Becke , Ma hias J. e al.
Resea ch Repo
Decoding An isemi ism: An AI-d i en S udy on Ha e
Speech and Image y Online
Discou se Repo , No. 6
P o ided in Coope a ion wi h:
WZB Be lin Social Science Cen e
Sugges ed Ci a ion: Becke , Ma hias J. e al. (2024) : Decoding An isemi ism: An AI-d i en S udy on
Ha e Speech and Image y Online, Discou se Repo , No. 6, Technische Uni e si ä Be lin, Zen um
ü An isemi ismus o schung, Be lin,
h ps://doi.o g/10.14279/deposi once-19748
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/312153
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by/4.0/
Feb ua y 2024
Discou se Repo | 6
Pilo P ojec
Decoding An isemi ism:
An AI-d i en S udy on Ha e
Speech and Image y Online
Cen e o Resea ch on An isemi ism,
Technical Uni e si y Be lin
P incipal In es iga o :
D Ma hias J. Becke
Cen e o Resea ch on An isemi ism, Technical Uni e si y (TU) Be lin
Co-In es iga o :
P o . Helena Mihalje ić
Da a Science, Uni e si y o Applied Sciences (HTW) Be lin
Resea ch Team TU, HTW & WZB Be lin:
D Lau a Ascone
D Ma hew Bol on
Alexis Chapelan
Pia Haupel sho e
Alexa K ugel
S e an Munnes
Ka olina Placzyn a
Milena Pus e
Ma cus Scheibe
Vic o Tschiskale
P ojec Manage :
P o . U a Jensen
Cen e o Resea ch on An isemi ism, TU Be lin
In collabo a ion wi h Ca di Uni e si y’s Ha eLab,
Uni e si y o Michigan—School o In o ma ion, and he
Depa men o Digi al Humani ies a King’s College London
Funded by he Al ed Landecke Founda ion
P ojec Managemen :
D Lau a Ascone
Sec e a y:
And ea Rellin
P ojec A ilia es:
Ve onika Bundzíko á (Masa yk Uni e si y, B no, Czechia)
Iael Ku jan (Fede al Uni e si y o San a Ca a ina, B azil)
Mohamed Salhi (Goe he Uni e si y F ank u , Ge many)
Technical Uni e si y Be lin
Cen e o Resea ch on An isemi ism (Z A)
Kaise in-Augus a-Allee 104 –106
10553 Be lin
Con ac : [email p o ec ed]
Web: decoding-an isemi ism.eu
Ci a ion ecommenda ion:
Becke , Ma hias J.; Ascone, Lau a; Bol on, Ma hew; Bundzíko á, Ve onika; Chapelan, Alexis; Haupel sho e , Pia;
K ugel, Alexa; Ku jan, Iael; Mihalje ić, Helena; Munnes, S e an; Placzyn a, Ka olina; Pus e , Milena; Salhi, Mohamed;
Scheibe , Ma cus; Tschiskale, Vic o (2024). Decoding An isemi ism: An AI-d i en S udy on Ha e Speech and Image y
Online. Discou se Repo 6. Be lin: Technical Uni e si y Be lin. Cen e o Resea ch on An isemi ism.
DOI: h ps://doi.o g/10.14279/deposi once-19748
Ad iso y Boa d
P o . Johannes Ange mulle , Discou se, Languages
and Applied Linguis ics, The Open Uni e si y, UK
D Ildikó Ba na, Depa men o Social Resea ch Me ho-
dology, Eö ös Lo ánd Uni e si y, Budapes , Hunga y
P o . Michael Bu e , Ame ican Li e a y and Cul u al
His o y, Ebe ha d Ka l Uni e si y o Tübingen, Ge many
P o . Manuela Consonni, Vidal Sassoon In e na ional
Cen e o he S udy o An isemi ism, Heb ew Uni e si y, Is ael
P o . Ni a Elkin-Ko en, Facul y o Law, Tel A i Uni e si y,
Is ael
P o . Ma in Emme , Ins i u e o Media and Communica-
ion S udies, F ee Uni e si y Be lin; Weizenbaum Ins i u e,
Ge many
P o . Da id Feldman, Bi kbeck Ins i u e o he S udy
o An isemi ism, Uni e si y o London, UK
D Joel Finkels ein, Ne wo k Con agion Resea ch Ins i u e;
P ince on Uni e si y, USA
Shlomi Hod, HIIG’s AI & Socie y Lab, Be lin, Ge many
P o . Gün he Jikeli, Ins i u e o he S udy o Con empo a y
An isemi ism, Indiana Uni e si y Blooming on, USA
D Lesley Kla , Depa men o Law & C iminology,
She ield Hallam Uni e si y, UK
P o . Jö g Meibaue , Johannes Gu enbe g Uni e si y
Mainz, Ge many
P o . Claudine Moïse, Labo a o y o Linguis ics and
Didac ic o Fo eign Languages and Mo he Tongues,
S endhal Uni e si y, G enoble 3, F ance
P o . Ch is oph Neube ge , Ins i u e o Media and
Communica ion S udies, Depa men o Poli ical and Social
Sciences, F ee Uni e si y Be lin; Weizenbaum Ins i u e,
Ge many
D And e Obole , Online Ha e P e en ion Ins i u e, Aus alia
D Da id Reichel, Resea ch and Da a Uni , Eu opean Union
Agency o Fundamen al Righ s (FRA), Aus ia
P o . Ma in Reisigl, Depa men o Linguis ics, Uni e si y o
Vienna, Aus ia
Ad . Ido Rosenzweig, The Mine a Cen e o he Rule o
Law unde Ex eme Condi ions, Uni e si y o Hai a, Is ael
P o . Eli Salzbe ge , The Mine a Cen e o he Rule o
Law unde Ex eme Condi ions, Uni e si y o Hai a, Is ael
D Cha les A. Small, Ins i u e o he S udy o
Global An isemi ism and Policy; S An ony’s College,
Uni e si y o Ox o d, UK
D Abe Swei y, Home O ice, UK
P o . Gab iel Weimann, Depa men o Communica ion,
Uni e si y o Hai a, Is ael
D Ma k Wei zman, Wo ld Jewish Res i u ion O ganiza ion,
USA
P o . Ha ald Welze , No be Elias Cen e o T ans-
o ma ion Design & Resea ch (NEC), Uni e si y o Flensbu g;
Fu u zwei. S i ung Zukun s ähigkei , Ge many
D Juliane We zel, Cen e o Resea ch on An isemi ism (Z A),
Technical Uni e si y Be lin, Ge many
Michael Whine MBE, UK & Bu eau Membe ,
Eu opean Commission Agains Racism and In ole ance,
Council o Eu ope; Eu opean Jewish Cong ess, Belgium
P o . Ma hew L. Williams, C iminology; Ha eLab,
Ca di Uni e si y, UK
Table o Con en s
Execu i e summa y 6
1. Decoding An isemi ism: he jou ney so a 8
2. The Hamas e o a acks on 7 Oc obe 14
2.1 Immedia e online esponses o he e en s 14
2.1.1 UK 14
2.1.2 F ance 16
2.1.3 Ge many 17
2.2 Compa ison o Oc obe 2023 and May 2021 18
2.2.1 UK 19
2.2.2 F ance 20
2.2.3 Ge many 21
2.3 Reac ions o he Is aeli o ensi e a e 13 Oc obe 22
2.3.1 UK 22
2.3.2 F ance 23
2.3.3 Ge many 24
2.4 An isemi ic eac ions on Ins ag am 25
2.4.1 UK 25
2.4.2 F ance 28
2.4.3 Ge many 31
2.5 Quali a i e analyses o esponses in o he coun ies 33
2.5.1 I aly 33
2.5.2 Spain 34
2.5.3 Poland 36
2.5.4 Slo akia 37
2.5.5 Romania 39
2.5.6 Mo occo 40
3. Elon Musk’s online s a emen s abou Jewish indi iduals and o ganisa ions 42
3.1 Geo ge So os 42
3.2 ADL 44
3.3 “The A is Fo me ly Known as E ic” 46
3.4 Media Ma e s o Ame ica, IBM, The Whi e House 47
4. Ou wo kshop se ies on web- ela ed ha e s udies: conclusions and akeaways 48
5. Au oma ed de ec ion o an isemi ic ex s: is con ex all we need? 52
5.1 A emp s o op imise exis ing BERT-based model 52
5.2 Expe imen s wi h p omp -based models 54
5.3 Expe imen al esul s 55
5.4 Conclusion and u u e wo k 57
6. Quan i ying an isemi ism: app oaches, challenges and p omises 58
Execu i e summa y
1. This epo examines an isemi ic discou se online du ing an ex ao dina ily
umul uous ime. The Hamas a acks o 7 Oc obe and he subsequen Is aeli
mili a y esponse in Gaza gene a ed a huge le el o online discussion as well
as inci emen , unp eceden ed in he p ojec ’s ime ame, e en in poli ically
mode a e con ex s. The epo sha es ou indings om an analysis o o e
19,500 commen s pos ed online in he UK, F ance and Ge many since he
a acks, wi h an addi ional 6,000 commen s p esen ed in a six-coun y com-
pa a i e s udy.
2. An isemi ic con en pos ed in he i s week ollowing he a acks (7–13 Oc obe )
ep esen s a u ning poin in an isemi ic online communica ion, cha ac e ised
by open celeb a ion and a i ma ion o iolence in lic ed on Is aeli ci ilians by
Hamas.¹ In he UK co pus, on a e age 27% o an isemi ic commen s a i med
he a acks; in he F ench da a, i was a s agge ing 55%. By con as , he
Ge man co po a p ominen ly ea u ed he an isemi ic concep o is ael’s sole
guil o he con lic (29.6% o all an isemi ic commen s), pinning esponsibil-
i y o he 7 Oc obe a acks on Is ael i sel .
3. To ge a be e sense o he no el y o hese immedia e eac ions, we com-
pa ed hem o ou p e ious analysis. While he o e all an isemi ism le els
emained simila , a much wide ange o an isemi ic concep s was employed
in he 2021 co pus. One eason o his may be a ise in social accep ance
o an isemi ic exp ession online be ween he wo e en s, educing he need
o use a a ie y o s e eo ypes o analogies o demonise – o jus i y ac ion
agains – Is ael. In he 2023 co pus, exp essions o an isemi ic esen men
appea o be mo e o e ly s a ed.
4. In he subsequen phase o he 2023 con lic (13–31 Oc obe ), by which ime
he Is aeli mili a y esponse in Gaza had become he cen e o media a en-
ion, he pa e n o an isemi ic discou se g adually e e ed o ha seen in
p e ious escala ion phases in he Middle Eas . Ac oss all h ee language
communi ies, claims ha Is ael was e il, a e o is s a e o genocidal, a ious
o ms o he nazi analogy, as well as no ions o Jewish powe and in luence,
eplaced he p io a i ma ion o iolence as he op concep .
1 – Fo ela ed esul s,
please ind ou addi-
ional p elimina y s udy
on Facebook and
YouTube commen s sec-
ions soon a e 7 Oc o-
be : Becke , Ma hias J./
Ascone, Lau a/Bol on,
Ma hew/Chapelan,
Alexis/Haupel sho e ,
Pia/K ugel, Alexa/
Placzyn a, Ka olina/
Scheibe , Ma cus/
Tschiskale, Vic o , 2023.
Celeb a ing Te o :
an isemi ism online
a e he Hamas
a acks on Is ael:
P elimina y Resul s I.
Be lin: Technical
Uni e si y Be lin.
Cen e o Resea ch
on An isemi ism,
h ps://doi.o g/
10.14279/
deposi once-19143.
7
Decoding An isemi ism
5. Th oughou he a e ma h o he Hamas a acks, Ins ag am played a c ucial
ole in he online discou se, pa icula ly amongs younge web use s.
Analysis o esponses o pos s by bo h p o-Is aeli and p o-Pales inian Ins a-
g am in luence s e ealed ha he mos equen an isemi ic ideas we e claims
ha Is ael is inhe en ly e il, is commi ing genocide and has no igh o exis .
These we e o en en angled in conspi a o ial na a i es poin ing o supposed
Jewish poli ical o economic powe and in luence.
6. Ou analysis o 2,000 social media commen s ela ed o Elon Musk’s con o-
e sial s a emen s h oughou 2023 highligh ed he phenomena o a i ma ion
and denial o an isemi ism, he e igge ed by he use o classical opes o e il
and conspi acy by in luen ial public igu es.
7. Fo he i s ime, he p ojec expanded i s ocus o o he language commu-
ni ies, examining web use esponses o 7 Oc obe in I aly, Spain, Poland,
Slo akia, Romania and Mo occo. This analysis p o ided a glimpse in o he
dis inc discou ses o each language communi y: om a no ably high le el
o an isemi ic slu s in he Romanian co pus o he p ominence o conspi acy
heo ies in he Polish da a.
8. In he au umn o 2023, we an a se ies o online wo kshops which e lec ed
on he me hods and aims o ou p ojec , es ablished a dialogue wi h o he
expe app oaches, and opened a deba e on he p ac ical challenges o
con on ing ha e speech. This endea ou illumina ed bo h he ad ancemen s
and sho comings in su ounding esea ch a eas.
9. As demons a ed by ou wo k on he AI-based S ep 2, au oma ed de ec ion
o an isemi ic con en online con inues o pose signi ican challenges and
ea i m he impo ance o con ex uali y. E en s a e-o - he-a models like
Cha GPT, which ha e p ocessed as po ions o he in e ne , encoun e
subs an ial di icul ies. A econside a ion o he app oach and examina ion o
en i e discussion h eads migh imp o e he cu en esul s.
10. As pa o he quan i a i ely o ien ed S ep 3 o he p ojec , we de eloped an
in e ac i e da a isualisa ion, accessible on ou homepage. Using hese ools,
a ious discou se e en s can be compa ed in de ail, allowing use s o inspec
he equencies and co-occu ences o labels assigned in quali a i e anno a-
ion du ing S ep 1, as well as ex speci ic an isemi ic keywo ds.
1. Decoding An isemi ism:
he jou ney so a Ma hias J. Becke
The impe a i e o in es iga ing online an isemi ism
has seldom been as compelling as a he ime o p e-
pa ing his epo . The e en s o 7 Oc obe plunged
he Middle Eas in o chaos once again and ha e
b ough he issue o global an isemi ism, bo h on- and
o line, o he o e on o public a en ion. The emo-
ional u gency wi h which an i-Is ael demons a o s
pou ed on o he s ee s ollowing he Hamas a o-
ci ies, some imes e en be o e any subs an ial Is aeli
mili a y esponse, he allies a Wes e n uni e si ies
o en explici ly exp essing solida i y wi h Hamas,
calls o he elimina ion o Is ael and physical a acks
on Jewish people, symbols and synagogues, was
ma ched by an ee ily cold de achmen om he phys-
ical iolence, o en laced wi h misogyny, in lic ed on
Is aeli ci ilians on 7 Oc obe . Fo some, he Hamas
a oci ies we e a cause o ou igh celeb a ion.2
O he s sough o dis o o deny he eali y o he
a acks, e using o gi e Is aelis he s a us o ic ims,
and ea ing demands o he elease o hos ages
as li le mo e han con ec ed Is aeli p opaganda.3
The cen al ole played by social media in acili-
a ing he wa e o Is ael- ela ed an isemi ism ha
has engul ed global socie ies o e he las mon hs
emphasises once mo e he singula impo ance o
examining he pa icula o ms o an isemi ic exp es-
sion and concep ual con en ound online. This is
ue no only o dis o ed deba es abou Is ael bu
also o ano he deeply conce ning phenomenon,
namely, indica ions ha young people a e inc eas-
ingly suscep ible o dis o ions and e en doub he
e aci y o he Holocaus al oge he .4
While he poli ically mode a e legacy media is by no
means de oid o an i-Is ael bias, pa icula ly wi hin
le -leaning ou le s, i is s ill hank ully a e o ind
an isemi ic concep s, such as s e eo ypes and o e
demonising analogies, wi hin mains eam epo ing.
