprovedrapidlyduetosocialmediaandremotesensingtechnologies,
theyhavealsoproducedbiased,flawed,orotherwiseinvasiveresultsthatmademigrants’lives
moredifficultintransit.Thisreviewarticleexplorestherecentdebateontheuseofsocialmedia
datatotrainmachinelearningclassifiersandmodifythresholdstohelpalgorithmicsystemsmoni
torandpredictviolenceandforcedmigration.Ultimately,itidentifiesanddissectsfiveprevalent
explanationsintheliteratureonlimitationsfortheuseofsuchdataforA.I.forecasting,namely
‘policyengineeringmismatch’,‘accessibility/comprehensibility’,‘legal/legislativelegitimacy’,‘poor
datacleaning’,and‘difficultyoftroubleshooting’.Fromthisreview,thearticlesuggestsanonymiza
tion,distributedresponsibility,and‘righttoreasonableinferences’debatesaspotentialsolutions
andnextresearchstepstoremedytheseproblems.
Keywords:conflict;forcedmigration;artificialintelligence;eventdata;bigdataethics
1.Introduction:OpportunitiesandPitfallsofExtractingInformationfromViolent
andMigrationProneRegions
Masshumandisplacementhasoftenbeenabyproductoforganizedviolence.Some
ofthelargestforcedmigrationeventsinworldhistoryhavebeentriggeredbyanthropo
genicdisasters,andinternationalorintrastateconflictsstillremainamongthemostim
mediatecausesofrefugeecrises(LozanoGraciaetal.2010;Schon2019;Selbyand
Hoffmann2012).Sincecivilwarsareespeciallyfoughtnearpopulatedareastoestablish
controloverpopulationcenters,theyareparticularlypronetogeneratingmassdisplace
ment(Steele2009;Lichtenheld2020).Indeed,frequentexchangesofterritoryasaresult
ofconflict—suchasinUkraine,Syria,SouthSudan,Myanmar,DRCongo,Somalia,Cen
tralAfricanRepublic,Afghanistan,andIraq—generateanoverwhelmingmajorityofthe
worldrefugeepopulation.Bothinhumanhistoryandtoday,armedconflictsremainone
ofthemostsignificantpredictorsofforceddisplacementand,often,thesizeofforced
migration.
Violenteventsleadtothemassuprootingofnoncombatantsfortworeasons.First,
theimmediatelifethreateningpotentialofconflictsleadstothedepartureofentirevil
lages,towns,andsometimescities,fleeingintosaferregions(BasuandPearlman2017).
Often,largescalepopulationdisplacementgetstriggeredbecauseciviliansexpectharsh
treatmentortargetingbytheconquerorsandleavetoescapefromsuchfate(Conteand
Migali2019).Second,theaftershocksofaconflictcreatesignificantinfrastructureandsus
tenanceproblemsthatleadtothemassdepartureofcivilianstogainaccesstoessential
resourcesandservices(Humphrey2013;Rajabalietal.2009).Quitefrequently,secondary
effectssuchasthedestructionofhousing,sanitationandelectricalinfrastructure,absence
Citation:Unver,HamidAkin.2022.
UsingSocialMediatoMonitor
ConflictRelatedMigration:A
ReviewofImplicationsforA.I.
Forecasting.SocialSciences11:395.
https://doi.org/10.3390/socsci11090395
AcademicEditor:CarlosArcila
Calderón,TubaBircan
andNigelParton
Received:19May2022
Accepted:22August2022
Published:1September2022
Publisher’sNote:MDPIstaysneu
tralwithregardtojurisdictional
claimsinpublishedmapsandinstitu
tionalaffiliations.
Copyright:©2022bytheauthors.
Submittedforpossibleopenaccess
publicationunderthetermsandcon
ditionsoftheCreativeCommonsAt
tribution(CCBY)license(https://cre
ativecommons.org/licenses/by/4.0/).
Soc.Sci.2022,11,3952of15
oflawandorder,anddisruptionsinfoodsuppliesgenerateescalatinghardshipsinapop
ulationcenter,creatingsustained,lowintensitymigration(SowersandWeinthal2021;
Crush2013).Thisisnottosuggestthatarmedviolenceautomaticallygeneratesforced
migration(sometimescivilianseitherchoosetostayintheclashzoneforvariousreasons,
orhavemobilityproblemsthatpreventsuchprospects),butonecouldsafelyassumethat
organizedviolencetendstocreatepoverty,grievance,andthreatstimulithatusuallycon
tributestolocals’assessmentsofstayingorleaving(Tellez2022;Epstein2010).
Violentorganizedconflictisnotsolelyanindependentvariableinthestudyofdis
placement.Often,climateandseasonaladversities(likebadharvest),inadditiontonatu
raldisasters,maytriggerforcedmigration,andareexacerbatedbypreexistingornewly
emergingformsofconflict(AshandObradovich2020).Insuchcases,althoughviolence
doesnotnecessarilyinitiatemigrationevents,itnonethelessaffectsthetempo,duration,
anddirectionofmigrationflows(Burkeetal.2015).Suchdisastersandclimaterelated
effectsalsocontributetoincreasedviolenceduetodwindlingaccesstoresources,gener
atingadditionalcompetitionoverwater,food,andmedicalsupplies.Dormantgrievances
reemergeduetoawidearrayofdisplacementtriggers,andethnic/religiousgroupsmay
targeteachotherduringmigration,ortheymaybetargetedbyoutsidehostilearmedac
torswhileonthemove(Salehyan2007).
Sincearmedconflictiseither(orboth)anindependentandinterveningvariablein
migrationresearch,reliefagenciesandgovernmentshavebeenexploringwaystoquan
tifyandlogeventdatatoproducemoreinformedanalysesaboutbothformsofinterlinked
crises.Fromtheneedtoestablishmoredetailedandrobustmechanismstomonitorand
forecastforcedmigrationandtopreventviolence,arosemacroleveleventdatasets
(Lubecketal.2003;Nam2006;Eck2012).Thesedatasetsinitiallyloggedviolenceinbinary
forms,andinadyadicfashion,butintandem,anumberofhighlygranularandmulti
directionaldatasetshavebegunappearinginscientificspace(someexamplesare
Chojnackietal.2012;Weidmann2013;DemarestandLanger2018).Inthelastdecade,
thesedatasetshavebeenusednotjustforexplainingconflictandmigration,butalsoto
forecastthem(BlairandSambanis2020;Carammiaetal.2022).
Theneedtoforecastorganizedviolenceanduseitinturntopredictforcedmigration
haslongbeenapriorityforstateactorsandinternationalorganizations.Priortothepro
liferationofthescientificandprivateeventdatarevolution,informationaboutconflicts
andmigrationwasthuslargelysuppliedbyformalmilitaryobservers.Themainproblem
withthesemilitaryreportswasthattheywerepronetocensorshipandreportingbias,as
statestendedtomisrepresentfieldeventsinlinewiththeirnationalinterests(Banerjee
andMacKay2020).Fromthemid20thcenturyonwards,theonsetofhighcirculation
newspapers,radio,andtelevisionchangedthenatureofwarreportingandextractingin
formationaboutconflictzonesbytheriseofwarreportersandaidworkersasanim
portantadditionalsourceofconflicteventdata.Withoutaformalmilitarychainofcom
mandconstraints,warreportersandaidagencieshavebegungeneratingamorediverse
arrayoffieldinformationforbothmilitaryandnonmilitarypurposes,whichwasawel
comeaddition,sincetheysufferedfromfewernationalinterestcensorshipconstraints.
However,ultimately,theybegansufferingfromotherformsofobservationbias.Report
ershadeditorialconstraints(Knightley2002),whereasaidworkerswereincreasinglycon
strainedbytheirsuperiorswhenfieldeventsdidnotconformtoeithereditorialor
agency/donorinterests(Bunceetal.2019).Gradually,moreacademic/scientificprojects
tookpacebythe1970s,suchasDavidSinger’s‘CorrelatesofWarProject’,whichbrought
local,national,andinternationalnewsreportstogethertogeneratearicheraccountoffield
eventsthroughtraditionalmediareports(Eck2012).