By con as , bo om-up na a i es eme ging om
web communi ies, cha ged by anonymi y and algo-
i hmic p ocesses o adicalisa ion, ha e c ea ed he
condi ions o he ee a icula ion o a discou se o
an isemi ism laden wi h (ancien and upda ed) s e e-
o ypes and his o y-dis o ing analogies.5 In addi ion
o he decline o he p o essional sa egua ding once
o e ed by mains eam jou nalis s and media ou le s,
web use s a ely ace se ious consequences when
ehabili a ing an isemi ic opes in public discou se,
o when pu suing he u hlessly simplis ic e alua-
ion o he highly complex A ab-Is aeli con lic in a
black-and-whi e dicho omy.
2 – Becke , Ma hias J. e al., 2023.
3 – Rich, Da e, 2023. The pos e wa . E e yday Ha e, 29 Oc obe 2023
h ps://e e ydayha e.subs ack.com/p/ he-pos e -wa
(las accessed on 15 Decembe 2023).
4 – The Economis , 2023. One in i e young Ame icans hinks he Holocaus is
a my h. The Economis , 7 Decembe 2023, h ps://www.economis .com/uni -
ed-s a es/2023/12/07/one-in- i e-young-ame icans- hinks- he-holocaus -is-a-
my h (las accessed on 18 Decembe 2023); Ramgopal, Ki , 2023. Su ey inds
‘shocking’ lack o Holocaus knowledge among millennials and Gen Z. NBC News,
16 Sep embe 2023, h ps://www.nbcnews.com/news/wo ld/su ey- inds-shock-
ing-lack-holocaus -knowledge-among-millennials-gen-z-n1240031 (las accessed
on 15 Decembe 2023); Weisman, Jona han/Igielnik, Ru h/McFadden, Alyce,
2023. Poll Finds Wide Disapp o al o Biden on Gaza, and Li le Room o Shi Gea s.
New Yo k Times, 19 Decembe 2023, h ps://www.ny imes.com/2023/12/19/
us/poli ics/biden-is ael-gaza-poll.h ml (las accessed on 13 Janua y 2024).
5 – Becke , Ma hias J., 2021. An isemi ism in Reade Commen s.
London: Palg a e Macmillan; Hübsche , Monika/ on Me ing,
Sabine (eds.), 2022. An isemi ism on Social Media. London: Rou ledge.
15
Decoding An isemi ism
isible below he line o news s o ies po aying
Is aelis and/o Jews as ic ims, and/o po en ially
ulne able.14 As no ed in he ollowing compa a i e
s udy below (2.2), p e ious escala ion phases o
he con lic had no ea u ed high le els o explici
sel -posi ionings by web use s openly endo sing,
applauding, and celeb a ing he e o a acks.
The cen ali y o such celeb a ion wi hin he online
esponse o he 7 Oc obe a acks he e o e seems o
indica e a u ning poin in an isemi ic web discou se
ega ding Is ael.
Use s o en exp essed such a i ma ion di ec ly:
“way o go HAMAS.. we suppo
you..” (TELEG-FB[20231009]);
“Ve y good Pales ine
🇯🇴
👏
good job God bless you
(TELEG-FB[20231007])”,
“long li e HammasLong li e
Pales ine
🇯🇴
We s and wi h
hem and app ecia e hei
e e y s ep hey ake is ael is
a e o is s a e and dese e
such kind o des uc ion”
(BBC-FB[20231009]).
Al e na i ely, hey suppo ed he a acks by depic ing
hem as he ac ions o ‘ eedom igh e s’ – “Figh ing
agains an occupying o ce is no e o ism, i ’s igh -
ing o eedom” (INDEP-FB[20231008]); “F eedom
Figh e s a en’ e o is s” (TELEG-FB[20231010]).
O he commen e s glee ully p edic ed ha he a acks
we e only he i s s ep in a long campaign ha would
lead o Is ael’s des uc ion: “do you hink hey en e ed
i wi h no plan B,C,D...enjoy he las days o is-no -
eal” (BBC-FB[20231008]). This commen combines
a i ma ion wi h he denial o is ael’s igh o exis
h ough he use o wo dplay.
Such denials made up he second-la ges
ca ego y o an isemi ic commen s in he co pus
(18%). Commen e s o en combined a delegi imisa-
ion o Is ael wi h a h ea o Is ael’s u u e e adica-
ion: “ he e is no such hing as is ael and no ma e
how many yea s will ake hey will be expeled om
pales ine” (BBC-FB[20231009]); “All Islamic s a es
mus join hands wi h Pales ine. Is ael mus pe ma-
nen ly dele ed” (BBC-FB[20231008]). Accusa ions
ha Is ael was en i ely esponsible and has sole
guil o he con lic – including he 7 Oc obe
a acks hemsel es – we e also common (16.9%):
“Bu doing wha he Is aeli’s ha e been doing o
75yea s in Pales ine, and expec ing no comeback
as a esul o hei ac ions is ou ageous.Kick a dog
o long enough, i will bi e you.I can say ha he
dog is o blame because he dog is he one mis-
ea ed, and mos people would say good o he
dog” (TELEG-FB[20231009]).
O he commen e s ook he epo s o he Hamas
a acks as an oppo uni y o exp ess hei condem-
na ion o Is ael in e e mo e ex eme ways.
Compa isons be ween Nazi Ge many and
Is ael we e p ominen – “sel -p oclaimed Fascis
go e nmen which belie eds in shoo ing and e i y-
ing chuild en and has as i ’s xcen al policy a ‘Final
solu ion’ o ex e mina ion o he na i es.70 yea s
ago, we we e FIGHTING People like ha .Wha
changed?” (GUARD-FB[20231012]).
14 – In addi ion, he s ikingly compa able obse a ions made
by Mo is should be men ioned he e: “The cu ious hing is ha
an isemi ic inciden s inc eased exponen ially a e he massi e
Hamas slaugh e o 7 Oc obe , e en be o e he Is aelis launched
hei coun e a ack on Gaza. In o he wo ds, i is signs o Jewish –
meaning Is aeli – weakness ha ini ially igge ed he ecen
an isemi ism in he Uni ed S a es, a he han pe cei ed Is aeli
w ongdoing agains A abs, hough ha su ely didn’ help. In his,
he e a e echoes o 1930s Poland, whe e an isemi ism inc eased
in esponse o Ge man a acks on Jews and signs o Jewish
weakness.” (Mo is, Benny, 2023. Muslim An isemi ism and
he Wes e n Le . Quille e, 7 No embe 2023,
h ps://a chi e.ph/ubbeN#selec ion-455.0-455.40).
O he s ex ended he ‘ eedom igh e ’ image o
make compa isons be ween Hamas and esis ance
igh e s o he Nazis, he eby indi ec ly ac i a ing
he nazi analogy be ween Is ael and he Thi d Reich:
“Pales ine s ands condemned, o wha ? Doing wha
he B i ish would ha e done i Ge many had made
i ac oss he channel, wha he F ench esis ance
did do” (GUARD-FB[20231011]). Some web use s
ac i a ed classic an isemi ic associa ions o Jews
wi h he de il: “I s Is ael who has been commi ing
genocide o Muslims since decades and all he de il
powe s a e wi h he ...
” (TELEG-FB[20231012]), o
made dehumanising claims ha “Zionism is a cance ”
(TELEG-FB[20231010]).
2.1.2 F ance
As in he UK, Hamas’s a acks we e widely co e ed
by he F ench media. Wi hin he F ench da ase ,15
16% o commen s we e conside ed an isemi ic.
Simila ly o he B i ish co pus, di ec o indi ec a i -
ma ion o iolence ca ied ou by Hamas ep e-
sen ed by a he mos p ominen o m o an i-
semi ism (55.2% o all an isemi ic commen s).
This again sugges s ha while o e all le els o
an isemi ism emained s able, i s se e i y and in en-
si y inc eased. Explici suppo o Hamas’s ac ions –
such as “Long li e Hamas” [“Vi e le Hamas”] – was
o en dilu ed by mo e gene ic suppo o he Pales-
inian cause: “To al suppo o Pales ine” [“Sou ien
o al à la Pales ine”] (FRA-FB[20231009]). Tha
no ion was also a icula ed h ough he allega ion
ha he a ack was mo ally jus i ied as sel -de ence
o e ibu ion o Is aeli iolence: “I is only payback
o he Zionis pa asi es” [“Ils n’on ai que end e
la monnaie de leu pièce aux pa asi es sionis es”]
(MONDE-FB[20231008]).
The nazi analogy was also a popula means
o con eying suppo o Hamas (10.4%).
Hamas was egula ly compa ed o he F ench
Resis ance h ough commen s such as “How did we
end up u ning esis ance in o e o ism? Du ing he
occupa ion in he 1940s, he Resis ance ook up
a ms agains he Ge mans and hey we e glo i ied?”
[“Commen on a i e à ans o me la ésis ance
en e o isme? Pendan l’occupa ion dans les
années 40 la ésis ance on bien p is les a mes
con e les allmands e on les a glo i ié?”] (LEFIG-
FB[20231008a]). Ano he commen ac i a es
he nazi analogy bu also maps on o nega ionis
na a i es:
“Hamas is paying i back o
hei people killed since
48 by
commi ing genocide
(Holocaus ), bu a eal one,
and he pe pe a o s a e he
su i o s o 45
”
[“Hamas end la monnaies de son peuple
qui se ue pa les
depuis 48 aisan un /
génocide (holocaus e) mais un ais e les
bou eaux son les escapés du 45
”]
(LEFIG-FB[20231008b]).
By con as ing he “ eal” genocide o Pales inians
wi h he supposedly ‘ ab ica ed’ one o he Jews,
his commen in e es ingly ep esen s a p ojec ion o
Nazi a oci ies on o he Jewish s a e, while denying
he exis ence o he Nazi genocide. In addi ion o
hese di ec o ms o e e ence, he analogy can also
be a icula ed h ough puns, such as in “Is a Heil”
(MONDE-FB[20231008]).
2. The Hamas e o a acks on 7 Oc obe
15 – The da ase s ems
om he Facebook
p o iles o Le Monde,
Le Figa o, Le Pa isien,
Libé a ion, BFMTV,
F ance24
and 20Minu es.
We analysed he i s
150 commen s om
each h ead.
17
Decoding An isemi ism
Ano he equen ly ac i a ed concep is he
denial o is ael’s igh o exis (9.5%): “Is ael has
no igh o peace and secu i y. Fo he simple eason
i has no igh o exis ” [“Is ael n’a d oi ni a la paix
ni a la secu i e. Pou la simple aison q’il n’a pas
d oi a l’exis ence”] (LEFIG-FB[20231008a]). This
concep is o en implied h ough he allega ion
ha Is ael is a e o is s a e (6.5%):
“How many Pales inians killed
by he e o is s a e o
Is ael !?!!
”
[“Combien de pales iniens mo pa
l’é a e o is e d’Is aël !?!!
”]
(LEFIG-FB[20231008b]).
In o he e ms, he eac ions o he a acks pe pe-
a ed by Hamas in Oc obe 2023 seem o ocus
mo e on he app ecia ion o Hamas (o Pales ine,
when con ex ually unde s ood as a me onymy o he
e o is g oup), han on he demonisa ion o Is ael.
This shows ha e en posi i e s a emen s – posi i e
om a lexical pe spec i e – can ac ually ep esen a
o m o ha e speech and, in his case, o an isemi ic
discou se. The con ex o he s a emen s is he e o e
c ucial o de e mine hei na u e.
2.1.3 Ge many
The Hamas a acks on Is aeli ci ilians gene a ed a
highly ampli ied esponse also on Ge man social
media. Wi hin he da ase ,16 a o al o 12.6% o
all commen s we e iden i ied as an isemi ic,
al hough he e was a wide ange o 8% o 29% o
an isemi ic commen s in he indi idual h eads.
E en hough he a e age p opo ion o an isemi ic
commen s is compa able o o he discou se e en s in
Ge man-speaking coun ies (see p e ious discou se
epo s and sec ion 2.2), some hema ic and
concep ual peculia i ies a e s iking. Fi s o all, he
an isemi ic s a emen s a icula ed wi hin he i s
phase o he con lic jus i y he cause o Hamas’
a ack by e e ence o Is ael’s beha iou , a ib-
u ing o Is ael sole guil o he con lic (29.6%
o all an isemi ic commen s): “Wha has is eal
done wi h all he Pales inians?! Think i s !!! Wha
happened be o e. Whe e all his ange comes
om...” [“Was ha is eal mi den ganzen Paläs in-
ense gemach ?!E s mal nachdenken!!! Was o he
geschehen is . Wohe diese ganze Wu komm ....”]
(WELT-FB[20231008]). This is legi imised by he
accusa ion ha he a ack is due o Is ael being
a e o is s a e (12.7%) – “The Pales inians a e
igh ing agains he e o is s a e o Is ael” [“Die
Paläs inense kämp en gegen den Te o s aa Is ael”]
(BILD-FB[20231009a]), o commi ing genocide17
(5.8%) agains he
Pales inians – “I
someone wan s
o libe a e hei
coun y and sa e
hei people om
genocide, hey a e
no a e o is bu
a eedom igh e !”
[“Wenn jemand
sein Land be eien
will und sein Volk
16 – The da ase con ains en h eads, wi h 150 commen s
analysed om each and s ems om he Facebook p o iles o
De Spiegel, DieWel , Bild and n- .
17 – In iden i ying dis o ed s a emen s, such as he claim ha
Is ael would commi a genocide o s uc u ally esemble Sou h
A ica’s apa heid, we ollow he p inciple ha asse ions p edic -
ing simila scena ios in and a ound Is ael in he u u e a e no
o be classi ied as an isemi ic (see Becke , Ma hias J./T oschke,
Hagen/Bol on, Ma hew/Chapelan, Alexis (eds.), 2024.
Decoding An isemi ism: A Guide o Iden i ying An isemi ism
Online. London: Palg a e Macmillan/Sp inge Na u e.
Howe e , while he e is no in e na ional poli ical and legal
consensus ye ha would ully legi imise he genocide accusa ion,
we acknowledge he ongoing, u gen deba e a ound his opic.
2.2 Compa ison o Oc obe 2023 and May 2021
In o de o e alua e he no el aspec s o he 7 Oc obe
eac ions, we sough o compa e i wi h online
esponses o a p e ious ou b eak o iolence in he
Middle Eas . The escala ion phase o he Is ael-Hamas
con lic in May 2021 was he mos se ious ou b eak
o iolence in he yea s p eceding he 7Oc obe
a acks. A ha ime, Hamas i ed hund eds o ocke s
om Gaza in o Is ael a e clashes be ween Pales ini-
ans and Is aeli police a he Al-Aqsa Mosque,
ollowed by Is aeli mili a y e alia ions in Gaza.
While nei he he Hamas a acks no he Is aeli
esponse came close o he se e i y o ecen e en s,
May 2021 ne e heless ep esen s he closes ch on-
ological poin o compa ison o an e alua ion o he
simila i ies and di e ences in online esponses.
The co pus o he May 2021 case s udy was again
made up o 4,500 commen s (1,500 om each lan-
guage communi y) aken om he Facebook pages
o leading media ou le s.18
2. The Hamas e o a acks on 7 Oc obe
o einem Genozid e en will is kein Te o is son-
de n ein F eihei skämp e !”] (BILD-FB[20231009b]).
The p esen a ion and legi ima ion o Hamas’s ac ions
as a igh o eedom pe aded he en i e discou se:
“In Uk aine, hese people a e called eedom igh e s”
[“In de Uk aine nenn man die Leu e F eihei skämp-
e ”] (BILD-FB[20231009c]). As a esul , Is ael’s
igh o sel -de e mina ion is o en comple ely
denied (24% o all an isemi ic commen s):
“The Jews a e he p oblem. We all know ha .
They don’ own he coun y” [“Die Juden sind das
P oblem. Wissen wi doch alle. Ihnen gehoe das
Land nich ”] (WELT-FB[20231008]); “ques ion ,his-
o y who owns is ael , eedom o pales ine” [“ age
,geschich e wem gehö is ael , eihei ü paläs ina”]
(BILD-FB[20231009b]).
Mo eo e , hese a emp s o legi imise Hamas’s
iolence h ough Is ael’s ac ions culmina e in open
a i ma ion o he e o is a acks – “Awesome
wha he Pales inians ha e pulled o , chee s o
Hamas o a g ea ac ion” [“Geil was die Paläs in-
ense da abgezogen haben ein hoch au Hamas
geile Ak ion”] (BILD-FB[20231009b]) – o a icula-
ion o he desi e o u u e and e en mo e ex ensi e
iolence agains Is ael and/o Jews:
“I hope Is ael becomes
his o y like Uk aine”
[“ich ho e Is ael wi d Geschich e
so wie die Uk aine]
(BILD-FB[20231009c]).
18 – Fo mo e de ailed in o ma ion abou
ou analyses o he p e ious escala ion
phase in May 2021, see Becke , Ma hias J.
e al., 2022. Decoding An isemi ism:
An AI-d i en S udy on Ha e Speech and
Image y Online. Discou se Repo 2.
Be lin: Technical Uni e si y Be lin.
Cen e o Resea ch on An isemi ism.
19
Decoding An isemi ism
2.2.1 UK
The o al le el o an isemi ism wi hin he May
2021 h eads19 was, a 26.9%, simila o ha
ound in he Oc obe 2023 h eads. Howe e , he
ange o he an isemi ism le els was much na owe ,
wi h all h eads bunching a ound he 20% ma k and
no high ou lie s. The analysed h eads we e di ided
in o wo main ca ego ies, one ocusing on Hamas’s
ocke a acks and he o he on he IDF’s e alia o y
ac ions. The pa e n o an isemi ic s e eo ypes in
he commen s sec ions shi ed depending on he
emphasis in he a icles. When he media highligh ed
Hamas’s ac i i ies, he e was a no able inc ease in
commen s accusing he media o a ou i ism owa ds
Is ael. This esponse was clea ly dis inc om eac-
ions o media epo ing o he 7 Oc obe a acks – in
he la e , he e was li le sign o conce n abou media
bias, bu a he , as no ed abo e, di ec a i ma ion o
he a acks (27% o an isemi ic commen s).
Con e sely, ew i any esponses o he Hamas
ocke s o May 2021 sough o a i m o cele-
b a e he a acks. Specula i ely, his migh be due
o he ela i e ailu e o he ocke a acks in compa -
ison wi h he iolence o 7 Oc obe : had he May
2021 ocke a acks esul ed in he same physical,
emo ional and psychological impac o he ecen
Hamas a oci ies, pe haps he esponse would ha e
been mo e celeb a o y han conspi a o ial. Ano he
possibili y o his di e ence may be an inc ease in
social accep abili y o an isemi ic exp ession online
be ween he wo e en s. I o e an isemi ism is ind-
ing g ea e social accep ance, he e is less need o
he use o s e eo ypes o analogies as means o jus-
i ica ion. Al e na i ely, while he May 2021 escala-
ion was no p eceded by a build-up o ension, bu
seemed o come ou o he blue, he Oc obe 2023
a acks ollowed a yea in which he inclusion o a -
igh ex emis s in he new Ne anyahu go e nmen ,
in e nal Is aeli poli ical con lic s o e judicial e o m,
and spo adic bu egula iolence on he Wes Bank,
ecei ed low-le el bu pe sis en media co e age in
all h ee coun ies. This may ha e had a adicalising
e ec on global pe cep ions o Is ael o Jews as
whole, which we e hen ac i a ed in he esponses o
he iolence o 7 Oc obe .
On he o he hand, when he May 2021 a icles cen-
ed on Is ael’s esponse o he ocke a acks, Is aelis
and Jews hemsel es became he a ge o a i-
ous an isemi ic s e eo ypes, wi hou e e ence o
media a ou i ism. The mos equen ly exp essed
s e eo ypes in he co pus we e no ions o Jewish
o Is aeli e il (39% o all an isemi ic commen s)
and a ibu ing sole blame o he con lic o Is ael
(27%). The la e concep was s ongly ep esen ed in
he Oc obe 2023 co pus as well (16.9%), as com-
men e s who may be uneasy wi h di ec a i ma ions
o he Hamas a acks sough o jus i y hem.
The concep o e il was ela i ely in equen in he
Oc obe 2023 co pus (6.4%). This may be due o he
ocus o he co pus being on Hamas a acking Is ael,
a he han ice e sa. I may also be due o use s
p e e ing o condemn Is ael by means o he closely
ela ed e o is s a e accusa ion (8.3%), which
o en ea u es as a esponse o media a icles which
depic Hamas as a e o is o ganisa ion. Ano he
concep o ea u e p ominen ly in bo h co po a was
he denial o is ael’s igh o exis , al hough his
was mo e common in he Oc obe 2023 commen s
(18%) han in he May 2021 co pus (7.7%). Again,
he di e ence o his dispa i y may lie in he ela i e
success o he espec i e Hamas a acks. As seen
abo e, many web use s saw he 7 Oc obe a acks
as he i s s ep owa ds he inal des uc ion o Is ael,
and hei eupho ia may ha e mo i a ed jus i ica ion
o , p edic ions o h ea s o Is ael’s ul ima e demise.
19 – The da ase con ains en h eads, wi h 150 commen s
analysed om each, and s ems om he Facebook p o iles
o BBC News, Daily Exp ess, Daily Mail, Financial Times,
The Gua dian, The Independen , Daily Mi o , The Spec a o ,
The Teleg aph and The Times.
2.2.2 F ance
Compa ed o Oc obe 2023, he online commen a y
on May 202120 exhibi ed sligh ly lowe an isemi-
ism le els: 12.9%, as opposed o 16% in 2023.
Mo e impo an ly, di e ences eme ged in he con en
o hese an isemi ic web commen s. In 2021, mos o
he an isemi ic commen s a ge ed Is ael and Is aelis
a he han Jews. He e, he mos equen an isemi ic
s e eo ype was he one p esen ing Is ael as e il:
“The ba ba ism o his
S a e canno be hidden by
2–3 lies o adi ional
media” [“Ce n’es pas 2 - 3
mensonges des médias adi-
ionnels qui pou on cache
la ba ba ie de ce e é a ”]
(MONDE-FB[20210511]);
“I ’s he only people in
he wo ld ha boas s abou
killing human beings”
[“C’es le seul peuple au monde qui se
an e de ue des ê es humains”]
(MONDE-FB[20210511]).
This s e eo ype appea ed in 49.7% o an isemi ic
commen s in he 2021 co pus, bu only 13.9% in
2023. The same endency was iden i ied o he
allega ion ha Jews a e child mu de e s and ha
is ael has no igh o exis : “They’ e esis ing and
de ending hemsel es agains he eal e o is s
who bomb child en” [“Ils on de la ésis ance e se
dé enden con e les ais e o is e qui bomba de
des en an s”] (LEPAR-FB[20210511]); “Is ael doesn’
exis and will ne e exis ” [“Is aël
n’exis e e n’exis e a jamais”]
(LEPAR-FB[20210511]).
The 2021 co pus was also cha ac e ised by com-
men s accusing Is aelis o holding powe (5.2%)
o e he media and being lia s (4.7%) and:
“I ’s clea ha le monde is paid by hese demons”
[“on oi bien que le monde es payé pa ces
demons”] (MONDE-FB[20210510]); “You’ e he
champions in lying and alsi ying his o y” [“Vous
ê es champions dans les mensonges e la alsi ica-
ion de l his oi e”] (LIB-FB[20210512]).
Meanwhile, he 2023 co pus p esen s many mo e
e e ences o is ael’s sole guil in he con lic
(9.9% in 2023 and only 2.6% in 2021): “Is ael
ne e wan ed peace, hey’ e jus eaping he
bene i s o hei poli ics” [“Is aël n’a jamais oulu la
paix, ils écol en jus e les ui s de leu poli ique”]
(LEPAR-FB[20231009]). As his commen shows,
he concep is o en used o legi imise and suppo
Hamas’s a acks as p edic able and dese ed due
o Is ael’s own ac ions. The a i ma ion o iolence
pe pe a ed by Hamas, which as s a ed be o e con-
s i u es he mos p ominen o m o an isemi ism in he
2023 co pus (53%), was absen in 2021. The e,
commen e s did no exp ess hei suppo o he
hund eds o ocke s launched om Gaza bu a he
called o a u u e iolen ac ion agains Is ael:
“you day will come and you
will be mu ila ed, massac ed,
killed like he housands o
innocen s Pales inians killed
by Is ael. This day will come”
[“ on jou a i e a e u se as mu ilé
masac é ué comme les millie s de
pales iniens innocen s ué pa Is aël
Ce jou a i e a”]
(LEPOI-FB[20210512]).
2. The Hamas e o a acks on 7 Oc obe
20 – We examined he i s
100 commen s om h eads on he
Facebook p o iles o Le Monde,
Libé a ion, Le Figa o, Le Pa isien, Le Poin ,
L’Exp ess and 20Minu es.
21
Decoding An isemi ism
2.2.3 Ge many
While a signi ican o sligh inc ease in an isemi ic
commen s can be obse ed wi hin he B i ish and
F ench media, he le el o an isemi ic commen s
in he Ge man discou se ac oss he wo esca-
la ion phases emained qui e s able.21 Wi hin
he 2023 measu emen pe iod, 12.6% o all
coded commen s we e iden i ied as an isemi ic –
in con as o 13.4% in 2021. While i s quan i y
was ela i ely cons an , he quali y o an isemi ic
communica ion has changed s ikingly. In May
2021, Is ael was demonised p ima ily ia wo
an isemi ic concep s.
Fi s ly, i was cha ac e ised as he e il o he wo ld
by implying ha Is ael’s ac ions a e essen ially
wicked o ha i would pu pose ully and p oac i ely
cause ex ensi e ha m o o he s. While 39.7% o
all an isemi ic commen s in 2021 con ained he
concep , in 2023 i was only 4.2%. As in May 2021,
commen s in he Oc obe 2023 co pus con inue o
po ay Is ael nega i ely, o blame i unila e ally
o he con lic : “Yeah igh , i ’s Hamas’s aul . I ’s
Is ael’s aul , hey’ e been e o ising Hamas o
yea s. Bu nobody ca es. Now all o a sudden i ’s
poo Is ael. Hamas is igh . Ti o a ” [“Aha die
Hamas is schuld.Is ael is schuld die machen
schon Jah e lang Te o gegen die Hamas. Nu das
in e essie keinen. Je z au einmal heiß es A mes
Is ael. Die Hamas ha Rech .Wie du mi so ich di ”]
(WELT-FB[20231007]). Ins ead o essen ialising
Is ael ac oss he boa d wi h he concep o e il, com-
men e s a e eso ing o mo e nuanced jus i ica ions –
which explains he decline in he use o he e m
compa ed o 2021, especially since he a icula ion
o o he s e eo ypes mo e han doubled in 2023:
while in 2021 11.3% o all an isemi ic commen s
exp essed is ael’s sole guil in he con lic ,
29.6% ealised his s e eo ype in 2023: “Is ael has
been s a ing his o 70 yea s” [“Is ael äng sei
70Jah e an”] (BILD-FB[20231009d]).
Secondly, in 2021, Is ael was inc easingly accused
o exe ing in luence on bo h he media and
poli ics (11.3%) – a concep a ea ha is less
p onounced, bu s ill consis en ly used in 2023:
7.9% o an isemi ic commen s in 2023 made accu-
sa ions such as
“No wonde Bild epo s all
his abou Is ael, since Bild
GmbH belongs o Jews”
[“Kein Wunde das Bild das alles
übe Is ael be ich e weil doch das
Bild GmbH Juden gehö ”]
(BILD-FB[20231009e]).
In con as , use s in 2023 exp ess hei a i ma ion
o iolence mo e clea ly. While in 2021 only 3.4%
o all an isemi ic commen s con ey such a i udes
owa ds Is ael and/o Jews, o legi imise he iolen
ac ions o Hamas: “Ac ion ~ Reac ion” [“Ak ion ~
Reak ion”] (WELT-FB[20231008c]). The idea o
is ael’s sole guil in he con lic he e o e pe ades
bo h discou se e en s.
21 – In 2021, we examined 150 commen s
each om en h eads aken om he
Facebook p o iles o Bild, FAZ, Focus,
n- , RP Online, De Spiegel, Süddeu sche
Zei ung, az, Die Wel and Die Zei .
In o de o con inue acking esponses o he un old-
ing si ua ion in he Middle Eas , a u he co pus o
1,000 commen s pe language communi y was buil
om esponses o epo s (published 13 o 31 Oc o-
be ) o he Is aeli mili a y bombing campaign and
he hen-p ospec i e in asion o Gaza ha ollowed
he 7 Oc obe a acks.
2.3.1 UK
The UK da ase anged om accoun s o Is aeli
bombing aids o UK P ime Minis e Rishi Sunak’s
isi o Is ael.22 Among he esponses om he
eade ship, 19.9% we e classed as an isemi ic.
Compa ed wi h he commen s pos ed in he imme-
dia e a e ma h o he Hamas a acks, he e was a
ma ked d op in he le el o a i ma ion and/
o celeb a ion o an isemi ic iolence. As he
co e age e u ned o he s anda d pa e n ollow-
ing p e ious escala ion phases (wi h he cen al
ocus now being Is aeli mili a y ac ions a he
han Hamas iolence), so oo did he pa e n o
an isemi ic exp ession e u n o he mean. The mos
equen ly exp essed an isemi ic concep was ha
o Is aeli/Jewish e il (27% o an isemi ic com-
men s). Use s decla ed ha “humani y will win he
igh agains he Zionis lobby,we will,” posi ioning
Is ael as ou side o and opposed o humani y,
and added ha “you migh own he poli icians bu
you don’ own he people” – he eby making a
claim o “Zionis ” powe /in luence o e non-
Is aeli poli ical sys ems (BBC-FB[20231018]).