However,evenwhenusedaspartofscientificdatasets,traditionalmediadatahasits
biaslimitations:Primarily,bothlocalandinternationalnewsoutletshaveeditorialinter
estsandpoliticalprerogativesthatoftenmurktheirreportingefforts(Ravi2005).Inmost
cases,mediagroupschoosetoornottoreporteventsbasedontheirpreexistingpolitical
tiesoreditorialbiases.Newspaperoutletshavevaryingdegreesofautonomyand
Soc.Sci.2022,11,3953of15
editorialexposuretogovernmentalpressures,renderingtheindependenceofmedia
basedeventreportshighlyskewed(LeeandMaslog2005).Second,mediadatatendsto
focuson‘bigevents’thateitheraffectalargenumberofpeople,orawideswatheofterri
toryinordertomaximizeitsreadershippotential(BaumandZhukov2015).Thisleaves
out‘smallerevents’thatarestillfieldeventsbutareomittedduetoreadershiporsales
considerations.Third,mediagroupstendtooverfocusoneventsthatconcerntheirhost
countryandleaveoutpotentiallyimportanteventsthatdonotconcernthemdirectly
(Weidmann2016).These‘irrelevantevents’canoftenbediscarded—again,forhome
countryreadershipconsiderations—butmayotherwisecontainimportantdetailstoinfer
theinnerworkingsofconflicts.Finally,asstipulatedintheHerman–ChomskyPropa
gandaModel(ChomskyandHerman2002;MullenandKlaehn2010),therearefivefilters
throughwhicheventshavetogothroughinordertoappearinthenews.Theseare(a)
mediaownershipinterests,(b)advertiserinterests,(c)censorshipornudgesfromthegov
ernment,(d)threatsagainstestablishedpower(corporateorstate),and(e)limitingthe
mobilizingeffectofdissent.Thesefiltersformsignificantbarriersagainsttheutilization
ofmainstream,‘established’medianewsstreamsandasillustratedinthebombingofSer
biain1999,theyresultinthesignificantoverrepresentationofviolenceconductedbyUS
adversaries.
Largelyduetotheselimitationsof‘classical’crisismediaeventdatacollectionproto
cols,theadventofsocialmediawasheraldedasanimportantnovelty.Themostcritical
noveltysocialmediabroughtwasits‘disintermediated’nature.Disintermediationisthe
reductionorremovalofintermediariesinasystem(inthiscase,newsandcommunication)
andwithinthecontextofdigitalcommunication,itdefinesthereducedimportanceof
traditionalnewsintermediaries—editors,reporters,censors—thatserveasthemiddle
menbetweenthenewssourcesandnewsconsumers(Sampedroetal.2022).Socialmedia
enablesaviolenteventoramigrationmovementontheground,tobedocumentedby
participantsthemselves,orpasserbycivilians,throughvideos,images,andtextsinreal
time,withoutanyintermediaries,andverylittle(andobscure)contentmoderation,di
rectlyintothenewsfeedsofdigitalusersacrosstheworld(Eriksson2018).Socialmedia
platformshavebothoverlappingandnichecapabilities.Twitterexcelsinshorttext/short
mediaformat,whereasFacebookcanbeusedforgallerystyleextendeddocumentation.
Instagramoffersamorevisualplatformexperience,whereasTikTok’scomparativead
vantageisshortterm‘mood’videos(Jaidka2022).WhileTwitterdatahassofarbeenused
extensivelyforcrisisresearchduetoitsgranularAPIsystem,researchersareincreasingly
experimentingwithothersocialmediaplatformsforcrisiseventdata.
Thedownsidesofsocialmediadatahavepreciselybeenbyproductsofthesamedis
intermediationthatalsoservesasitsstrength.Duetotheabsenceofverificationandfact
checkingfilterssuchasinvestigativereportersordeskeditors,socialmediahasbeenrife
withdisinformationandredundancy(Gohdes2018;Zeitzoff2018).Additionally,ithas
beenfallingpreytoanotherformofavailabilitybias,wherebysocialmediafielddatacan
onlybeproducedwheresmartphones,internetaccessorcellphonetowersarepresent.
Thisisanimportantfactorbecauseinwaranddisasterzoneswhereallthreecanoftenbe
missing,socialmediaascrisiseventdatacanbeverydifficulttoproduce.Anumberof
recentstudiesdemonstratethisclearlyascellphonecoverageandinfrastructurehasa
strongimpactonthevolumeandreliabilityofsocialmediadatacomingfromdifficultto
accessregions(PierskallaandHollenbach2013).
Theremainderofthispaperwillfocusonthefollowingissues:
Advantagesanddisadvantagesofdifferentformsofeventdatathatfeedcurrently
deployedA.I.monitoring/forecastingmodels.
Adiscussionofwhysocialmediaisbecomingmorepopularasfielddatathatisbeing
usedtotrainforecastingmodels.
EthicalconsiderationsinusingsocialmediadatatotrainA.I.conflict/migrationfore
castingclassifiers.
Soc.Sci.2022,11,3954of15
2.The‘AchillesHeel’ofForecasting:DataReliability
Forecastingcriticalandresourceintensiveeventslikecrises,migration,violence,and
warshavelongbeenindemandamonggovernment,military,andinternationalorgani
zationcircles.Buildingpredictionandearlywarningsystemsallowsgovernmentsand
agenciestoprepareforrelief,aid,andlawenforcementplanning,andbuildresilience
againstmajorshocks.Giventhelimitedresourcesofemergencyresponseandreliefinsti
tutions,forecastingcanbeanimportantcostoptimizationprocess,allowingsuchagencies
tobereadyduringtimesensitiveepisodesthatrequiresubstantialresources.
Forecastingisalsoviewedwithwavesofsuspicioninsocialsciences.Mostdirect
criticismisthattheverybasisofforecasting:Collatingpriorinstancesofcasestoinferthe
timingandtypeoffutureevents,hasbeenviewedasunrealistic,orunabletoproperly
capturesocialuncertainty(Dowding2021).Whilemoretransformativeandlargescaleof
suchevents—‘blackswans’(Ahmadetal.2021)—haveindeedbeendifficulttoforecast,
proponentsofthisapproachviewedforecastingasnonethelessusefulinestimatingthe
futuretrajectoryofexistingevents.Thisimpliesthat,foranyforecast,atleasttwotypes
ofinputarerequired:historicaldata,whichcanbeintheformofevent,population,meas
urementorsurvey;andamodel,whichwillformthefoundationoftheestimatorthatwill
determinethewayhistoricaldatawillbeprocessedtoproducetheoutput,namely,pre
diction(Efendietal.2018).
Eventdatahasbeenanimportantinputinviolenceandmigrationforecasting.Given
thefactthatsnowmeltsduringspringtime,insurgencyandcounterinsurgencyopera
tionsrecordanuptickinmountainousareasduringtheseperiods.Monsoonseasonswit
nessheavyrainfall,whichrestrictsthemovementofarmorinmuddyregions,resulting
inreducedoffensivesbystatemilitaries.Massmigrationawayfromtheconflictareasis
expectedwhenthemonsoonseasonends,thesoildries,andheavyarmorcanonceagain
moveintocombatnearpopulationcenters.Heavyartilleryshellingoftenindiscriminately
targetscivilianpopulationcenters,andthereforethemovementofmobileartilleryunits
closertosuchpopulationcenterstendstogeneratemassexodus.Inwinter,heavysnowfall
andfrostrendercombatoperationsinruggedareasmoredifficult,socombatantsprepare
forspringoffensivesinordertopursuetheirtacticalobjectivesaroundtheseareas.As
forcedmigrationoftenhappensaroundtheseseasonalvariances,battlerepertoires,and
violencepatterns,theseindependentvariablesprovideresearcherswithsomedegreeof
confidencethatpastinstancesofeventscanbeharnessedtopredictunrealizedoutcomes.
Itisimportanttoemphasizethatconflictsandmigrationarespatiotemporallymultide
pendenteventsandstandardlinearforecastsoftenfailtoaddressthenuancesofsuch
variance(Christiansenetal.2021).
Givenitshighlevelgranularityanddatavolumeadvantagescomparedtomoretra
ditionalformsofcrisisinformation,socialmediadataisbeingincreasinglyleveragedto
monitor,aswellasforecast,conflictsandmigrationevents.Itisalsobeingusedaspartof
thebroadersuiteofadditionaldatainputstotrainartificialintelligence(A.I.)classifiers
thatarebeingdeployedtomonitorandforecastmigration(Alexanderetal.2022;
Willekens2018;Ningetal.2019;Salah2022;BircanandKorkmaz2021).Givenbiaslimi
tationsoftraditionalmilitaryobserver,warreporter,classicalmedia,andaidworker
basedfielddata,conflictandmigrationforecastingbigdataandA.I.approachesarebe
comingincreasinglyinterestedinharvestingsocialmediaasanadditionallayeroffield
reportseitherinproducingnewdatasets,ortocomplementexistingones.Socialmedia
datacanincreasetherobustnessofpredictionmodelsandforecastingconfidenceasit
providesresearcherswithagreatervolumeoffielddatafromcrisisevents,andifcleaned
forredundancyanddisinformationsufficiently,theycanstrengthenthevalidityofclaims
inferredthroughforecasts(Kaufholdetal.2020;Shanetal.2019).