The second mos equen concep (17.5% o
all an isemi ic commen s) was he accusa ion
ha Is ael was commi ing genocide agains
he Gazan popula ion. Some commen e s
claimed ha he Is aeli genocidal a ack on Gaza
was unp eceden ed in i s sa age y, using lan-
guage ha e oked images o he Holocaus and
hus exp essed an indi ec nazi analogy: “2.5mil-
lion people a e subjec ed o a wa o ex e mi-
na ion and a siege ha has ne e happened in
his o y” (BBC-FB[20231018]); “Shame on he
wo ld! They kep wa ching when Hi le did i and
now when Ne anyahu is doing i ! A genocide o
Pales inians!” (INDEP-FB[20231013]).
Use s made equen e e ences o Is ael as a
e o is s a e (11%), a common esponse o he
depic ion o Hamas as a e o is o ganisa ion
o o compa isons be ween he Hamas a oci ies
o 7Oc obe and hose o Da’esh: “Zionis s is
e il Zionis s is he i s e o is O ganiza ion in he
wo ld” (DAILY-FB[20231020]); “is ael he eal isis”
(BBC-FB[20231017]). The denial o is ael’s igh
o exis was o en he ul ima e conclusion o
an isemi ic commen s (10%) – “ac ually wha is
no easy o belie e Is ael when i ’s whole exis ence
is a lie” (BBC-FB[20231017]) – while use s egula ly
eso ed o he nazi analogy (6.5%) o unde line
hei condemna ion o Is aeli ac ions: “Is ael is he
new nazism” (INDEP-FB[20231013]); “ he e no be -
e hen he Nazis” (DAILY-FB[20231020]); “You’ e
wo se han he e y same people you c y abou
e e y yea in you sel -pi ying memo ial se ices –
wi hou ha ing lea n -all. No e en he Nazis mas-
e ed p opaganda and ha nessed he powe o he
media in he way ha Zionis s ha e in he mode n
age” (INDEP-FB[20231017]); “And you hink hi le
is he wo s .. now i unde s and why he did he
Holocaus ” ([INDEP-FB[20231013]).
2. The Hamas e o a acks on 7 Oc obe
2.3 Reac ions o he Is aeli o ensi e a e 13 Oc obe
22 – The commen s we e aken om he
Facebook p o iles o he mains eam UK
media ou le s BBC News, The Independen ,
The Times and Daily Mail.
23
Decoding An isemi ism
2.3.2 F ance
In he F ench co pus,23 8.9% o he commen s
p esen ed an an isemi ic s a emen , meaning ha
he p opo ion o an isemi ic commen s dec eased
compa ed o he i s days a e he a acks. As wi h
he UK esul s, analysis e ealed ha he se e i y
and in ensi y o he an isemi ic commen s diminished
somewha : a i ma ions o iolence pe pe a ed
by Hamas we e e y p ominen in eac ions o he
media a icles published immedia ely a e Hamas’s
a acks (53%), while in his second phase o he
con lic hey cons i u ed only 10% o he an isemi ic
commen s.
In his co pus, he mos p e alen concep is he s e-
eo ype o e il (20.2%). The commen s demonise
ac ions o he na u e o Is ael i sel , o example by
claiming i is “a e y mu de ous coun y” [“Is aël es
un pays ès meu ie ”] (MONDE-FB[20231101])
o “an in asi e me as asised cance ” [“un cance
in asi ayan mé as asé”] (LEFIG-FB[20231101]).
The las example, in which Is ael is demonised
h ough he disease me apho , po aying i as an
en i y wi h a disin eg a ing e ec , is eminiscen o he
19 h- and 20 h-cen u y o ms o an isemi ism.
Al hough in much smalle p opo ion han in he i s
phase, one o he mos equen an isemi ic con-
cep s is assigning Is ael sole blame o he con lic
(12.3%). Commen s end o explain his by lis ing he
ac ions unde aken by Is ael ha ha e allegedly led
o such a d ama ic si ua ion as “ he expansion o he
illegal Is aeli colonisa ion” [“l’expansion des colo-
nies is aéliennes illegals”] (MONDE-FB[20231101]).
The example illus a es he linkage o he concep
o is ael’s guil in he con lic wi h colonialism
analogies, a phenomenon which usually co-occu s
wi h he o me . Implici ly, his consis en ly leads o a
denial o is ael’s igh o exis , unde pinned by he
alse claim ha Is ael was es ablished on Pales ine:
“ he Zionis na ion came wi h an a my o se le by
o ce on Pales inians’ land” [“la na ion sionis e es
enue a ec une a mée s’ins alle de o ce su la e e
des pales iniens”] (LEFIG-FB[20231101]).
23 – The F ench co pus consis s o he
i s 100 commen s pos ed in eac ion o en
a icles published on he Facebook pages
o he ollowing media ou le s: Le Monde,
Le Figa o, Le Pa isien, F ance24 and CNEWS.
2.3.3 Ge many
While he i s phase o Ge man epo ing ocused
on Hamas’s a oci ies, he second cen ed on domes-
ic eac ions o he e en s. The co pus con ains com-
men s sec ions esponding o a ange o he la e ’s
media epo s: om p o-Pales inian demons a ions
in Ge many o assessmen s o hese by he O ice
o he P o ec ion o he Cons i u ion, o he Fede al
Chancello ’s isi o Is ael.24 The second phase o he
s udy pe iod is cha ac e ised by a signi ican ly lowe
p opo ion o an isemi ic commen s: only 8.5% o
all analysed commen s we e iden i ied as an i-
semi ic. This may be due o he change in epo ing,
as web use s o en used epo s on he ac ions o he
Ge man go e nmen as an oppo uni y o c i icise
o exp ess displeasu e abou domes ic poli ics – in
o he wo ds, o shi he opic o discou se.
I is pa icula ly no iceable ha is ael’s sole guil in
he con lic is a ibu ed signi ican ly less equen ly
han in he ini ial phase – only 8.2% o all an isemi ic
commen s ealise his s e eo ype. The high le els o
a i ma ion o iolence, which accoun ed o 20.1%
o an isemi ic commen s in he i s phase, has also
allen signi ican ly, o 12.5%. Ins ead, use s e u n o
amilia pa e ns and cha ac e ise Is ael – simi-
la o he indings om May 2021 (see 2.2) – as
a s a e cha ac e ised by essen ial wickedness,
mos equen ly desc ibed as a e o is s a e
(17.6%): “Fo o e 70 yea s, he e o is S a e o
Is ael has ampled on human igh s” [“Sei übe
70 Jah en i de Te o s aa Is ael das Menschen-
ech mi Füßen”] (WELT-FB[20231031a]), o as
he “ ogue S a e o Is ael” [“Schu kens aa Is ael”]
(SPIEGEL-FB[20231031]).
2. The Hamas e o a acks on 7 Oc obe
In addi ion, he concep o in luence is o en
in oked (14.1% o all an isemi ic commen s)
when depic ions o Hamas a oci ies a e nega ed –
“Ei he hey we e ake pic u es o pic u es o
mu de ed child en om Pales ine. Is ael has some
eally nice ai y ales in s o e” [“En wede wa en es
akebilde ode Bilde on e mo de en Kinde aus
Paläs ina. Is ael ha ich ige schöne Mä chen au
Lage ”] (WELT-FB[20231031b]) – o p esen ed as
disin o ma ion ins umen alised by Is ael in o de o
comple ely annex Gaza:
“This whole epo has
no u h all wi hou bodies
only a ew suspec pic u es
all nicely planned by is ael
so hey can occupy he las
open p ison”
[“Diese ganze Be ich is schon keine
Wah hei alles ohne Leichen nu ein
paa Ve däch igen bilde alles schön
on is ael geplan dami sie das le z e
o en Ge ängnis bese zen können”]
(FB-WELT[20231031c].
This idea – o disin o ma ion a ising om Is aeli/
Jewish powe o e public opinion (16.4%) – was
exp essed much less equen ly in he i s phase.
This can p obably be explained by he ac ha a
jus i ica ion o demonising Is ael’s ac ions is s ill
needed in o de o ob Is ael o any legi imacy o
mili a y ac ion. Howe e , his jus i ica ion mus no
(any longe ) glo i y Hamas’s iolence, as such
glo i ica ion is socially unaccep able.
24 – The da ase con ains en h eads,
wi h 100 commen s analysed om each,
and s ems om he Facebook p o iles o
De Spiegel and Die Wel .
31
Decoding An isemi ism
2.4.3 Ge many
While in he pas he Hamas a acks on Is aeli ci il-
ians igge ed ew eac ions in he Ge man-speak-
ing communi y, he e en s o 7 Oc obe gene a ed
a huge esponse on Ins ag am. Wi hin he da ase
o 3,000 commen s, 16.8% we e iden i ied as
an isemi ic – signi ican ly mo e han on Facebook
o he en i e analysed pe iod. In o de o co e
he de elopmen s o he con lic in i s en i e y and
o ensu e compa abili y wi h he Facebook anal-
yses, he co pus included commen s sec ions om
he pe iod be ween 9 Oc obe and 4 No embe .
While hei poli ical s ance canno always be clea ly
de e mined, we aimed o main ain a balance
be ween he di e en oices, wi h 15 p o-Is aeli
and 15 p o-Pales inian h eads examined o ou
analysis.31 While some h eads a emp o explain
and con ex ualise he con lic and he subsequen
deba es – such as he posi ion o he Ge man
go e nmen , o F idays o Fu u e ac i is G e a
Thunbe g – o he h eads me ely exp essed suppo
o one o he pa ies in he con lic o highligh ed
he su e ing o all ic ims, so ha he p opo ion
o an isemi ic commen s in he indi idual h eads
anged om 1 o 42%. The la ges p opo ion o
an isemi ic commen s could be iden i ied in he
p o-Pales inian h eads.
Al hough he pe cen age dis ibu ion di e s, he mos
equen ly exp essed concep s a e simila o he
indings in he B i ish and F ench da ase s, as well as
he esul s o he Facebook analyses in he p e ious
sec ions. The mos p e alen an isemi ic concep s
we e accusa ions o e il (26.7% o an isemi ic
commen s), genocide (16.8%), a i ma ion o
iolence (11.9%), ee pass (9.7%) and in luence
(9.7%). Howe e , hese opes o en co-occu ed
wi h each o he , as well as wi h concep s such
as an isemi ic conspi acies (7.5%) and e o is
s a e (6.1%), and hei equency depended on he
aming o he o iginal pos .
This is pa icula ly e iden in he ollowing example,
which combines a wide a ie y o an isemi ic ideas:
“Suppo ing Is ael’s igh o exis means ha you sup-
po impe ialism, colonialism, apa heid (Amnes y
In e na ional, Human Righ s Wa ch, B’Tselem), geno-
cide, expulsion, e c. Shows e y well whe e you s and.
Ge many mus s op blaming Pales inians o hei pas ”
[“Exis ens ech Is ael zu un e s ü zen bedeu e , dass
ih Impe ialismus, Kolonialismus, Apa heid (Amnes y
In e na ional, Human Righ s Wa ch, B’Tselem), Völke -
mo d, Ve eibung usw. un e s ü z . Zeig seh gu wo
ih s eh . Deu schland muss au hö en ih e Schuld aus
de Ve gangenhei den Paläs inense n in die Schuhe
zu schieben”] (FFFG-IN[20231020]). The commen
links Is ael’s igh o exis o nega i e a ibu ions,
consequen ly calling ha igh in o ques ion: by p e-
supposing he apa heid analogy, impe ialism anal-
ogies and genocide, i s igh o exis can be nega ed
inso a as he a ibu ions mee wi h social ejec-
ion. A he same ime, he commen alludes o he
Holocaus when insinua ing ha Ge many’s suppo
o Is ael is solely due o his o ical guil , which suppos-
edly blinds i o Pales inian su e ing in he p esen .
The na a i e o Is ael as an absolu e e il pe ades
he en i e discou se and is mos ly used o asc ibe
Is ael wi h wil ul and delibe a ely c uel in en ions
owa ds he Pales inians. The Jewish s a e is ei he
accused o inhumane beha iou :
“I is ob ious ha ‘Is ael’
does no conside human li e
wo h p o ec ing”
[“Es is doch o ensich lich das
Is ael’ menschliches Leben nich
ü schü zenswe häl ”]
(VICE-IN[20231018]),
31 – The co pus was
compiled using he
commen s sec ions o he
Ins ag am pages o he
ollowing Ge man in lu-
ence s, ac i is accoun s
and (media) ins i u ions
as well as poli icians:
VICE au Deu sch,
El on TV, F idays o
Fu u e Ge many,
Düzen Tekkal,
Robe Habeck,
Volks e pe ze ,
Amadeu An onio
S i ung, E ik Ma qua d ,
Paläs ina Sp ich ,
Paläs ina Kampagne,
Ta ek Bae,
Jü gen Todenhö e ,
Gene a ion Islam,
Za a Sec e .
o cha ged wi h ac i e in e es in causing as much su e ing
as possible: “O cou se he Is aelis ca e abou he Pales-
inians. They hink e e y day how bes o make he people
in Gaza su e ” [“Na ü lich sind die Paläs inense den
Is aeli en nich egal. Die machen sich ag äglich Gedanken
da übe , wie man die Menschen im Gaza am bes en leiden
lassen kann”] (TAREK-IN[20231021]), o simply essen ial-
ised as an undemoc a ic e o is s a e unde which he
Pales inians ha e o su e on a daily basis: “I ’s e o ! I is
bad! And wi hou men ioning a bu , he long lis o names
o he innocen , slaugh e ed 3,000 child en in Gaza is also
e o ! Te o o a go e nmen ha is being sold he e as
democ a ic” [“Es is Te o ! Es is schlimm! Und ohne ein
Abe zu nennen, die lange Namenslis e de unschuldigen,
abgeschlach e en 3000 Kinde n in Gaza is ebenso Te o !
Te o eine Regie ung die hie als demok a isch e kau
wi d”] (DUZEN-IN[20231020]). These alleged e o s o
cause hu usually culmina e in insinua ions ha ep esen a
mode n classic o an isemi ic asc ip ions, namely he claim
ha Is ael is ca ying ou a genocide, which, simila ly o
he indings in he F ench co pus, is o en communica ed
openly and explici ly “Is ael is commi ing genocide agains
he Pales inians!” [“Is ael begeh einen Genozid an den
Pales inänse n!”] (VOLKS-IN[20231020]).
This is supposedly accep ed wi hou ese a ion by he in e -
na ional communi y. Some commen s sugges ha “Is ael has
oolp oo eedom wi h he Wes . The poli ical eli e will no
condemn Is ael, no ma e wha i does” [“Is ael ha beim
Wes en na en F eihei . Die poli ische Eli e wi d Is ael egal
was sie mach nich e u eilen”] (JUERG-IN[20231017]),
ei e a ing he idea ha Is ael and/o Jews ha e a
manipula i e in luence (o e en absolu e con ol) o e
poli ics and he public: “We can ename ou sel es he
Fede al Republic o Is ael. O hand o e he login de ails
o he go e nmen - un social media channels o he IDF.
Then you sa e you sel he wo k o emails and phone calls”
[“Wi können uns umbenennen in Bundes epublik Is ael.
Ode übe gib gleich die Login-Da en de egie ungs-
ge üh en Social Media Kanäle an die IDF. Dann spa
man sich die A bei mi Emails und ele onie en”] (TAREK-IN
[20231021]). The p ominence o he alleged ee pass in
he Ge man co pus, compa ed o he Facebook analyses as
well as he B i ish and F ench analyses, could be explained
by he ac ha many o he p o-Pales inian h eads da e
om he end o Oc obe 2023 – a a ime when Is ael’s
mili a y in e en ions we e al eady ad anced. I seems ha
such commen s ake his as an oppo uni y o un es ic edly
a ibu e exagge a edly ep ehensible beha iou o Is ael,
and jus i y hese a ibu ions wi h he concep o a ee pass.
All hese asc ip ions, which a e used o condemn he ac ions
o Is ael, ul ima ely se e o legi imise he iolence o Hamas,
so ha he a i ma ion o iolence, which appea s p om-
inen ly in he en i e discou se om his pe iod, is ed by he
ne wo k o concep s desc ibed abo e:
“The yea s o inac ion and
cease i e ha e only led o inno-
cen Pales inians being mu de ed
and bombed almos daily and u he
unjus ly occupied, while he Wes
has p edic ably played i s h ee
monkey ole… Un o una ely he e
is only one language ha Is ael
and he Zionis s unde s and”
[“Die jah elange a losigkei und Wa ens ills and
ha nu dazu ge üh das as äglich unschuldige
Paläs inense e mo de bomba die und wei e
zu Un ech meh bese z wu den wäh end de
Wes en wie e wa e ih e 3 A en Rolle gespiel
haben.. Es gib leide nu eine Sp ache die
Is ael bzw die Zionis en e s anden we den”]
(JUERG-IN[20231017]).
2. The Hamas e o a acks on 7 Oc obe
33
Decoding An isemi ism
In o de o b oaden, di e si y and en ich he scope
o ou egula esea ch, we ca ied ou analyses o
six u he language communi ies and hei online
eac ions in he immedia e a e ma h o he Hamas
a acks.32 The six da ase s we e collec ed om he
Facebook pages o mains eam media ou le s in I aly,
Spain, Poland, Slo akia, Romania and Mo occo,
di ec ing ou ocus owa ds a icles pos ed be ween
7 and 10 Oc obe 2023 o examine he spon a-
neous eac ions o web use s o hese ini ial e en s.
Each da ase comp ises 1,000 use commen s (100
i s commen s om each o en h eads); we ol-
lowed he usual p ojec me hodology o anno a e
and classi y he da a. The addi ional six housand
commen s and hei analyses p esen ed below a e
no jus a b ie glimpse in o he unique discou ses o
each coun y, bu also p o ide ma e ial o obse -
ing he commonali ies o an isemi ic discou se ac oss
languages and cul u es. Wi h his ex ended con as-
i e dimension o he 7Oc obe e o a acks, ou
co pus analysis ep esen s he la ges in es iga ion
so a in o Is ael- ela ed an isemi ism online du ing
he con lic ’s cu en escala ion phase.
2.5.1 I aly
Lau a Ascone
Wi h only 9.7%, he I alian da a33 con ained
signi ican ly ewe an isemi ic esponses han
he o he language communi ies analysed in
his epo . Howe e , while he pe cen age o
an isemi ic commen s is lowe in he I alian da ase ,
he dis ibu ion o he concep s exp essed ollows
he o e all endency iden i ied in he o he coun ies.
The analysed commen s p esen ei he an an isemi ic
concep , ypically aimed a Is ael, o he exp ession
o suppo o Hamas, which – as s a ed be o e –
ep esen s a o m o an isemi ism. In he en h eads
unde in es iga ion, he pe cen age o an isemi ic
eac ions anges om 5% (Il Messagge o) o 17%
(Co ie e della Se a).
F equen ly, commen e s e e o is ael’s sole
guil in he con lic (19.6%), p esen ing i as he
only pa y esponsible o he e o is a acks. When
commen s accuse Is ael o sole esponsibili y o he
con lic , hey end o e e o he Is aeli ac ions which
supposedly led o o jus i ied such a iolen a ack:
“S op gi ing money o Is aelis, who p o oke Mus-
lims on a daily basis” [“Bas a da e soldi ai ascis i
is aeliani che quo idianamen e p o ocano i musulm-
ani”] (LIBER-FB[20231009]); “Is aelis a e paying o
he yea s o abuses on Pales inians… so y o
he dead… bu Is ael is he one esponsible o all
his bloodshed” [“Gli is aeliani pagano anni di
sop usi con o i Pales inesi ....dispiace pe i mo i...
ma Is aele ha u e le colpe di ques o sangue e -
sa o...”] (LASTA-FB[20231008a]). As men ioned
abo e, his an isemi ic concep some imes leads o
he jus i ica ion o Hamas’s a acks: “The s inke s
ha e been killing he beas s34 o se en y yea s...
In he end he beas s ha e eac ed. Easy…” [“Da
se an ’anni che le puzzolen e s anno uccidendo le
bel e ..Alla ine le bel e hanno eagi o . acile”]
(LIBER-FB[20231009]); “wha goes a ound comes
a ound, i ’s called ka ma” [“quello che di a si ice e,
si chiama Ka ma”] (LASTA-FB[20231008b]).
2.5 Quali a i e analyses o esponses in o he coun ies
32 – Schwa z-F iesel and Reinha z analysed an isemi ic s a emen s in le e s
and emails o Jewish and Is aeli ins i u ions ac oss se en Eu opean coun ies:
Aus ia, Swi ze land, The Ne he lands, Spain, Belgium, England, I eland, and
Sweden. Schwa z-F iesel, Monika/Reinha z, Jehuda, 2017. Inside he An isemi ic
Mind: he language o Jew-Ha ed in con empo a y Ge many. Chicago: Uni e si y o
Chicago P ess, 194–203.
33 – The I alian co pus consis s o he i s 100 commen s pos ed in eac ion o
10 a icles sha ed on he Facebook pages o i e I alian mains eam media
(La Repubblica, Co ie e della Se a, Il Messagge o, La S ampa, and Libe o).
Two h eads pe ou le we e analysed.
34 – The commen e uses he same e m used in he i le o he a icle. Howe e ,
hey use i in an i onic way, implying ha he eal beas s a e he Is aelis.
As we ha e seen in he p e ious sec ions, ins ead
o di ec ly a acking Is ael some use s a i m
he iolence o Hamas’s a acks (18.3%). Mos
o he ime, his is done h ough he exp ession o
suppo o Pales ine, which cons i u es he e a
me onym o Hamas and i s a acks and, as a con-
sequence, a o m o an isemi ism: he slogans “Go
Pales ine” [“Fo za Pales ina”] and “F ee Pales ine”
[“Pales ina libe a”] appea in se e al h eads. In
some cases, his o m o an isemi ism is e balised
in an explici way: “Go Hamas” [“W Hamas”]
(LIBER-FB[20231008]); “i Hamas coo dina es
he igh agains Is ael, may i come!” [“se Hamas
coo dina il con as o a Is aele, ben enga!”].
Commen e s jus i y he Hamas a ack by a ib-
u ing male olen , e il cha ac e o Is ael (16.5%).
They ei he demonise Is aeli ac ions hemsel es
(“Is ael has been bombing Pales ine o 70 yea s”
[“Sono 70 anni che Is aele colpisce la Pales ina a
suon di bombe sulla ci à”] (REPUB-FB[20231008]))
o compa e hem o he Hamas massac e on
7Oc obe , he eby bo h downplaying and legi -
imising he la e : “Is ael has been doing his
o 80 yea s...” [“Is aele lo a da 80anni...”]
(REPUB-FB[20231009]).
Ano he way o demonising Is ael is h ough
he nazi analogy (11.3%) which, again, is e -
balised in wo di e en ways: ocusing on Is ael’s
alleged na u e (“Is ael is Nazism,” [“Is aele è il
nazismo”] (MESSA-FB[20231007])) o compa -
ing Is ael’s ac ions o hose o he Nazis (“Tha ’s
inc edible how Jews who ha e su e ed Hi le a e
beha ing like hei pe secu o ” [“Inc edibile come
gli eb ei che hanno subi o Hi le si compo ano
come il lo o pe secu o e”] (CORRI-FB[20231009])).
In some cases, no only is Is ael compa ed o he
Nazis, bu p esen ed as being
e en wo se: “E en he Nazis didn’ ”
[“Roba che manco i nazis i”]
(REPUB-FB[20231008]).
Fu he mo e, Is ael is also desc ibed as a e o is
s a e (10.3%), which in some cases leads o he
denial o i s igh o exis (9.2%). The analysis shows
ha , h oughou he da ase , he ocus is on he demoni-
sa ion o Is ael mo e han on he p aise o Hamas.
2.5.2 Spain
Iael Ku jan
Wi hin he Spanish Facebook da ase , 17% o
analysed commen s we e classed as an ise-
mi ic.35 Much o he commen a y bo h explici ly and
implici ly jus i ied he Hamas a acks as a legi ima e
esponse o p e ious Is aeli ac ions – claiming he
iolence ep esen ed Pales inian sel -de ence, due
o Is aeli opp ession, in o de o ame he no ions o
ic im/opp essed s guil y/opp esso .
By a he mos equen ly exp essed an isemi ic
concep was he accusa ion o Is ael being a
e o is s a e (16.5% o all an isemi ic commen s).
Acco ding o such commen s, Is ael massac es o
ex e mina es Pales inians: “The e o is s a e o Is ael
exe cises he occupa ion and ex e mina ion o Pales ine”
[“el es ado e o is a de Is ael eje ce la ocupación y
ex e minio de Pales ina”] (ABC-FB[20231010]). Some
web use s combined he e o is s a e accusa ion
wi h indi ec o ms o he nazi analogy, e. g. by
equa ing Gaza wi h Auschwi z:
“ he e o is s a e o Is ael
is ca ying ou a o al ex e -
mina ion in Gaza, Gaza has
been con e ed in o Auschwi z”
[“el es ado e o is a de Is ael, eje ce
un o al ex e minio sob e Gaza, Gaza
la han con e ido en un Auschwi z”]
(ABC-FB[20231010]).
2. The Hamas e o a acks on 7 Oc obe
35 – 1,000 use commen s pos ed
on Facebook p o iles o i e Spanish
mains eam media ou le s we e anno a ed:
El País, CNN en Español, ABC España,
El Mundo and La Vangua dia.
35
Decoding An isemi ism
Ins ances o he nazi analogy comp ised 9.4% o all
an isemi ic commen s: “Jewish Zionis s epea wi h he Pal-
es inians wha hey su e ed a he hands o he Nazis” [“Los
sionis as judíos epi en con los pales inos lo que ellos su ie on
a manos de los nazis”] (ABC-FB[20231007]), o “You can
ell ha Is aeli es ha e lea ned a lo abou Nazism du ing he
Second Wo ld Wa . Then hey play ic ims a ound he wo ld”
[“Se e que los is aeli as ap endie on mucho del nazismo de
la Segunda Gue a Mundial. Luego an de íc imas po el
mundo”] (ELPAI-FB[20231007]). This commen uses i ony o
imply ha jews ha e no lea ned om he pas , while o he s
s a e ha Is aelis’ “DNA was con amina ed wi h Nazism” [“su
ADN se con aminó con el nazismo”] (LAVAN-FB[20231010]).
The hi d mos equen concep was he accusa ion ha
Is ael is commi ing genocide on he Pales inians (8.8%):
“cowa ds a e hose who en e a land ha is no hei s and
commi genocide, hey co ne hem and expel hem, Jews will
always be ha ed by humani y, e y b a e now bu in Be lin
in 1944 hey pooped hei pan s” [“coba des es el que se
me e en una ie a que no es suya y le hacen genocidio, lo
aco alan y los expulsa, los judíos siemp e se án odiados po
la humanidad, muy alien es aho a pe o en Be lín en 1944 se
hacían caca en los pan alones”] (ABC-FB[20231010]). As in
his case, many genocide commen s inco po a ed he nazi
analogy, along wi h e e ences o child abduc ion o
e en delibe a e child mu de (6.5%):
“I is wha he ‘chosen ones’
sough as God’s people; massac ing
ou Pales inian b o he s and wha
is wo se, imp isoning child en. No
one jus i ies iolence, bu hese
Is aelis a e wo se han Nazis”
[“Es lo que busca on los ‘elegidos’ como pueblo
de dios; masac ando a nues os he manos
pales inos y lo que es peo enca celando a
los niños. Nadie jus i ica la iolencia, pe o
es os is aelíes son peo es que los nazis”]
(ELPAI-FB[20231007]).
Al hough his Facebook commen does no pe se amoun
o child mu de , he commen e claims he delibe a e imp is-
onmen o child en and hus pe pe ua es a cen u ies-old
p ejudice.
I is also wo h no ing ha a conside able numbe o
commen s exp essed indi ec a i ma ion (15.3%) o
he Hamas a acks ia gene alised suppo o Pales-
ine, wi h slogans such as: “Resis , ee Pales ine” [“Resis e
Pales ina lib e”] (CNN-FB[20231007a]). O he s used he
Hamas a acks as an oppo uni y o deny is ael’s igh
o sel -de e mina ion (5.3%). An isemi ic commen s o en
con eyed he idea ha Zionis s a e in ade s and ha hey
mus lea e Pales inian lands – no jus he Wes Bank bu he
en i e egion. Is ael was widely e e ed o, in essen ialised
e ms, as an “occupie .” Th oughou he da ase , he e was
a clea associa ion be ween he gene alised use o ‘occu-
pie ’ and e ms such as ‘in asion’: “Is ael, he in ade and
colonialis in he 21s cen u y” [“Is ael in aso y colonialis a
en pleno siglo XXI”] (ELPAI-FB[20231009]), o : “Is ael is he
coun y ha in aded Pales ine” [“Is ael es el país que ha
in adido Pales ina”] (ELMUN-FB[20231009]).
2.5.3 Poland
Ka olina Placzyn a
In he Facebook commen s sec ions o Polish
news ou le s, an isemi ism le els a e aged
17.7%, eaching a high o 38% unde a news pos
desc ibing he e en s a he Supe no a es i al.36
In bo h his h ead and he en i e da ase , e enly
dis ibu ed ac oss all ou le s, he op concep was
conspi acy (19.8% o all an isemi ic commen s),
as web use s e used o accep Hamas a acks
had happened wi hou Is ael’s aci ag eemen
o e en ins iga ion: “This ac ion is a p o oca ion
by Mossad, he e was a h ea o ci il wa in Is ael
so hey had o eac ” [“Ta akcja o p owokacja
Mosadu, w Iz aelu g oziła wojna domowa więc
zeba było za eagować”] (FAKT-FB[20231008a]).