Inthemigrationandconflictforecastingcontext,A.I.couldbedefinedas“agrowing
resourceofinteractive,autonomous,selflearningagency,whichenablescomputational
artifactstoperformtasksthatotherwisewouldrequirehumanintelligencetobeexecuted
successfully”(TaddeoandFloridi2018).Amongtheseselflearningtasksarethe
Soc.Sci.2022,11,3955of15
methodicalassessmentofalgorithmsanddatainputstoimprovetheirknowledgeofper
formancewithexperience(Flach2012).A.I.basedanalyticsandforecastingmethodshave
beenincreasinglyusedbyinternationalorganizationsandgovernmentstocontroland
managehumanmobility.Thesemethodsincludeidentityrecognition,automatedborder
monitoring,andanalysisofasylumapplications(Beduschi2021).Recently,thesemethods
alsobegantoharvestconflicteventdata(includingsocialmediadata)inordertotrain
A.I.modelstopredictmigration(Molnar2019a).Algorithmicmigrationmonitoringisbe
ingincreasinglyreliedonintheUnitedStates,UnitedKingdom,Canada,andmostEuro
peanUnionmembersstates,andformsthebasisofnextgenerationmigrationtechnology
investmentinmajorinternationalorganizations.
A.I.basedanalysisandforecastingsystemsrequiresignificantvolumesofdata—es
pecially‘bigdata’intheformofnotjustlargevolumeinformation,butalsohighgranu
larity,highvelocity,andhighcomplexityinformation.Whilethereisnotalwaysadirect
relationship,datasizeisoftenanimportantvariableinmoresophisticatedandaccurate
A.I.basedsystems.Tothatend,trainingmachinelearningclassifiersisadatahungryen
deavor,whichrequiresnotjustagreaternumberofevents,butmultiplereportsofthe
sameeventforgreaterrobustnesscheckanddatavalidity.Inaddition,greaterdatainput
canrenderdecisionandoutputchainsmoreefficientinfurtheriterationsbytriangulation,
asresultsaremorerobustresultsincrementallytrainfurtheriterationswithgreaterpreci
sion(Tarasyevetal.2018;Quinnetal.2018).
3.EthicsofSocialMediaasForecastingData
Forecastingsystemsthatarebuiltonmachinelearningprinciplescangrowincreas
inglymoreaccurateintheirpredictionsandcanultimatelymakeinferencesaboutlarge
scalehumanflowsduringconflictsornaturaldisasters.Thiscouldaidintheoptimization
ofreliefaidorrefugeecamppreparations,aswellasproperstaffingofmigrationpro
cessingcenters(Molnar2019b).
However,forecastingmigrationbroadly,inandofitself,maynotalwaysgenerate
relieforientedresults.Mentioningthesegeneralpitfallsareimportantpriortoconnecting
thisargumenttosocialmediadata.Countriescanclosebordersorengageindirectpre
ventativeactiontorenderrefugeepathwaysmoredangerous(PécoudanddeGuchteneire
2006;Vives2017).Rebelgroupsmayuseforecastresultspublishedonlinetocrackdown
onciviliansandcutofftheirescaperoutes(Bruce2001;LarsonandLewis2018).Anim
provedabilitytopredictmigrationmayalsocauselocalpopulationstofleeearlierandin
greaternumbers,giventhefactthatsuchpredictionsareoftensharedthroughsocialme
diaandcanbeseenbythelocalsthroughsmartphones(Dekkeretal.2018).This,inturn,
playsintothehandsofsmugglers,whomayprovidelesserknownpathwaysacrosscoun
triesformigrantsescapingstateprecautions,endangeringrefugees(Sanchez2017).These
prospectsgrowmoreproblematicasinsufficientlyoptimizedsystemsgetdeployedinde
cisionandanalyticsroles,andendupmisidentifying,miscalculating,andmisjudgingref
ugeesandtheiractions.Furthermore,cyberattacksanddataprotectionproblemsmay
leadtomigrationrelateddatasetstobestolen,leaked,andusedbymaliciousactors.Itis
importanttounderlinethattheEuropeanDataProtectionSupervisor(EDPS)initscon
sultationwiththeEuropeanAsylumSupportOffice(EASO)(D(2019)1961C20181083)
hadconcludedthatsocialmediamonitoringtopredictmigrationpatternsisnotinline
withtheEUregulations.Yet,initsdisclosure,theEASOhasrevealedthatithasbeen
activelyharvestingsocialmediadatainordertosupporttheoperationsoftheDepartment
ofOperations,CountryofOriginInformation(COI),InformationandAnalysisUnit
(IAU),andtheDepartmentofAsylumSupport(DAS),amongothers.
Bringingtheargumentclosertothescopeofthisarticle,theuseofsocialmediaas
migrationforecastingdataontheotherhand,bringsinanevenlargersetofethicalcon
siderations.First,thescaleofmigrationdataavailablethroughsocialmediaandmobile
phonesignalsyieldsafalsesenseofsophisticationinanalyses,thatoftenleadsdecision
makerstotreatthesedatasourcesasrepresentative,orformattedsufficientlytotrain
Soc.Sci.2022,11,3956of15
machinelearningclassifiers(StewartandWilson2016).Anumberofimportantstudies
havetimeandagaindemonstratedthat,evenwithitsimmensevolume,socialmediaor
cellphonedatafromforcedmigrationscannotbetreatedasrepresentative,orarenotro
bustenoughintheircurrentformtoproperlytrainclassifiersthatleadtorealworlddeci
sions(SilverandAndrey2019).Thislackofrepresentativenessisamajorcontestationarea
betweenresearchersthatwarnagainstusingmostmediabaseddatasystemsontheirown
toguidedecisions,anddecisionmakersthataimtoleveragethesurfacesophisticationof
A.I.modelsbuiltonflaweddatatomarketthe‘validity’oftheirdecisions(Lachlanetal.
2016).Sincepredictiveandanalyticalmodelsarebasedoninferentialstatistics,theyoper
ateondegreesofprobability,andthusthe‘acceptable’thresholdfordecisionsbecomesa
politicalbenchmark,ratherthantechnical(Schroederetal.2013).Determiningwhich
thresholdis‘sufficient’tomakelifealteringinferencesfromdatausuallybecomesa
murkyprocessthatdecisionmakersfindtoocomplicatedtothinkabout,andgetsdele
gatedtoengineerswhohaveneitherthepoliticalorlegallegitimacy,norappropriate
trainingtomakesuchdecisions(Rahwan2018;Cunneenetal.2019).Giventhecomplexity
oftheprocess,importantdetailssuchasthevariablesandmeasurementsusedinbench
markingbecomesobscuredfrompublicdebateandcreatesaproblematicethicalgap.
Althoughalgorithmicdecisionmakingstructurescanoftenbeusefulwithsufficient
oversightandcontrolmechanisms,itisdifficulttoassertthatbigdatamigrationforecast
ingprotocolshavesufficientsafeguardsinplace,renderingtheverycoreoftheprocess
detachedfrom,andinaccessibletotheverypeopleitdealswith(Zednik2021).Ultimately,
A.I.predictiveandforecastinganalyticsoperateoninferencesbasedonanumberofcor
relations,buttowhatextentthesecorrelationsrepresentcausalmechanismsisnotalways
straightforward(Milanoetal.2021).Giventhedatasizeandspeedofsocialmediabased
information,establishingflexibleandadjustablecausalmechanismsbecomesverydiffi
cult,whichoftenforcesengineerstoofferchoicealternativesfordecisionmakersbased
solelyoncorrelations(BuhmannandFieseler2021).Incaseswheredecisionmakersdonot
havethesufficientbackgroundoradvisorsupporttodigdeeperintothequestioningof
suchcausalmechanisms,aswellastheirdirectionandweight,bigdatamigrationfore
castinganddecisionproceduresbecomeopaqueandpoorlycalibrated.