Such commen s also linked he supposed “p o -
oca ion” wi h a wish o “des oy Pales ine... who
c ea ed hamas i no is ael...” [“zniszczyc pal-
es yne... k o s wo zył hamas jak nie iz ael...]
(INTER-FB[20231008]). O he commen s in his
ca ego y o en e e enced he ‘Hea enly Je usalem’
p ojec – a Jewish ese lemen conspi acy heo y:37
“Jus as planned. Uk ainians
will lood Poland and Jews
will se le in Uk aine.
The global plan is slowly
coming o ui ion
”
[“Czyli ak jak miało być .
Uk aińcy zalewają Polskę,
a Żydzi zamieszkają na
Uk ainie.Globalny plan
pomału się ealizuje
”]
(RMF-FB[20231008]);
“Ano he alse lag pe o med by he sidelocks.
I eckon hey’ll make us in o a second Pales ine
now” [“Kolejna als lag w wykonaniu pejsa ych.
Mysle że z nas e az z obią 2 Pales yne”] (FAKT-
FB[20231008b]). The condescending e e ence o
Jews/Is aelis used in he la e example e lec s he
ela i ely high sha e o slu s p esen in he co -
pus (4.5% o an isemi ic commen s), compa ed o
he UK, F ance o Ge many.
The second mos common an isemi ic ope in
he co pus (18.6%) also accused Is ael o ne a -
ious in en , po aying i as e il o immo al –
ei he h oughou i s en i e his o y: “Is ael has been
mu de ing Pales inians o yea s, he wo ld is silen !!”
[“Iz ael od la mo duje Pales yńczyków, świa mil-
czy!!”] (POLSAT-FB[20231007]), o on a global
scale: “Is ael is he cause o all e il in he wo ld”
[“Iz ael o p zyczyna całego zła na świecie”]
(ZET-FB[20231008]). This some imes spilled o e
in o claims ha Is ael is a e o is s a e: “The e
a e no bigge e o is s han hose unde he S a
o Da id. A ipe ibe” [“wiekszych e o ys ów
niż spod gwiazdy Dawida o nie ma. Żmijowe ple-
mię”] (RMF-FB[20231008]), he e ein o ced wi h a
dehumanising insul and i s biblical conno a ions
o eache y o disloyal y ( hough such eligious
e e ences we e a e, wi h commen s o en a acking
eligion as a sou ce o con lic o igno ance, echoing
he cu en an i-Ca holic sen imen s in he socie y).
Some cons uc ed Is ael as genocidal “on a pa
wi h he Hi le i es. The same me hods o ex e mina-
ion” [“na owni z hi le owcami. Te same me ody
eks e minacji”] (POLSAT-FB[20231007]). The nazi
analogy (9% o an isemi ic commen s) was
exp essed wi h a me e allusion by his eade o he
le -leaning OKO.p ess “Adol ’s diligen s uden s”
[“Pilni uczniowie Adol a”] (OKO-FB[20231008]),
and a di ec accusa ion in a use commen om he
cen e- igh Wp os : “Li e ally a slow Holocaus . Bu
hey’ e Jews, and Jews a e allowed o” [“Dosłownie
polowolny holokaus . Ale o żydzi, a żydom
2. The Hamas e o a acks on 7 Oc obe
36 – The co pus con ains commen s om he
Facebook accoun s o he daily newspape
Fak , he weekly magazine Wp os , online
news pla o m In e ia, in es iga i e jou nalism
pla o m Oko.p ess, TV s a ion Polsa and
wo adio s a ions: RMF FM and Radio Ze .
37 – Simila claims ha e appea ed in e. g.
Romanian discou se (c . 2.5.5) and Ge man
a - igh spaces online: h ps://www.isd-
global.o g/digi al_dispa ches/an-an isemi -
ic-conspi acy- heo y-is-being-sha ed-on- el-
eg am- o-jus i y- ussias-in asion-o -uk aine/
(las accessed on 15Janua y 2024).
37
Decoding An isemi ism
wolno”] (WPROST-FB[20231008]). He e, he sa -
cas ic punchline hin s a he s e eo ype o a ee
pass supposedly enjoyed by Jews and/o Is ael
(6%); such commen s a e a clue o he linge ing
pe cep ions o Jews as un us wo hy and p i ileged,
despi e he con empo a y a emp s o dismiss an i-
semi ism as ei he a hing o Poland’s pas , o only
e e a p oblem o o he na ions.
Alongside he conspi a o ial explana ion
behind Hamas a acks, some commen s amed
hem as simply well-dese ed, blaming Is ael
ei he o he a acks o o he whole con lic
(13% o all an isemi ic commen s). This ca e-
go y pa icula ly s ongly co ela ed wi h a lack o
empa hy (al hough his was ound in non-an isemi ic
commen s oo), he e succinc ly con eyed ia he
idiom “[w]ho sows he wind eaps he s o m” [“K o
sieje wia , en zbie a bu ze”] (FAKT-FB[20231009]).
Some web use s cynically a i med he an isemi ic
iolence (7% o an isemi ic commen s): “Is ael
has li ed o eg e hese decades o mu de ing Pales-
inians. Lo d, you’ e no swi , bu you a e igh eous!”
[“Doczekał się Iz ael za e dziesią ki la mo dowa-
nia Pales yńczyków. Panie Boże jes eś nie ychliwy
ale za o sp awiedliwy!”] (POLSAT-FB[20231007]).
Some commen s in his ca ego y sugges ed he
a acks would no ha e happened had Is ael no
“c ea ed a ic i ious s a e on Pales ine’s land, his
illegal c ea ion was legalised hanks o he USSR
and he USA” [“s wo zył ikcyjne pańs wo na e e-
nach nalezacych do pales yny I dzięki ZSRR o az
usa en nielegalny wó zos ał zalegalizowany”]
(INTER-FB[20231008]). Ano he commen sim-
ila ly denies is ael’s igh o exis , pain ing a
con emp uous pic u e o i s ci izens as e e nally
homeless and alien “[ ]he e a e no Is aelis, he e
a e Jews. They ha e no hing o hei own, because
whe e e hey appea ed hey ha e been chased
away” [“nie ma Iz aelczyków, są Żydzi. Nie mają
nic swojego, bo gdziekolwiek się zjawili o ich p ze-
ganiano”] (ZET-FB[20231008]).
2.5.4 Slo akia
Ve onika Bundzíko á
The a e age le el o an isemi ism in he Face-
book commen s o Slo ak media ou le s s ands
a 14.9%, e en hough i eaches up o 23.7% in
some h eads.38 Compa ed wi h some Wes Eu o-
pean coun ies, p o-Pales inian commen s ha would
in ac disguise he suppo o Hamas’s a ack only
a ely occu in he commen s sec ions o he ana-
lysed Slo ak media ou le s. Also, he explici suppo
o he a oci ies pe pe a ed by Hamas is p esen in a
e y limi ed scope; only one use eac s by pos ing:
“G ea , inge s c ossed o Pales ine, i a Pales ine”
[“Supe d žím palce Pales íne, i a Pales ína”]
(NOVYC-FB[20231008]). Ins ead, he p o-Pales-
inian sympa hies a e disclosed as a by-p oduc o
Is ael’s ili ica ion. They usually su e om o e sim-
pli ica ion o he Middle Eas e n poli ical con ex ,
aming he Is aeli-Pales ine con lic in a black-and-
whi e manne . Is ael is po ayed as he immo al
agg esso (6%) o a p i ileged s a e ha e ades
any c i ique o i s misconduc hanks o holding
a ee pass o i s ac ions (10%).
Occasionally, pa allels be ween he Russian-
Uk ainian wa and he si ua ion in Is ael a e
d awn in he Slo ak an isemi ic commen s. This migh
be enhanced by he ac ha Slo akia is Uk aine’s
neighbou and has been a ec ed by he wa
looming a i s bo de s. One use s a es: “When he
Russians ook o e Uk ainian soil, he whole wo ld
condemned hem bu nobody condemned Is ael,
some hing s inks he e” [“Keď Rusi zab ali Uk ajine
územie sko o cely s e ich odsúdil Iz ael nie niečo
u sm dí”] (NOVYC-FB[20231008]). Also, a
hand ul o commen s (5%) equa e he si ua ion
in Pales ine igh a e he Is aeli coun e a ack
wi h he su e ing o Jews du ing he Holocaus ,
which ep esen s an indi ec nazi analogy: “One
F ench jou nalis no ed ha i eminds him o he
38 – The Slo ak
da ase con ains
commen s eplying o
pos s published by he
ollowing media ou le s:
Denník N, SME,
Ak uali y, P a da,
Plus Jeden Deň, Topky
and No ý Čas.
liquida ion o he Jewish ghe o in Wa saw by he ascis s!”
[“Jeden ancúzsky no iná sa yjad il, že mu o p ipomína
lik idáciu žido ského a ša ského ghe a ašis ami!”]
(AKTUA-FB[20231009]).
A he concep ual le el, s ongly ep esen ed a e
he a acks based on denying Is ael’s s a ehood
and he Jewish igh o sel -de e mina ion. This idea
is exp essed in 40% o he commen s classi ied as
an isemi ic. The use s ei he deny Is ael’s his o ical igh
o he land (“Un o una ely, Is ael s ole his land om Mus-
lims. They used o li e he e in ha mony wi h Ch is ians.
Then Jews came, and ha was he end” [“bohužiaľ Iz ael
moslimom uk adol u o zem. Oni si am kedysi s k esť-
anmi súlade žili. Po om p išli židia a bol o ich koniec”]
(PLUSJ-FB[20231009])), o ame he Is aelis as occupie s
(“Jews ha e been occupying hese lands and killing Pales in-
ians o decades, and he wo ld is quie ”) [“Židia desať očia
okupujú územia a aždia Pales ínčano a s e je icho”]
(NOVYC-FB[20211008]). Some conspi acy-led com-
men s (4%) highligh he ole o he USA in he es ab-
lishmen o Is ael, which would no ha e exis ed o he wise.
Such commen s, hough a e, a e closely ela ed o
he accusa ion ha Is ael is a o eign o disin eg a ing
elemen in he egion (10%): “because hey a e na i e, in
con as o he ew millions o hese Talmudic pseudo-He-
b ews ga he ed om all a ound he wo ld, who occupy i
he e wi h hei a i icially es ablished s a e Is ahell” [“lebo
od iaľ pochádzajú, na ozdiel od ých niekoľko milióno
almudických pseudoheb ejco pozbie aných z celého s e a,
k o í o am okupujú ako bol umelo y o ený š á Iz ahell”]
(PRAVD-FB[20231008]).
Fo use s pos ing such commen s, Is ael has no jus i ica ion
o i s exis ence and hence is no wo h any pi y a e he
a ack on 7 Oc obe . Such a gumen a ion is augmen ed
by he belie ha he Jews (o Is ael) a e he ul ima e
e il (18.8%). These s a emen s o e blow he du a ion o
he scope o hei misdeeds, wi h he ollowing example
demons a ing how an i-Is aeli and an i-Jewish sen imen s
may o e lap: “They occupy his land and opp ess he Pal-
es inians since 1947! This should s op, and Jews should
s op sp eading e il in he wo ld as hey ha e been doing
since he Old Tes amen un il nowadays!” [“Od oku 1947
okupuju o o uzemie a u lačaju pales inčano ! To o by mali
skinci a židia by msli p es a ši i zlo po s e e od s a eho
zakona po dnes!”] (PLUSJ-FB[20231009]). This goes hand
in hand wi h ano he an isemi ic concep ha appea s in he
commen s: blaming jews o an isemi ism (16%) o solely
is ael o he con lic (12.8%), usually accompanied
by lack o empa hy and schaden eude. In such com-
men s, use s p e end o pu he Hamas a ack in o he bigge
pic u e bu end up swapping he pe pe a o s o he ic-
ims: “I ’s ho ible. Bu hones ly, Is ael has been pe pe a ing
genocide o he Pales inian na ion o decades. So, un o -
una ely, we could ha e expec ed his o happen” [“Je o
s ašné. Ale pop a de, Iz ael pácha genocídu pales ínskeho
ná oda celé desať očia. Takže sa o žiaľ dalo očaká ať”]
(PLUSJ-FB[20231009]). Las ly, some commen s sugges ing
a sec e conspi acy behind he Hamas a ack ha bou an i-
semi ism: hey accuse Is ael o Jews o being behind he
a oci ies o decei ully p o i om hem (9.4%):
“Maybe i ’s no ailu e o he
in elligence se ices bu a pe ec
accomplishmen . Now, hey can
use he a ack o e adica e ‘wi h
impuni y’ he Gaza S ip om he
map o e e ”
[“Mozno o obec neni ziadne zlyhanie ajnych
sluzieb, ale naopak dokonaly uspech. Te az
pod zamienkou napadnu ia mozu ‘bez es ne’
ygumo a pasmo gazy z mapy na zdy”]
(PROJE-FB[20231007]).
2. The Hamas e o a acks on 7 Oc obe
39
Decoding An isemi ism
2.5.5 Romania
Alexis Chapelan
In he Romanian da ase ,39 18.2% o all com-
men s we e classi ied as an isemi ic. This pe -
cen age showcases ha he Middle Eas con lic is
a signi ican igge o an isemi ism on Romanian
social media. Howe e , he na u e o he discou se
di e s om o he Eu opean con ex s: Is ael- ela ed
concep s a e less p ominen ; ins ead, classical
an isemi ic opes as well as slu s highligh a unda-
men al con inui y be ween olde pa e ns o demoni-
sa ion and o he ing o Jews and pe cep ions o Is ael
as a ‘collec i e Jew.’
This is e lec ed, o example, in he widesp ead
use (in almos 30% o all an isemi ic commen s) o
he e m ‘Jews’ o desc ibe Is aelis. This con la ion
o poli ical and e hnic/ eligious iden i y no
only enac s a o m o collec i e esponsibili y
o Jews o he Is aeli go e nmen ’s policy, bu
ans e s he nega i e image y linked o Jews on o
Is ael. Slu s his o ically e e ing o Jews (such
as ‘jidan,’ oughly ansla ed as ‘kikes’) a e also
applied o Is aelis mo e equen ly han in o he
coun ies (a ound 5% o all an isemi ic commen s).
Concep s o dehumanisa ion and epulsi eness
o en wo k in syne gy: “His o y augh us hey a e
no hing bu a bunch o pa asi es, e en i uses ha e
hei p ide, Jews don’ ” [“Iis o ia ne-a in a a ca nu
sun doa o sleah a de pa azi i!In compa a ie, i usii
au mand ia lo ,e eii nu!”] (STIRI-FB[20231007]).
One use desc ibes Is aelis in ways eminiscen o
19 h cen u y eugenics, ocusing on g eed, lack o
cleanliness and diso de ly sexuali y: “They make
6–7 kids, hey a e loud, di y, hey li e e e y-
whe e hey go, hey wan e e y hing cheape o
e en o ee!! You should ac ually lis en o hose
who had o deal wi h his people called Jews!!”
[“Fac ca e 6 -7 copii, san galagiosi, san mu da i
si lasa mize ie pe unde ec, si o o ul ca mai
ie in sau chia pe degeaba!! Voi cei ca e o bi i
din ibunal a i bine sa ascul a i oamenii ca e au
a u angen a cu aces neam numi e ei!!”]