ThisisinlinewithMolnar(2021),wheretheadventofCOVID19pandemichasgen
eratedgreaterrelianceonbiosurveillance(virustargetingrobots,phonetracking,A.I.en
abledthermalcameras),whichconnectstoTendayiAchiume’s(2021)claimthatsurveil
lancetechnologiesarereinforcingspatialracism.Byusingpandemicrelatedmotives,bor
dercontrolagenciesarerelyingincreasinglyonautomatedforecastingandprediction
modelstokeeprefugeesandmigrantswithinconfinedspaces—camps,borderwallareas,
orsegregatedprocessingcenterswithincities.ThesearewhatJasondeLeon(2015)calls
as‘landofopengraves’,whereauthoritiesconfinerefugeemovementandsettlementinto
dangerousareaswhereselfsustenanceisoftendifficult.Althoughmoretraditionalbor
derprotectiontacticsseeksimilaroutcomes,newertechnologiesrenderautomatedpush
backdecisionslessaccountablebyexportingtheauthorityofdecisionstoambivalent
modelsbuiltonmurkytrainingdata.AsunderlinedinMolnar(2021),theexistingtrack
recordofautomatedtechnologiesonraceandgender,openupsthepathforthedeploy
mentofsimilarmethodsonmigrationsurveillance.
Second,algorithms(asopposedtohumansupervisedinferencemethods)areincreas
inglybeingusedinaselflearningfashionduringcrisisforecastingandanalysis.Most
conflictandmigrationrelatedsocialmediadataarebeingusedtotrainmachinelearning
classifiers,whichthen‘learn’fromthislimiteddataandextractfutureinstancesofdata
collectionprotocols(Wachteretal.2017).Automateddatacollectionandanalysisparam
etersarethusdoublydetachedfromhumanagency,control,andlegalresponsibility:not
onlydoesitperforminanautomatedfashion,butitgrowsincreasinglylessinfluencedby
theoriginalhumancontroloveritthatformeditsprimarylegalandethicalbasis
(Mittelstadtetal.2016).Whilethisproblemremainsintactevenforpoliticallyandlegally
wellcontrolleddataprotectionmechanismsliketheGDPR,formigrationandrefugee
Soc.Sci.2022,11,3957of15
forecasting,itisfurtherproblematicallyhiddenfromvulnerablepopulations,whocannot
seeorchallengetheinferencemechanismthatleadstoaparticulardecisionthatconcerns
theirlives.Sincethespecificprocessesbywhichthecollecteddataareusedtogenerate
likelihoodsofanoutcomearemurky,civiliansareexposednotonlytodecisionsmadeby
inconclusiveevidence,butfromalegalstandpoint,cannotreliablychallengesuchdeci
sionsbecauseparametersthatcreatesuchdecisionsareabstractedfromhumanagencyby
degrees(iterations)ofselftrainingprocessesthatgeneratemachinelearningclassifiers
(Zarsky2013).
Indiscussingtheethicsandlegalityofmigrationanalyticsandforecastingprotocols,
twokeyconceptsformthecoreofthedebate.Thesearetransparencyofthedatacollection,
modeling,anddecisionchain;andtheexplainability/comprehensibilityofthetechnicalities
thatliewithinthislink.Often,theseprotocolsandclassifiers,aswellasthedatastreams
thatareintegratedintothem,areacquiredsecondhandfromsecondorthirdpartysup
pliersaspartof‘integratedsolutions’thatcontainpresetclassifiersanddecisionthresh
olds(Renda2019).Whenusingsuchsecondhandsolutions,stateinstitutionsorinterna
tionalorganizationsriskbeingchallengedontheaccessibilityandcomprehensibility
grounds,sincequiteoftentheseinstitutionsthemselvesdonothavetheengineeringca
pacitytounderstandthedetailedanalyticsprotocolsthemselves(Mittelstadtetal.2016).
Sinceinstitutionshavelowtechnicalcapacitytoalterthesethresholds,theyquiteoften
operatewith‘onesizefitsall’parameters,thatfrequentlyproduceinaccurateforecasting
andanalytics.Whilethereisanongoingdebateonwhetherthisaccessibilitygapisinten
tionalornot,thereisnonethelesssuchalegalandethicalgapthatremainsinplacewith
suchsolutions.
Third,therichestaspectofsocialmediadata—thatitisdiverse,highvolume,and
representingabroadrangeofviewsandperspectivesfromfieldevents—hasthedanger
ofrunningcountertotheveryfundamentalbasisofA.I:itregularlymodifiesandalters
itsdetection,datacollection,processing,andmeasurementapproachesinaselflearning
fashion(VallorandBekey2017).Thisrenderslegalandjudicialoversightandlegitimacy
amovingtarget,asitbringsinacriticalquestion:Whichparameter,threshold,andconfi
denceintervalrangewillbethebasisoflegal,ethicalandpoliticalresponsibility?If,for
example,aparliamentoracourtapprovesthedeploymentofaparticularalgorithmic
structuretobedeployedduringemergencies,migrationcrisesandinmonitoringviolent
conflictatt=0,andifthealgorithmsadjustitsfundamentalentityrecognition,statistical
inference,andconfidenceintervaloptimizationparametersast+1,t+2howwillthis
approvaltranslateoverneweriterationsofthealgorithm?Willthealgorithmbeapproved
‘asis’,orcanitbediscussedanddeliberatedwithinaspecificparameteroscillationrange,
expectingitsfutureselflearningalterations?
Giventhewealthofsocialmediadataanditsdramaticpeaksandplateausduring
keyevents,usingsuchdataasmachinelearningtraininginputwilllikelycausesignificant
episodicchangestohowitcollectsandmodelsemergencyinformation.Thefundamental
questionthusbecomes:areparliaments,governments,andcourtsequippedtodealwith
thequestionofwhethertheirt=0approvalofanalgorithmwillbevalidornotatt+1
andbeyond?Ifthealgorithmdecidestocollectdatabeyonditsinitiallegalandpolitical
confinestooptimizeitsrobustness,wherewillthelegalandelectoralresponsibilitylie?In
politicaldiscourse,thesealgorithmsaregenerallyconstructedas‘semisupervised’sug
gestingthattherewillalwaysbeanengineertooptimize/overseetheseparameter
changes,recentscholarshippositsthatthisisnotalwaysthecaseafteranalgorithmis
legallyandpoliticallyapproved(Shneiderman2016;CastetsRenard2019;ElkinKoren
2020).Lawyers,bureaucrats,andpoliticians,inturn,usuallydonothavetheknowledge
oradvisoryassistancetomakesuchassessmentsbeyondthemediumtermaswell,sug
gestingthatselflearningalgorithmsthatarevettedpriortotheirapprovalmaysteerfur
therawayfromtheirinitialparameters,andbecometoolargetoretrospectivelytrouble
shoot,astimeprogressesanddatasizeandvariabilityincreases.
Soc.Sci.2022,11,3958of15
Fourth,theuseofsocialmediadataasoneoftheinputsofclassifiertrainingrisksa
highdegreeofdatacleaningandwranglingproblems.Often,datastreamsthatharvest
socialmediaplatformsandfeeditintotrainingandanalyticsdashboardshavelessthan
optimaldatacleaningpractices,whichleadtotheinclusionofpotentiallyredundant,mis
leading,andsometimesincoherentinformationintothetrainingchain(Chuetal.2016;
Jainetal.2020;Pavlyshenko2019).Whenaclassifieristrainedbasedon—forexample—a
socialmediadatapoolthatincludesheavybotcampaign,highvolumedisinformation,or
misleadinginformation,itisgoingtogeneratesignificantlymismatchedinferencescom
paredtothegroundreality.Toillustrate,anattackerduringacivilwarcanspreadsocial
mediadisinformationaboutthepresenceof‘potentialjihaditerrorists’amongthefleeing
civilianpopulace,whichmaybepickedupbyautomatedmigrationcrisisdataextraction
protocolsandgenerateaflawedinferenceabouttherefugees,generatingdisproportion
atelypreventativeorharshtreatmentbystatesecurityservices,refugeecampadministra
tions,orevenbythelocalsofthetownsandvillagesalongthemigrationpath.Sincea
propercleaningandformattingofsocialmediadatatakestime,currentA.I.protocolshave
ahighlikelihoodofgeneratingnoisyandflaweddecisionsduringcriseswithtimecon
straints.
Thisincreasestherisksofdiscriminationinbigdatamigrationresearch.WhileRomei
andRuggieri(2014)suggestthatacontrolleddistortionoftrainingdata,introductionof
antidiscriminationinputsintotheclassifiers,postprocessingofclassifiersforasecond
roundcheck,andmodifyingfairnessrelatedparametersinfurtheriterations,thisprotocol
becomestrickywithsocialmediadatathatrendersthesesafeguardsextremelytimecon
suming.Giventhefactthatrefugeecrisesareextensivelysecuritizedintraditionaland
socialmediadata,narrativesofmigrationtendtobehighlysecuritizedaswell(Colombo
2018).Usingsocialmediadirectlyasatextinputalsoinjectsthisdiscursivediscrimination
intoanalyticalandforecastingchains,creatingautomaticbiasesaboutmasshumanmo
bility.Althoughtechnicallysocialmediatextdatacanbepostprocessedtomitigateorat
leastsoftentheseunderlyingdiscursivebiases,itisbothverytimeconsuming,andalso
highlycontextspecificgiventhepeculiaritiesofforeignlanguages,slang,andsatire/sar
casmdynamics(BatrincaandTreleaven2015).WhiletheseparametersarewellsetinEng
lish,muchmoreworkisrequiredtoconductthiscleaningworkinotherlanguages—es
peciallyinlanguagesinwhichdatacleaningengineersarenotproficient.Thisissuenot
onlycreatesageneralizedunfairnesssinceanalgorithmactivelybeingtrainedonsocial
mediadatawillcreateimmediatebiasesagainstrefugees,butitalsoproducesasecond
setofbiasandunfairnessduetothealgorithm’stextcorpusvariancesacrossdifferent
languages.
Fifth,andconnectedtothethird,thesheersizeoftheA.I.forecastingandanalytics
datainfrastructures,rendersreverseengineeringofmistakesverydifficultandtimecon
suming(Sejnowski2020).Oncesuchdecisionsleadtoflawedoutcomes—especiallythat
resultinharmagainstrefugees—returningtotrainingdataandtryingtounderstand
whichcomponentcausedtheflawedoutcomebecomesadauntingtaskwithsocialmedia
data.Sincesuchtroubleshootingandfixingworkrequiresextensivestaffing,time,and
financialresources,decidingtoengageinsuchenterpriseitselfbecomesapoliticaland
bureaucraticdecision(Agrawaletal.2018).Evenifsuchdecisionistaken,tryingtofind
theneedleintheproverbialdatahaystacknecessitatesredirectingresourcesfromdaily
monitoring,analytics,anddecisionproducingprotocols.Suchdiminishingreturnsren
derstheentiretyofthetroubleshootingprocessdeterringandpotentiallyunsustainable,
causinganalyticsandpredictionteamstogoaheadwithinsufficientlytrainedclassifiers
thatcontinuetoproduceflawedresults—andworse:learnfromtheseflawedpriors
(CoglianeseandLehr2016).Whensuchcasespersist,itbecomesnearlyimpossibletofind
whichdatastreamcausedtheflawedinterpretation,whichengineeringdepartmentwas
responsibleforit,andmostimportantly:whoistobeheldlegallyaccountableforsuch
flawedinterpretations.Thiscreatesnotjustanethicalproblemwherebythesourceof
harmagainstmigrants/refugeescannotbeidentifiedreliably,butitalsoposesasignificant
Soc.Sci.2022,11,3959of15
legalproblemwhereresponsibilitycannotbetraced,andbindingrevisionscannotbeap
pliedtoproperindividuals,teams,andinferencemechanisms.Notonlythatthismecha
nismputspeopleatriskintoharm,butitalsocreatesadigitallyauthoritarianconstruct
whichinvisiblyshutsthedoortocomplaints,informationrequestsandlegalreparation
options.
Thereareotherproblemswithbigdatamigrationandviolenceforecastingthatis
beyondthescopeofthispaper,whichfocusesonusingsocialmediaasclassifiertraining
data.Theseproblemsinclude(butarenotlimitedto)problemswiththeanalysisandmod
elingrobustnessaspectofalgorithms,i.e.,thefactthatthestatisticalparametersthatgen
eratedecisionsorhelpthealgorithmlearnsecondarytasksmaynotalwaysbecalibrated
wellenough,oraccurate.Sincethereisaknowhowandformationaldisconnectbetween
engineerswhocalibratethesemodelsandthedecisionmakerswhoactonthefindingsof
suchmodels,inferencemodelaccuracybecomeslessofamathematicalortechnicalen
deavorandmoreofapoliticalandethicaloneinwhichconfidenceintervals,variable
weights,directions/strengthofcorrelations,andhowwithinsampleevidenceisusedto
generateoutofsampleforecastsbecomebiased,andpotentiallydiscriminatory.
4.AvenuesforEthicalSocialMediaDataUseinMigrationandViolenceForecasting
Theaboveaccountmaysuggestthatthenumberofissuesthatcurrentlyobscurethe
useofsocialmediadataasareliableemergencyinformationfortrainingA.I.basedsys
temsandassistingintheirinferenceandforecastprotocolsaretooinsurmountable.This
doesnotimplythatsocialmediadatacannotbeusedforsuchpurposes,butanumberof
broaderethicalandlegaldecisionshavetobemadebeforemoretechnical/programming
optimizationworkcanbeundertaken.
Oneofthefirstpathwaystorendersocialmediadata‘moreuseable’inalgorithmic
formatwouldbetofindwaystoanonymizeandrendersuchdatauntraceablebacktoa
vulnerableperson,oragroup.Whilesomeanonymizationapproachesareavailablefor
accountswhosesocialmediapostsarepartoftrainingclassifiers,suchanonymizationbe
comesharderinvideoandimagedata,wherethecontentiscrucialforinference,andcan
oftencontaininformationaboutactualidentityofindividualsandgroupsthatareboth
vulnerableandindanger(Beigietal.2018;TownsendandWallace2017;Zhangetal.
2018).Anewanonymizationprotocolisnecessarytocreateadisconnectbetweenactual
identitiesandtheinferenceproducedasaresultofit,aswellasfurtherstudiesthatex
plorehowlackofanonymitycreatesunintendeddiscriminatoryinferencesbyalgorithms.
Thislogic,however,needstobedissectedwhenconsideringindividualversuscol
lectiveharm.Harmanddiscriminationtowardsone,orasmallnumberofmigrantsfits
intoadifferentanonymizationdebatecomparedtoharmanddiscriminationtowardsa
largerflowofmigration.Intheformercontext,anonymizationcanprovideaviablesolu
tion(asitisalreadydeployedassuchbytheEASOandUNHCR),althoughitstartsto
becomelessrelevanttothemigrantprotectiondebatewithinthecontextoflargescale
flows.Thispointbecomesalargerpoliticalandideologicalissue(antiimmigration),as
opposedtoatechnicalandhumanrightsrelateddebatebecausediscriminationofoneor
agroupofmigrantswithinthebasisofethnicity,religionandbiometricssitsinadifferent
corneroftheanonymizationdebatecomparedtoextractingdatafromamassexodus.In
thelattercontext,identityrelatedvariablesformasecondaryconsiderationinlightofthe
moreimportantfirstconsideration:wherethemigrationflowhasoriginated,whereitis
headedto,andwhichpathwaysittransitsthrough.Thedifferencebetweenthetwoalso
containsthedifferencebetweenflawedstateresponses(a)bydenyingtherighttoclaim
asylumbyawholesalepreventionormigration,and(b)bytargetingone,oragroupof
peopleinadiscriminatoryfashionusingdatacollectedthroughsocialmedia.
Thesecondpathistobroadenthejudicial,bureaucratic,andparliamentarydiscus
sionsontheresponsibilityofA.I.systems.Whilethisdecisionwilllikelydifferacross
countriesandpoliticalsystems,countrieswillmostlikelyoptforasharedresponsibility
infrastructureforA.I.baseddecisionerrors,distributinglegalandpoliticalerrormargins
Soc.Sci.2022,11,39510of15
betweenengineers,systemmanagersandA.I.protocols.Inaflaweddecisionthatcreates
asignificantlevelofhumansuffering,thelegalattributionchainwilllikelyfollowthrough
(1)theinstitution,(2)thedirector,(3)thesubmanager,(4)thechiefengineer,(5)thejunior
engineerthatruns,operates,andmaintainsthecodestructureofthealgorithm.Which
lineofthehierarchyistoblamewilllikelybeapoliticaldecisionthatdifferentcountries
willtakedifferentlybasedonpowerdynamicsbetweenvariousrelatedinstitutions,and
withintheinstitutionthathasundertakentheflaweddecision(Shah2018).Inrecentyears,
anewdebateemergedonthe‘moralagency’ofalgorithms—i.e.,whethernonhumanac
torsofanalgorithmsuchassequence,selection,anditerationcanbeheldresponsibleina
courtoflaw(Véliz2021;Cunneenetal.2019).ThisdebateriskswhatWachteretal.define
as‘deresponsibilizationofhumanactors’,whichinsimpletermsisto‘hidebehindthe
computer’andexportingtheresponsibilityofflaweddecisionstounprosecutableagents
(Löschetal.2017;Kirkpatrick2016).ThistrendrunsclosertoHannahArendt’sobserva
tionofhowinhumanandstructurallyhostileactionscandispassionatelybeundertaken
bybureaucraticnetworksandactorsthatarefollowingrulesandprotocolsthatcreate
automatediterationsofsufferingwithoutagencyontheirpart.