(STIRI-FB[20231007]).
Jews a e also accused o ha ing had a disin e-
g a ing impac on socie y h oughou his o y:
“Wipe Jews ou he ace o he ea h, hey des oyed
Romania” [“Săi adă de pe ața pămân ului pe
e ei au dis us România”] (RO.EU-FB[20231007]);
“Fuck o , scum! You des oyed en i e coun ies and
peoples! You will pay o i , scumbags!” [“Să ă ia
d acul de neno ociți ! Ați dis us ță i și popoa e ! O
să plă iți oi neno ocițilo !] (RO.EU-FB[20231007]);
“They a e no wi h God bu wi h he De il, ha ’s
why hey only wan wa s o des oy humani y”
[“Ad ia nu mai sun cu Dumnezeu și cu d acu deaia
o numa ăzboaie și sa dis ugă omeni ea...”]
(FB-DIGI[20231009]).
Cu en ac ions o he Is aeli go e nmen a e being
esi ua ed in o a b oade his o y o alleged Jewish
e il and c uel y, o en h ough eligious e -
e ences, hus c ea ing a clea na a i e ha such
iolence is on ologically ied o Jewish na u e: “You
will see how gene ous Jews a e, cheap p opaganda,
e e since Moses b ough hem om Egyp hey
massac ed all na ions li ing in he P omised Land,
a leas ha ’s wha he Bible says” [“o sa edeți
bună a ea e eilo , p opaganda de doi lei, de
când i-a adus Moise din Egip au macela i oa e
popoa ele ca e ăiau in a a ăgăduinței cel puțin
asa sc ie în Biblie”] (STIRI-FB[20231009]);
“You a e a clueless ool, Is ael is and will o e e be
a bunch o mu de ous immig an s. [...] Don’ o ge
who mu de ed Jesus! I wasn’ he Pales inians!”
[“Nu s i i nimic o bi i mul i sa a a la i in eaba,
is ael a i si a amane nis e imig an i c iminali […]
Si nu ui a cine la omo a pe isus! Nu ai os pales-
ine!”] (STIRI-FB[20231007]).
39 – The co pus
comp ises 1,000 use
commen s sampled om
he Facebook pages o
eigh popula Roma-
nian news ou le s in he
immedia e a e ma h o
he 7Oc obe a ack:
An ena3, Digi 24,
P o TV, S i i pe Su se,
Romania TV, Ade a ul,
Eu onews, and Ho news.
Simila o he Polish da a, he analysed h eads
also showcase a wide ange o an isemi ic
conspi acy heo ies (29% o all an isemi ic
commen s), om he adi ional allega ion o global
domina ion o he Khaza claim: “This is wha Is ael
always did, ind ways o ule he wo ld – i ’s in hei
eligion – and s oke up wa s” [“As a a s iu Is aelul
sa aca oa a ia a,sa cau e mijloace sa conduca
lumea – e in eligia lo – si sa p o oace azboaie !”]
(RO.EU-FB[20231007]); “Tha ’s why hey p o oked
he wa in Uk aine, so i is abandoned by Uk ainians
and hese Khaza s come and build he new Je usalem”
[“De aceia au p o oca așa zisul ăzboi în Uk aina
ca să ie pă ăsi ă de u ceainieni și să ină ei să con-
s uiască noul Ie usalim aceș i kaza i.”] (STIRI-FB
[20231009]). A p ominen conspi acy heo y also
asse s ha he 7 Oc obe massac e was an
Is aeli alse lag ope a ion o manipula e public
opinion in o accep ing he wa : “A cheap se -up, o
jus i y o he wo ld he expansion o sa anism” [“O
înscena e o dina ă ca e să jus i ice în ața omeni ii
expansiunea sa anismului”] (RO.EU-FB[20231007])”.
One use d aws a pa allel wi h Pea l Ha bo , hus
implying Jews we e also behind Ame ica’s in ol e-
men in Wo ld Wa II: “To jus i y hei a ack on I an
and p epa e and manipula e public opinion he Jews
needed ha . Some hing like Pea l Ha bo in he Sec-
ond Wo ld Wa ” [“Pen u a mo i a a acul impo i a
I anului si a p ega i si manipula opinia publica in
a oa ea azboiului e ei a eau ne oie de as a. Ce a
ca in al II -lea azboi mondial la Pea l Ha bo !”]
(STIRI-FB[20231007]). Some commen s also exp ess
he ea ha i Is ael is de ea ed Jewish people will
ese le and occupy Romania o Uk aine: “I see hem
massi ely coming o Romania, wi h he 600,000
which ecei ed Romanian ci izenship om he ai o
[P esiden o Romanian Sena e] Ciucă. I wonde i all
o hem a e going o come he e e en ually, i ’s ha d
wi h he A abs bu Romanians a e aw ully docile”
[“Vad ca in masi in Romania din e cei 600 000 de
e ei ce au p imi ce a enie omana de la gene alul
ada o Ciuca. Ma in eb daca pina la u ma in o i,
ca e g eu cu a abii, omanii insa sin p os i de docili”]
(STIRI-FB[20231007]).
2.5.6 Mo occo
Mohamed Salhi
While Mo occo does no sha e he same Jewish pe -
secu ion his o y as Eu ope, and despi e he ich Jewish
cul u e in Mo occo and he ins i u ional e o s o
engage wi h his his o y,40 i does su e om a signi -
ican an isemi ism p oblem. Among he explana ions
o con empo a y an isemi ism in Mo occo, a gues
Zho Rehihil, he di ec o o he Museum o Mo occan
Judaism, a e he lack o unde s anding and igno ance
o Jewish his o y, he i age, and cul u e in Mo occo.41
O he ac o s include he cen ali y o he Pales inian
cause o he popula and poli ical memo y,42 he
his o ically impo ed43 and pos -Zionis an isemi ism
(i. e. memo y o he Nakba and In i ada),44 and he
owned-upon Ab ahamic acco ds.45
2. The Hamas e o a acks on 7 Oc obe
40 – “U.S. Special En oy o Comba ing An i-Semi ism Commends Mo occan Model
in Te ms o In e ai h Dialogue,” Agence Ma ocaine de P ess, 15 No embe 2023
h ps://www.mapnews.ma/en/ac uali es/social/us-special-en oy-comba ing-an i-
semi ism-commends-mo occan-model- e ms-in e ai h (las accessed on 8 Janua y 2024).
41 – Bou chachene Wail, ”An i-Semi ism in Mo occo: A Complica ed Issue,”
Mo occo Wo ld News, 11 Ap il 2016 h ps://www.mo occowo ldnews.com/
2016/04/184021/an i-semi ism-in-mo occo-a-complica ed-issue
(las accessed on 8 Janua y 2024).
42 – Alexand e Aublanc, ”Au Ma oc, l’ancien p emie minis e islamis e
Abdelilah Benki ane dé ape dans l’an isémi isme,” Le Monde, 24 No embe 2023
h ps://www.lemonde. /a ique/a icle/2023/11/24/au-ma oc-l-ancien-
p emie -minis e-islamis e-abdelilah-benki ane-de ape-dans-l-an isemi-
isme_6202163_3212.h ml (las accessed on 8 Janua y 2024).
43 – The Documen a ion Cen e o No h A ican Jews, The Jews o Mo occo
du ing Wo ld Wa II, h ps://no ha icanjews-ww2.o g.il/en/jews-mo occo-du -
ing-wo ld-wa -ii (las accessed on 8 Janua y 2024).
44 – Dominique Vidal, “Quand « Le Jou nal » dénonce l’an i sémi isme,”
Le Monde Diploma ique, June 2022 h ps://www.monde-diploma ique.
/2002/06/VIDAL/8946 (las accessed on 8 Janua y 2024).
45 – Alexand e Aublanc, “Gue e Is aël-Hamas : au Ma oc, les
i aillemen s de la
communau é jui e,” Le Monde, 5 Decembe 2023
h ps://www.lemonde. /a ique/
a icle/2023/12/05/gue e-is ael-hamas-au-ma oc-les- i aillemen s-de-la-commu-
nau e-jui e_6204047_3212.h ml (las accessed on 8 Janua y2024).
47
Decoding An isemi ism
In No embe 2023, IBM announced ha i s ad e -
isemen s had been ound alongside Nazi pos s
and an isemi ic commen s on X.59 The o ganisa ion
Media Ma e s o Ame ica highligh ed ha ads om
so wa e gian s such as IBM, Apple, and O acle
appea ed on X nex o pos s wi h posi i e s a e-
men s abou Adol Hi le and Nazi ideology. IBM
decided o pause ad e ising on X – a decision ha
was ollowed by Apple and he media companies
Disney, Pa amoun , Wa ne B os and Lionsga e. The
EUCommission also announced ha i would empo-
a ily suspend ad e ising on X, wi h sha p c i icism
coming also om he Whi e House. Musk e e ed
o Media Ma e s as “pu e e il” and h ea ened a
“ he monuclea lawsui .”60
This co pus consis ed o 500 commen s aken om
he YouTube channels o bo h mains eam and
Conclusion
These explo a o y s udies once again unde sco e he pi o al ole
o opinion leade s like Elon Musk in he esu gence o commonly
exp essed an isemi ism. Mos o he commen s exhibi an a i ma-
ion o he an isemi ic wo ld iew exp essed in he ini ial pos , man-
i es ed h ough s e eo ypes o conspi acy heo ies. Al e na i ely,
commen e s deny he an isemi ic cha ac e o he wo ld iew, he eby
seeking o legi ima e and a i m i in an in e ed sense. Bo h can
be seen as a b idge o p ejudices agains Jews, independen ly
exp essed by he commen e s, as obse ed in ou analyses in he
o m o s e eo ypes, such as e il, disin eg a ion, powe , o he ins u-
men alisa ion o an isemi ism. In some o he examined YouTube com-
men s sec ions (no ably channels om mains eam news ou le s),
he language used by commen e s could ha dly be dis inguished
om he o ic wi hin he Whi e Sup emacy spec um, s udied by
o he esea ch p ojec s, illus a ing he condi ions o ‘mains eam-
ing’ openly ha e ul and exclusiona y ideas.
3.4 Media Ma e s o Ame ica, IBM, The Whi e House
al e na i e media.61 O hese, 16.8% we e classed
as an isemi ic. The wo mos p ominen concep s
we e he denial o he an isemi ic cha ac e o
Musk’s commen s (35.7%) and hei a i ma ion
(35.7%). A imes, hese wo eac ions occu in com-
bina ion: “Well, he does no ca e, and he is NOT
an i-semi ic jus because he wan s jus ice and no
singing he Song o he Zionis s” (SKY-YT[20231118]).
Elsewhe e, a commen combines a denial wi h a
blun allega ion ha an isemi ism is being ins u-
men alised (7%): “An onslaugh o exagge a ed,
some imes wholly baseless accusa ions o
an isemi ism has p o en o be b u ally e ec i e”
(MSNBC-YT[20231118]). Howe e , a i ma ion
was o en exp essed indi ec ly, h ough wide
app o al o Musk o his companies: “I ne e
hough o buying Tesla bu I hink I wan one now’
(CBSE-YT[20231118]).
59 – Wendling, Mike, 2023.
IBM suspends ads on X a e hey appea ed
nex o Nazi pos s, BBC, 17 No embe 2023,
h ps://www.bbc.co.uk/news/
wo ld-us-canada-67446797
(las accessed on 21Decembe 2023).
60 – Kolodny, Lo a, 2023. Musk h ea ens
“ he mo -nuclea lawsui ” agains media
wa chdog, calls ad e ise s “opp esso s.”
CNBC, 18 No embe 2023,
h ps://www.cnbc.com/2023/11/18/
musk- h ea ens- he monuclea -law-
sui -agains -media-wa chdog-calls-
ad e ise s-opp esso s.h ml
(las accessed on 21 Decembe 2023).
61 – Th eads we e aken om he YouTube
channels o BBC, MSNBC, Sky News,
Fox News, CSN E ening News.
4. Ou wo kshop se ies on web- ela ed
ha e s udies: conclusions and
akeaways Ka olina Placzyn a and Ma cus Scheibe
In he au umn o 2023, he Decoding An isemi ism p ojec
eam an se en online wo kshops on con empo a y online
ha e speech. The se ies was concei ed as a means o e lec
he me hods and aims o ou in e disciplina y p ojec , es ab-
lish a dialogue be ween ou esea ch expe iences and o he
expe app oaches, and open he deba e on he challenges
o ha e speech in a ious a eas o academic and non-aca-
demic p o essional p ac ice.
Wo kshop 1:
Quali a i e analyses o online ha e
speech and he challenge o g ey
a eas o meaning
In he opening e en o he se ies, mode a ed by Alexis
Chapelan (Decoding An isemi ism), we ocused on he ques-
ion o dis inguishing be ween ha e speech in i s he e oge-
nous o ms on he one hand, and legi ima e c i ique o social
s uc u e, inequali y, o human igh s iola ions on he o he .
In his alk, P o esso Joachim Scha lo h (Waseda Uni e -
si y) p esen ed he key concep s o denig a ion and in ec i -
i y, discussing he ea o social de alua ion, exclusion and
losing ace which d i e ep oduc ion o he social o de , bu
also esul in dispa aging speech owa ds he o he . P o esso
Vic o ia Guillén Nie o (Uni e si y o Alican e) examined
legally ac ionable in en ionali y h ough Speech Ac Theo y
and Rele ance Theo y, using he case o Uni ed S a es .
Wilcox o illus a e he esponsibili y o bo h he speake
and he hea e in ( e)cons uc ing he in ended meaning.
D Ma hias J. Becke (p ojec lead o Decoding An isemi-
ism) p esen ed he p ojec ’s axonomies, which help dis in-
guish be ween legi ima e c i ique and essen ialising o gen-
e alising ha e speech, using as an example he ela ionship
be ween an isemi ism and an icapi alism in he discou se
a ound Geo ge So os, as well as he dis inc ion be ween
c i icism o Is ael and Is ael- ela ed an isemi ism. In deciphe -
ing implici ness, p agma ic linguis ic app oaches in ol e
language, con ex , and wo ld knowledge. Ag eemen
among he expe s ex ended o obse a ions ha ha e
speech con eyed h ough indi ec ness and implici ness
poses a signi ican ly g ea e h ea han i s di ec and explici
coun e pa , as i can pene a e di e en milieus.
Wo kshop 2:
Discou se analy ical app oaches
o online ha e speech
The second wo kshop, mode a ed by Ma cus Scheibe
(Decoding An isemi ism) add essed online ha e speech
om he pos -disciplina y pe spec i e o discou se analy i-
cal app oaches, conside ing i s o ms as well as he social
changes impac ed by and impac ing on i s sp ead in he
ields o a , cul u e o poli ics. P o esso Benno He zog
(Uni e si y o Valencia) used he Spanish con ex o discuss
he con inui y and change in an isemi ic me apho s and
seman ics in he in o ma ion a chi ec u e o online spaces:
po en ially mo e democ a ic bu wi h lowe h esholds o
u h. Discou se s udies examine online an isemi ism on sym-
bolical and ma e ial le els and o e s a egies o coun e
i . P o esso Johannes Ange mulle (The Open Uni e si y)
used he UK case o an isemi ism accusa ions agains Je emy
Co byn o demons a e an isemi ism as a p oduc o linguis ic
and non-linguis ic (social) p ac ices and sys ems. He high-
ligh ed how a p es igious and highly isible posi ion no only
enables and ampli ies he impac o a poli ical in e en ion
bu is also c ucial in cons i u ing an ideological e ec such as
an isemi ism. In his iew, an isemi ism is no inhe en in ex s,
bu esul s om ex s used in a sys em o social p ac ices and
ela ionships; he e o e, an isemi ism canno be unde s ood
49
Decoding An isemi ism
independen ly om he social dynamics in which he ideo-
logical meanings, con en s and in en ions o he ac o s a e
cons uc ed. D Monika Kopy owska (Uni e si y o Łódź)
explo ed he e ec s o echnology and dis ance dynamics
on digi al meaning-making p ocess, and he implica ions o
he p oduc ion, dis ibu ion and consump ion o ha e speech.
She concluded ha he ans-spa ial, ans- empo al and ech-
no-discu si e cha ac e o online discou se shapes he social
impac o ha e speech, gi ing i a la ge pool o a gumen s
and a s ong emo ional po en ial. All speake s emphasised
ha discou se analysis as a me a- e lexi e app oach lends a
nuanced quali y o con en ional quali a i e analyses, as he
esea che ’s own s andpoin is also included in he analysis.
Wo kshop 3:
Quan i a i e and s a is ical analysis
o online ha e speech
In he hi d wo kshop, mode a ed by D Ildikó Ba na
(Eö ös Lo ánd Uni e si y), we explo ed how quan i a i e
and s a is ical analysis can shed ligh on, and expand he
ocus o , quali a i e analyses based on anno a ed co po a
o online da a. Daniel Miehling (Technical Uni e si y
Be lin) opened he discussion by p esen ing quan i a i e
and s a is ical analyses o con e sa ions on he mains eam
pla o m Twi e (now X) as well as in inge communi ies
on Teleg am. Using a ange o ools, including eg ession
models, he showed ha an isemi ism has inc eased o e
ime on bo h pla o ms, al hough hey di e signi ican ly in
e ms o he numbe o use s, size and olume o con en
gene a ed. An isemi ic con e sa ions abou Jews and Is ael
on X show a wide ange o di e en use s dissemina ing
such con en , each p oducing only a ew messages, while
Teleg am use s belonging o inge communi ies a e no
only mo e likely o p oduce mul iple messages, bu also
dissemina e hese messages mo e equen ly. D Julian
Ha g ea es (Wool Ins i u e) p esen ed measu emen s o
an isemi ism ac oss h ee s udies in o Google, Ins ag am
and Twi e in 2020–2021, commissioned by he Commu-
ni y Secu i y T us and An isemi ism Policy T us . His eam
ound high p esence o isual and ex ual an isemi ism on
all h ee pla o ms, including a p oli e a ion o Is ael- e-
la ed an isemi ism (pa icula ly conspi acy heo ies) on
Ins ag am. They concluded ha se ices such as Google’s
Sa eSea ch we e no su icien ly e ec i e, p oducing bo h
alse nega i es and alse posi i es. S e an Munnes (WZB
Be lin Social Science Cen e ) explo ed an isemi ic dynam-
ics in online discussions using Decoding An isemi ism’s
anno a ed da ase s, compa ing and con as ing da a om
di e en pla o ms, pe iods and coun ies. His analyses
e ealed a pa e n in which ini ial commen s end o be
mo e implici , inc easingly leading o explici an isemi ism
and igge ing u he and mo e h ea ening eplies. The
complimen a y esea ch app oaches o he h ee speake s
showed he p omise o quan i a i e and s a is ical analysis
when combined wi h a quali a i e elemen in he p ocess
o human anno a ion be o e he analyses, bo h expe and
non-expe , in o de o cap u e he implici and con ex ual
exp essions o an isemi ism.
Wo kshop 4:
No malisa ion o ha e speech
in online spaces
The ou h mee ing, mode a ed by D Ma hew
Bol on (Decoding An isemi ism), conside ed he
accep abili y and mains eaming o ha e speech,
i s e olu ion in he ecen decades, he inc easing
pola isa ion, and he ole o in e ac i e media in
hese p ocesses.62 Susan Benesch (Ha a d Uni-
e si y & Dange ous Speech P ojec ) showcased he
phenomenon o i ue alk: language ha con inces
he in-g oup ha hey a e c ucial, honou able mem-
be s o socie y du y-bound o p o ec women, chil-
d en, cul u al iden i y e c. (using as an example he
social media discou se o pa icipan s in he a ack
on he US Capi ol on 6 h Janua y 2021). As a posi-
i e, inwa dly di ec ed discou se, i is no immedia ely
iden i ied as ha m ul, and he e o e is pe missible in
public discou se. Mu ilo Camb uzzi (Obse a o y
on An isemi ism a Fondazione CDEC) ocused on
humou as a powe ul ehicle o no jus in-g oup
solida i y bu also o idiculing and beli ling he ou -
g oup. Jokes, i ony o schaden eude discou se shi
he line o wha is accep able, obscu e he speake ’s
in en ion, and o e a way o back ack i con on ed.
Bo h speake s emphasised ha i is misleading o
equa e ha e speech wi h ang y, emo ional ha e ul
language, as i aids g adual no malisa ion o he
o me , and he a i udes ha unde lie i , echoing he
h ea o implici ness deba ed in he i s wo kshop,
and calling a en ion o he u gen need o imp o ed
ools measu ing o ha e ideologies in hei whole
di e si y in o de o answe la ge ques ions o no -
malisa ion.
Wo kshop 5:
Mul imodal analysis o online
ha e speech
A he cen e o he i h e en , mode a ed by
D Lau a Ascone (Decoding An isemi ism), we e
alks ocusing on he g owing ins umen alisa ion o
memes, GIFs o ideos in enabling ha e ideologies
o in il a e digi al cul u e. D Janina Wild eue
and D Dimi is Se a is (Uni e si y o G oningen)
discussed bo h ‘so ’ and ‘ha d’ examples o ha e
speech in online en i onmen s, explo ed wi h wo-
s ep mul imodal analysis and a gumen a ion analy-
sis o e eal he in e play o se e al exp essi e o ms
in he meaning making, and he challenge o delim-
i ing con ex uali y. Ma cus Scheibe (Decoding
An isemi ism) in oduced he communica ion o ma
meme and showed how i is u ilised o an isemi ic
concep s. In pa icula , he highligh ed he empla e
cha ac e and demons a ed how memes can be
analysed om a discou se-semio ic pe spec i e.
D Ina i Sakki (Uni e si y o Helsinki) explo ed
mul imodal a e ac s in he o m o a TikTok ideo
c ea ed in he con ex o poli ical ha e communica-
ion. She placed emphasis on he complex in e play
o modali ies ( isual, e bal, au al) and he need o
me hodological wo k in he ield o music, as well as
he con ibu ion ha music can make o he cons uc-
ion o ha e. All speake s e e enced he impo ance
o including all modali ies in quali a i e analyses as
he only way o unde s and he cons uc ion o ha e
ideologies online, inso a as online communica ion
is always mul imodal. They s essed he need o
u he sys ema ic and ine-g ained analyses, la ge
co po a app oaches and compu a ional elemen s in
he ield.
4. The Decoding An isemi ism
62 – One o he scheduled speake s was
unable o join, hence only wo p esen a ions
a e summa ised he e.
51
Decoding An isemi ism
Wo kshop 6:
Au oma ed de ec ion o online
ha e speech
In he las wo e en s o he se ies, we wan ed o add ess he
applica ion o esea ch indings o p o essional p ac ice ou side
o academia. In he six h wo kshop o he se ies, mode a ed by
Jan Fillies (Ins i u e o Applied In o ma ics, Uni e si y o Leip-
zig) expe s conside ed means o au oma ically de ec ing
ha e speech and he cu en possibili ies and limi a ions o he
i e a i e exchange be ween human code s and AI models, wi h
he iew o signi ican ly ad ance social media s udies, bu also
o imp o e con en mode a ion. P o esso Helena Mihalje ić
(HTW Be lin & Decoding An isemi ism) alked abou he eam’s
wo k on de ec ion o online conspi acy heo ies and an isem-
i ism, based pa ly on da a sha ed by Decoding An isemi ism,
using a a ie y o models, including BERT, Llama and GPT-4.
She hypo hesised ha challenges o de ec ing online an i-
semi ism may be due o he mo e elusi e na u e o an isemi ic
exp essions and p ac ical aspec s such as he leng h o online
commen s, wo ld knowledge gap, o con ex uali y. D Sa as
Zanne ou (Uni e si y o Del & Max Planck Ins i u e o In o -
ma ics) desc ibed his wo k on de ec ing online an isemi ism, in
ex and images, on pla o ms such as 4Chan. He concluded
ha his model can play a ole no jus in iden i ying an isemi ic
con en bu also unde s anding i s e olu ion o e ime, bu ha
human in ol emen should s ill be pa o he mode a ion p o-
cess. The wo k Julia Mendelsohn (Uni e si y o Michigan)
p esen ed in he alk ocuses on coded language o en p esen
in oxic speech wi hin he public sphe e, and on in oducing
his p oblem o he ield o ha e speech de ec ion and compu-
a ional linguis ics. Toge he wi h he eam, she de eloped a
ypology o dog whis les, p epa ed a li ing glossa y wi h ich
con ex ual in o ma ion, analysed poli ical speeches, and e al-
ua ed ecommenda ions by la ge language models, wo king
wi h GPT3 p omp s and Google Pe spec i e API. The speake s
poin ed ou he ele ance o he shi away om adi ional
machine lea ning models owa ds new echnologies, impo -
ance o de ec ing exis ing oxic language as well as p edic ing
sub le and inno a i e exp essions o ha e ideologies, and he
u gen need o ex end he wo k om English o o he languages.
Wo kshop 7:
Educa ional s a egies o online
ha e speech awa eness, p e en ion
and eac ion
In he se en h wo kshop, mode a ed by Ka olina Placzyn a
(Decoding An isemi ism), we examined he lessons om
educa ional in e en ions and p ac ices. P o esso Özen
Odağ (Tou o Uni e si y Be lin) alked abou empowe men
o eme ging adul s and c ea ion o social change h ough
aining hem o ecognise and eac o online an isemi ism
wi hin a pa icipa o y amewo k, in o med by young peo-
ple’s own pe spec i es. D awing on a se ies o s udies as
pa o he RESPOND! p ojec , she iden i ied he challenge
o implici , low-in ensi y an isemi ism ha is di icul o young
people o ecognize and ha igge s disengagemen . She
emphasized young people’s iden i ies and loyal ies as
c ucial elemen s o media compe ence ainings o comba
an isemi ism. Miško S anišić (Te a o ming) discussed he
educa ional esou ces his o ganisa ion p o ides o coun e
s a e-sponso ed Holocaus dis o ion in Se bia (i. e. he
hijacking o his issue o p opaganda pu poses). He also
elabo a ed on he p ojec mapping an isemi ism iden i i-
ca ion capaci ies in Se bia, C oa ia and Romania wi hin
p o essional and social communi ies ha shape he public
discou se. Mal e Holle (Bildung in Wide sp uch) ga e a
p e iew o his o ganisa ion’s new websi e “An Allem Schuld”
(launched soon a e he wo kshop), a i ual place o in o -
ma ion and lea ning whe e eenage s and young adul s
can independen ly deal wi h cu en o ms o an isemi ism.
The esou ce p o ides basic knowledge abou an isemi ism
and i s mani es a ions, insigh s in o Jewish expe iences and
Jewish di e si y, bu also ways o manage emo ions and
psychological unc ions o an isemi ism, p omo ing ole ance
o ambigui y and, again, s imula ing e lec ion. All h ee
speake s ouched on he opics o young people g appling
wi h he di icul y and esponsibili y o con on ing an isem-
i ism, s essing he need o esea ch ha acks he ypes o
an isemi ic speech hey a e exposed o online.
In ou p e ious epo , we de ailed ou expe imen s
in ol ing ine- uning a BERT-based la ge language
model (LLM) o classi y ex s as ei he an isemi ic o
non-an isemi ic. Ou ocus was on he English-lan-
guage ex s and he u ilisa ion o a p e- ained
English-language model. Howe e , he imbalanced
na u e o he da a, wi h nea ly 10 non-an isemi ic
s a emen s o e e y an isemi ic one, p esen ed a
challenge common when a emp ing o iden i y
o ensi e language o ha e speech in mains eam
online media o social media pla o ms.
Ou ini ial model achie ed an F1 sco e o 0.69 o
an isemi ic ex s (posi i e class/class 1) and 0.96
o non-an isemi ic ex s (nega i e class/class 0).63
Ou objec i e was o employ a ious s a egies
o u he enhance he pe o mance o he exis -
ing model. Mo eo e , we u ilised he da a ha had
been anno a ed in he mean ime. Addi ionally, we
aimed o assess he obus ness o he model by
aining and e alua ing i on di e en spli s o
he a ailable da a. Alongside his, we explo ed
he capabili ies o con e sa ional models such
as Cha GPT ha ha e shown solid pe o mance
ac oss a ious ex classi ica ion asks.
5. Au oma ed de ec ion o
an isemi ic ex s: is con ex
all we need? Milena Pus e and Helena Mihalje ić
5.1 A emp s o op imise exis ing BERT-based model
Wo king wi h only app oxima ely 2,400 English-lan-
guage ex s labelled as an isemi ic posed a chal-
lenge due o he limi ed da a a ailable o aining.
This is especially he case gi en he b e i y, equen
usage o implici , coded speech, and agmen a ion
o he ex s in ou da ase . A he same ime, an isem-
i ism mani es s in a ious ace s wi hin hese ex s, as
e lec ed by he ex ensi e anno a ion scheme o he
p ojec . Thus, o e ec i e aining, a subs an ially
la ge olume o da a would be bene icial.
One s a egy o enhance he model in ol ed aug-
men ing he da a in he posi i e class as his is
signi ican ly unde ep esen ed. The i s app oach
in ol ed he eplacemen o some
wo ds by o he s wi h a supposedly
simila meaning in he sense o a p e- ained lan-
guage model, and he inse ion o wo ds ha sup-
posedly do no signi ican ly al e he o e all mean-
ing o he sen ence. The second app oach in ol ed
ansla ion as a s a egy. Tex s labelled as an isemi ic
we e ansla ed om English in o Ge man and hen
back o English. Addi ionally, posi i e da a om
o he languages we e ansla ed in o English. A an-
dom sample o ansla ions was manually checked.
Howe e , none o hese s a egies esul ed in a sig-
ni ican imp o emen in he F1 sco e o class 1. A
no able challenge s ems om he ac ha s anda d
p e- ained models may no e ec i ely cap u e he
sensi i e con ex in co esponding messages. Fo
ins ance, he wo ds ‘Is ael’ and ‘Pales ine’ migh be
63 – The F1 sco e is a measu e ha
combines how well a model inds ele an
ins ances ( ecall) and how p ecise i is in
iden i ying hem. I anges om 0 o