Thethirdpathfollowsthe‘righttoreasonableinferences’debateanditsintroduction
intothedebateontheuseofsocialmediadataforbigdatamigrationandviolencefore
castingpurposes(WachterandMittelstadt2019;Veroneseetal.2019).InGDPRcontext,
thisrightpositsthatindividualswhosedataisharvestedforinferenceandprediction
tasksthatconcerntheirlives(‘highriskinferences’),havetherighttoaskforjustification
bytheappropriatedatamanagertodisclosewhethersaidinferencehasbeenproduced
throughjustifiableandexplainableprotocols.Whenusedaspartofadecisionoraction
thatconcernsmigrantsandrefugees,thealgorithmmanagersandengineershavetopro
duceapublicexplanationthatoutlinestherationaleforusingthedatasettoproducesaid
output,whethertheformofinferenceandmodelingistrulyappropriateforsaiddecision,
andwhethertheoutputissufficientlyrobustandstatisticallymeaningfulwithinthelegal
andlegislativeparameterssetbythatcountry’slawsandregulations.
Toconclude,highvelocitysocialmediadatastreamscontainsignificantpotentialfor
emergencymonitoringandforecasting;however,manycontemporaryexamplesofsuch
attemptsremainethicallyandlegallyproblematic.Thisproblemisnotlimitedtotheuse
ofsocialmediadata,butconcernsseveralothercomponentsofbigdatamigrationand
conflictforecastingpractices,suchasflawedinferences,opaquedecisions,andpolicyen
gineeringmismatchbetweenthresholdsthatgenerateprofiling,monitoring,andscenario
building.Thispaperhasarguedthattheuseofsocialmediadatacanexacerbateexisting
problemswithexplainability,interpretability,andtransparencyofA.I.forecastingand
decisionsystemsthatdealwithviolenceandforcedmigration.Thispaperhasarguedthat
theuseofsocialmediadatacanexacerbateexistingproblemswithexplainability,inter
pretability,andtransparencyofA.I.forecastinganddecisionsystemsthatdealwithvio
lenceandforcedmigration.Largedatainjectionthroughtheinclusionofsocialmediadata
intoselflearningA.I.monitoringsystemsgenerateafalsesenseofsophisticationthat
forcesdecisionmakerstodisregardthevalidity,representativeness,androbustnessof
suchsystems.Sincetheyareselflearningsystems,learningclassifiersthatproducedeci
sionoptionsbecomedetachedfromthepeopletheyismonitoring,becomeunreachable
fromthemigrants’standpoint,andcanpotentiallyendangervulnerablepopulationsun
derduress.
Third,sincesocialmediadatacontainsrapidlychangingstances,discourses,andnar
ratives,A.I.systemsthatusesuchdataastraininginputriskssteeringoutsidetheconfines
oflegalandparliamentaryoversight.Whilecourtsandlegislatorsmayethicallyandle
gallyapproveaselflearningalgorithminitially,weeksandmonthsafterthesealgorithms
startlearningontheirown,theyhaveagreatlikelihoodofstrayingintoanoversight‘gray
zone’.FurtherproblemsarisefromthefactthatmostA.I.migrationandconflictmonitor
ingsystemshavesuboptimaldatacleaningpractices.Thisrendersmodelsvulnerableto
botcampaignsanddisinformationandrisksbiasanddiscriminationagainstmigrant
Soc.Sci.2022,11,39511of15
populations.Finally,althoughsocialmediadatabringsimmensedatadiversitytohelp
teachmachinelearningclassifiers,itssheerscalerendersretrospectivetroubleshooting
verydifficult,especiallywhenthesesystemsmakeacostlyflaweddecisionandreverse
engineeringbecomesessentialtofindoutwhichdatastreamscausedtheproblem.Alt
houghitisstillearly,oncetheseissuesarediscussedandpotentiallyresolved,socialme
diadatacanbeusedethicallyandcleanlyinfutureA.I.migrationandconflictmonitoring
tasks.
AnimportantstepinthesedirectionsweretakenbytheEuropeanDataProtection
Supervisor(EDPS)initsconsultationwiththeEuropeanAsylumSupportOffice(EASO)
(D(2019)1961C20181083),wheresocialmediamonitoringformigrationpredictionwas
deemedillegalwithintheparametersoftheEUlaw.However,itisimportanttokeepin
mindthattheEDPSconclusionisonlylimitedtoEASO’soperations,andassertsthat
trackingmigrationwithinthecontextofmigrantsmugglingandhumantraffickingmon
itoringarebeyondthescopeofEASO,andtherefore,leavesagrayareaforotherEUagen
ciesthatmaybemonitoringsocialmediaformigrationonthepretextthatsuchmonitor
ingisdoneto‘preventcrime’.
Still,however,thereissignificantdiscussionspaceastowhatextentEDPSjudgement
willhaveabindingeffectonotherEUnations’dailyborderprotectionpractices,orform
amodeltoemulatefortheborderprotectionagenciesoftherestoftheworld.Asthe
competitiontotrainmoreaccuratemodels—eitherinconflictormigrationprediction—
socialmediadatawillremainacontroversial,yetanincreasinglypopularchoice,asthe
debateoverhowbesttouseitethicallyandtomeetstrategicobjectiveswilllikelycontinue
inthefollowingyears.
Funding:ThisresearchwaspartiallyfundedbytheScientificandTechnologicalResearchCouncil
ofTurkey,ARDEB1001Program,GrantNumber120K986;andTheScienceAcademySocietyof
Turkey:2021BAGEPProgram.
InstitutionalReviewBoardStatement:Notapplicable.
InformedConsentStatement:Notapplicable.
DataAvailabilityStatement:Notapplicable.
ConflictsofInterest:Theauthordeclaresnoconflictofinterest
References
(Achiume2021)Achiume,E.Tendayi.2021.DigitalRacialBorders.AmericanJournalofInternationalLaw115:333–38.
(Agrawaletal.2018)Agrawal,Ajay,JoshuaGans,andAviGoldfarb.2018.PredictionMachines:TheSimpleEconomicsofArtificial
Intelligence.Boston:HarvardBusinessReviewPress.
(Ahmadetal.2021)Ahmad,Wasim,AliM.Kutan,andSmarthGupta.2021.BlackSwanEventsandCOVID19Outbreak:Sector
LevelEvidencefromtheUS,UK,andEuropeanStockMarkets.InternationalReviewofEconomics&Finance75:546–57.
https://doi.org/10.1016/j.iref.2021.04.007.
(Alexanderetal.2022)Alexander,Monica,KivanPolimis,andEmilioZagheni.2022.CombiningSocialMediaandSurveyDatato
NowcastMigrantStocksintheUnitedStates.PopulationResearchandPolicyReview41:1–28.https://doi.org/10.1007/s11113020
095993.
(AshandObradovich2020)Ash,Konstantin,andNickObradovich.2020.ClimaticStress,InternalMigration,andSyrianCivilWar
Onset.JournalofConflictResolution64:3–31.https://doi.org/10.1177/0022002719864140.
(BanerjeeandMacKay2020)Banerjee,Kiran,andJosephMacKay.2020.CommunitiesofPractice,ImpressionManagement,and
GreatPowerStatus:MilitaryObserversintheRussoJapaneseWar.EuropeanJournalofInternationalSecurity5:274–93.
https://doi.org/10.1017/eis.2020.11.
(BatrincaandTreleaven2015)Batrinca,Bogdan,andPhilipC.Treleaven.2015.SocialMediaAnalytics:ASurveyofTechniques,Tools
andPlatforms.AI&SOCIETY30:89–116.https://doi.org/10.1007/s0014601405494.
(BaumandZhukov2015)Baum,MatthewA.,andYuriM.Zhukov.2015.FilteringRevolution:ReportingBiasinInternationalNews
paperCoverageoftheLibyanCivilWar.JournalofPeaceResearch52:384–400.https://doi.org/10.1177/0022343314554791.
(BasuandPearlman2017)Basu,Sunkaya,andSarahPearlman.2017.Violenceandmigration:evidencefromMexico’sdrugwar.IZA
JournalofDevelopmentandMigration7:1–29.
(Beduschi2021)Beduschi,Ana.2021.InternationalMigrationManagementintheAgeofArtificialIntelligence.MigrationStudies9:
576–96.https://doi.org/10.1093/migration/mnaa003.
Soc.Sci.2022,11,39512of15
(Beigietal.2018)Beigi,Ghazaleh,KaiShu,YanchaoZhang,andHuanLiu.2018.SecuringSocialMediaUserData:AnAdversarial
Approach.InProceedingsofthe29thonHypertextandSocialMedia.NewYork:AssociationforComputingMachinery,pp.165–
73.https://doi.org/10.1145/3209542.3209552.
(BircanandKorkmaz2021)Bircan,Tuba,andEmreErenKorkmaz.2021.BigDataforWhoseSake?GoverningMigrationthrough
ArtificialIntelligence.HumanitiesandSocialSciencesCommunications8:1–5.https://doi.org/10.1057/s4159902100910x.
(BlairandSambanis2020)Blair,RobertA.,andNicholasSambanis.2020.ForecastingCivilWars:TheoryandStructureinanAgeof
‘BigData’andMachineLearning.JournalofConflictResolution64:1885–915.https://doi.org/10.1177/0022002720918923.
(Bruce2001)Bruce,Beverlee.2001.TowardMediatingtheImpactofForcedMigrationandDisplacementAmongChildrenAffected
byArmedConflict.JournalofInternationalAffairs55:35–57.
(BuhmannandFieseler2021)Buhmann,Alexander,andChristianFieseler.2021.TowardsaDeliberativeFrameworkforResponsible
InnovationinArtificialIntelligence.TechnologyinSociety64:101475.https://doi.org/10.1016/j.techsoc.2020.101475.
(Bunceetal.2019)Bunce,Mel,MartinScott,andKateWright.2019.HumanitarianJournalism.InOxfordResearchEncyclopediaof
Communication.Oxford:OxfordResearchEncyclopediaofCommunication.https://doi.org/10.1093/acre
fore/9780190228613.013.821.
(Burkeetal.2015)Burke,Marshall,SolomonM.Hsiang,andEdwardMiguel.2015.ClimateandConflict.AnnualReviewofEconomics
7:577–617.https://doi.org/10.1146/annureveconomics080614115430.
(Carammiaetal.2022)Carammia,Marcello,StefanoMariaIacus,andTeddyWilkin.2022.ForecastingAsylumRelatedMigration
FlowswithMachineLearningandDataatScale.ScientificReports12:1457.https://doi.org/10.1038/s41598022052418.
(CastetsRenard2019)CastetsRenard,Celine.2019.AccountabilityofAlgorithmsintheGDPRandbeyond:AEuropeanLegal
FrameworkonAutomatedDecisionMaking.FordhamIntellectualProperty,Media&EntertainmentLawJournal30:91.
(Chojnackietal.2012)Chojnacki,Sven,ChristianIckler,MichaelSpies,andJohnWiesel.2012.Eventdataonarmedconflictand
security:Newperspectives,oldchallenges,andsomesolutions.InternationalInteractions38:382–401.
(ChomskyandHerman2002).Chomsky,Noam,andEdwardHerman.2002.Apropagandamodel.InManufacturingConsent:The
PoliticalEconomyoftheMassMedia,2nded.NewYork:PantheonBooks.
(Christiansenetal.2021)Christiansen,Rune,MatthiasBaumann,TobiasKuemmerle,MiguelD.Mahecha,andJonasPeters.2021.
TowardCausalInferenceforSpatioTemporalData:ConflictandForestLossinColombia.JournaloftheAmericanStatistical
Association117:591–601.https://doi.org/10.1080/01621459.2021.2013241.
(Chuetal.2016)Chu,Xu,IhabF.Ilyas,SanjayKrishnan,andJiannanWang.2016.DataCleaning:OverviewandEmergingChallenges.
InProceedingsofthe2016InternationalConferenceonManagementofData.NewYork:AssociationforComputingMachinery,pp.
2201–6.https://doi.org/10.1145/2882903.2912574.
(CoglianeseandLehr2016)Coglianese,Cary,andDavidLehr.2016.RegulatingbyRobot:AdministrativeDecisionMakinginthe
MachineLearningEra.GeorgetownLawJournal105:1147.
(Colombo2018)Colombo,Monica.2018.TheRepresentationofthe‘EuropeanRefugeeCrisis’inItaly:Domopolitics,Securitization,
andHumanitarianCommunicationinPoliticalandMediaDiscourses.JournalofImmigrant&RefugeeStudies16:161–78.
https://doi.org/10.1080/15562948.2017.1317896.
(ConteandMigali2019)Conte,Alessandra,andSilviaMigali.2019.Theroleofconflictandorganizedviolenceininternationalforced
migration.DemographicResearch41:393–424.
(Crush2013)Crush,Jonathan.2013.Linkingfoodsecurity,migrationanddevelopment.InternationalMigration51:61–75.
(Cunneenetal.2019)Cunneen,Martin,MartinMullins,FinbarrMurphy,andSeánGaines.2019.ArtificialDrivingIntelligenceand
MoralAgency:ExaminingtheDecisionOntologyofUnavoidableRoadTrafficAccidentsthroughthePrismoftheTrolleyDi
lemma.AppliedArtificialIntelligence33:267–93.https://doi.org/10.1080/08839514.2018.1560124.
(deLeon2015)deLeon,Jason.2015.TheLandofOpenGraves:LivingandDyingontheMigrantTrail.Oakland:UniversityofCalifornia
Press.
(Dekkeretal.2018)Dekker,Rianne,GodfriedEngbersen,JeanineKlaver,andHannaVonk.2018.SmartRefugees:HowSyrianAsy
lumMigrantsUseSocialMediaInformationinMigrationDecisionMaking.SocialMedia+Society4:2056305118764439.
https://doi.org/10.1177/2056305118764439.
(DemarestandLanger2018).Demarest,Leila,andArnimLanger.2018.ThestudyofviolenceandsocialunrestinAfrica:acompar
ativeanalysisofthreeconflicteventdatasets.AfricanAffairs117:310–25.
(Dowding2021)Dowding,Keith.2021.WhyForecast?TheValueofForecastingtoPoliticalScience.PS:PoliticalScience&Politics54:
104–6.https://doi.org/10.1017/S104909652000133X.
(Eck2012)Eck,Kristine.2012.InDataWeTrust?AComparisonofUCDPGEDandACLEDConflictEventsDatasets.Cooperationand
Conflict47:124–41.https://doi.org/10.1177/0010836711434463.
(Efendietal.2018)Efendi,Riswan,NureizeArbaiy,andMustafaMatDeris.2018.ANewProcedureinStockMarketForecasting
BasedonFuzzyRandomAutoRegressionTimeSeriesModel.InformationSciences441:113–32.
https://doi.org/10.1016/j.ins.2018.02.016.
(ElkinKoren2020)ElkinKoren,Niva.2020.ContestingAlgorithms:RestoringthePublicInterestinContentFilteringbyArtificial
Intelligence.BigData&Society7:2053951720932296.https://doi.org/10.1177/2053951720932296.
(Epstein2010)Epstein,Andrew.2010.Educationrefugeesandthespatialpoliticsofchildhoodvulnerability.ChildhoodinAfrica2:
16–25.
Soc.Sci.2022,11,39513of15
(Eriksson2018)Eriksson,Mats.2018.LessonsforCrisisCommunicationonSocialMedia:ASystematicReviewofWhatResearch
TellsthePractice.InternationalJournalofStrategicCommunication12:526–51.https://doi.org/10.1080/1553118X.2018.1510405.
(Flach2012)Flach,Peter.2012.MachineLearning,1sted.Cambridge:CambridgeUniversityPress.
(Gohdes2018)Gohdes,AnitaR.2018.StudyingtheInternetandViolentConflict.ConflictManagementandPeaceScience35:89–106.
https://doi.org/10.1177/0738894217733878.
(Humphrey2013)Humphrey,Michael.2013.Migration,securityandinsecurity.JournalofInterculturalStudies34:178–95.
(Jaidka2022)Jaidka,Kokil.2022.CrossPlatform‐ andSubgroupDifferencesintheWellBeingEffectsofTwitter,Instagram,and
FacebookintheUnitedStates.ScientificReports12:3271.https://doi.org/10.1038/s4159802207219y.
(Jainetal.2020)Jain,Abhinav,HimaPatel,LokeshNagalapatti,NitinGupta,SameepMehta,ShanmukhaGuttula,ShashankMu
jumdar,ShaziaAfzal,RuhiSharmaMittal,andVitobhaMunigala.2020.OverviewandImportanceofDataQualityforMachine
LearningTasks.InProceedingsofthe26thACMSIGKDDInternationalConferenceonKnowledgeDiscovery&DataMining.New
York:AssociationforComputingMachinery,pp.3561–62.https://doi.org/10.1145/3394486.3406477.
(Kaufholdetal.2020)Kaufhold,MarcAndré,NicolaRupp,ChristianReuter,andMatthiasHabdank.2020.MitigatingInformation
OverloadinSocialMediaduringConflictsandCrises:DesignandEvaluationofaCrossPlatformAlertingSystem.Behaviour&
InformationTechnology39:319–42.https://doi.org/10.1080/0144929X.2019.1620334.
(Kirkpatrick2016)Kirkpatrick,Keith.2016.BattlingAlgorithmicBias:HowDoWeEnsureAlgorithmsTreatUsFairly?Communica
tionsoftheACM59:16–17.https://doi.org/10.1145/2983270.
(Knightley2002)Knightley,Phillip.2002.Journalism,ConflictandWar:AnIntroduction.JournalismStudies3:167–71.
https://doi.org/10.1080/14616700220129946.
(Lachlanetal.2016)Lachlan,KennethA.,PatricR.Spence,XialingLin,KristyNajarian,andMariaDelGreco.2016.SocialMediaand
CrisisManagement:CERC,SearchStrategies,andTwitterContent.ComputersinHumanBehavior54:647–52.
https://doi.org/10.1016/j.chb.2015.05.027.
(LarsonandLewis2018)Larson,JenniferM.,andJanetI.Lewis.2018.Rumors,KinshipNetworks,andRebelGroupFormation.
InternationalOrganization72:871–903.https://doi.org/10.1017/S0020818318000243.
(LeeandMaslog2005)Lee,SeowTing,andCrispinC.Maslog.2005.WarorPeaceJournalism?AsianNewspaperCoverageofCon
flicts.JournalofCommunication55:311–29.https://doi.org/10.1111/j.14602466.2005.tb02674.x.
(Lichtenheld2020)Lichtenheld,AdamG.2020.ExplainingPopulationDisplacementStrategiesinCivilWars:ACrossNationalAnal
ysis.InternationalOrganization74:253–94.https://doi.org/10.1017/S0020818320000089.
(Löschetal.2017)Lösch,Andreas,ReinhardHeil,andChristophSchneider.2017.ResponsibilizationthroughVisions.Journalof
ResponsibleInnovation4:138–56.https://doi.org/10.1080/23299460.2017.1360717.
(LozanoGraciaetal.2010)LozanoGracia,Nancy,GianfrancoPiras,AnaMariaIbáñez,andGeoffreyJ.D.Hewings.2010.TheJour
neytoSafety:ConflictDrivenMigrationFlowsinColombia.InternationalRegionalScienceReview33:157–80.
https://doi.org/10.1177/0160017609336998.
(Lubecketal.2003)Lubeck,M.,D.Geppert,andK.Nienartowicz.2003.AnOverviewofaLargeScaleDataMigration.Paperpre
sentedat20thIEEE/11thNASAGoddardConferenceonMassStorageSystemsandTechnologies,2003(MSST2003),SanDiego,
CA,USA,April7–10,pp.49–55.https://doi.org/10.1109/MASS.2003.1194835.
(Milanoetal.2021)Milano,Silvia,BrentMittelstadt,SandraWachter,andChristopherRussell.2021.EpistemicFragmentationPoses
aThreattotheGovernanceofOnlineTargeting.NatureMachineIntelligence3:466–72.https://doi.org/10.1038/s4225602100358
3.
(Mittelstadtetal.2016)Mittelstadt,BrentDaniel,PatrickAllo,MariarosariaTaddeo,SandraWachter,andLucianoFloridi.2016.The
EthicsofAlgorithms:MappingtheDebate.BigData&Society3:2053951716679679.https://doi.org/10.1177/2053951716679679.
(Molnar2019a)Molnar,Petra.2019a.NewTechnologiesinMigration:HumanRightsImpacts.ForcedMigrationReview61:7–9.
(Molnar2019b)Molnar,Petra.2019b.TechnologyontheMargins:AIandGlobalMigrationManagementfromaHumanRights
Perspective.CambridgeInternationalLawJournal8:305–30.https://doi.org/10.4337/cilj.2019.02.07.
(Molnar2021)Molnar,Petra.2021.Chapter10:Robotsandrefugees:thehumanrightsimpactsofartificialintelligenceandautomated
decisionmakinginmigration.InResearchHandbookonInternationalMigrationandDigitalTechnology.Cheltenham:EdwardElgar
Publishing.
(MullenandKlaehn2010)Mullen,Andrew,andJefferyKlaehn.2010.TheHerman–Chomskypropagandamodel:Acriticalapproach
toanalysingmassmediabehaviour.SociologyCompass4:215–29.
(Nam2006)Nam,Taehyun.2006.Whatyouusematters:Codingprotestdata.PS:PoliticalScience&Politics39:281–87.
(Ningetal.2019)Ning,Yue,LiangZhao,FengChen,ChangTienLu,andHuzefaRangwala.2019.SpatioTemporalEventForecast
ingandPrecursorIdentification.InProceedingsofthe25thACMSIGKDDInternationalConferenceonKnowledgeDiscovery&Data
Mining.NewYork:AssociationforComputingMachinery,pp.3237–38.https://doi.org/10.1145/3292500.3332291.
(Pavlyshenko2019)Pavlyshenko,BohdanM.2019.MachineLearningModelsforSalesTimeSeriesForecasting.Data4:15.
https://doi.org/10.3390/data4010015.
(PécoudanddeGuchteneire2006)Pécoud,Antoine,andPauldeGuchteneire.2006.InternationalMigration,BorderControlsand
HumanRights:AssessingtheRelevanceofaRighttoMobility.JournalofBorderlandsStudies21:69–86.
https://doi.org/10.1080/08865655.2006.9695652.
Soc.Sci.2022,11,39514of15
(PierskallaandHollenbach2013)Pierskalla,JanH.,andFlorianM.Hollenbach.2013.TechnologyandCollectiveAction:TheEffect
ofCellPhoneCoverageonPoliticalViolenceinAfrica.AmericanPoliticalScienceReview107:207–24.
https://doi.org/10.1017/S0003055413000075.
(Quinnetal.2018)Quinn,JohnA.,MargueriteM.Nyhan,CeliaNavarro,DavideColuccia,LarsBromley,andMiguelLuengoOroz.
2018.HumanitarianApplicationsofMachineLearningwithRemoteSensingData:ReviewandCaseStudyinRefugeeSettle
mentMapping.PhilosophicalTransactionsoftheRoyalSocietyA:Mathematical,PhysicalandEngineeringSciences376:20170363.
https://doi.org/10.1098/rsta.2017.0363.
(Rahwan2018)Rahwan,Iyad.2018.Societyintheloop:Programmingthealgorithmicsocialcontract.EthicsAndInformationTechnol
ogy20:5–14.
(Rajabalietal.2009)Rajabali,Alefiyah,OmerMoin,AmnaS.Ansari,MohammadR.Khanani,andSyedH.Ali.2009.Communicable
diseaseamongdisplacedAfghans:refugewithoutshelter.NatureReviewsMicrobiology7:609–14.https://doi.org/10.1038/nrmi
cro2176
(Ravi2005)Ravi,Narasimhan.2005.LookingbeyondFlawedJournalism:HowNationalInterests,Patriotism,andCulturalValues
ShapedtheCoverageoftheIraqWar.HarvardInternationalJournalofPress/Politics10:45–62.
https://doi.org/10.1177/1081180X05275765.
(Renda2019)Renda,Andrea.2019.ArtificialIntelligence.Ethics,GovernanceandPolicyChallenges.CEPSCentreforEuropeanPolicy
Studies.Availableonline:https://www.ceeol.com/search/bookdetail?id=829907(accessedon19May2022).
(RomeiandRuggieri2014)Romei,Andrea,andSalvatoreRuggieri.2014.AMultidisciplinarySurveyonDiscriminationAnalysis.
TheKnowledgeEngineeringReview29:582–638.https://doi.org/10.1017/S0269888913000039.
(Salah2022)Salah,AlbertAli.2022.CanBigDataDeliverItsPromisesinMigrationResearch?InternationalMigration60:252–55.
https://doi.org/10.1111/imig.12984.
(Salehyan2007)Salehyan,Idean.2007.RefugeesandtheStudyofCivilWar.CivilWars9:127–41.
https://doi.org/10.1080/13698240701207286.
(Sampedroetal.2022)Sampedro,Víctor,F.JavierLópezFerrández,andPatriciaHidalgo.2022.DigitalDisintermediation,Technical
andNationalSovereignty: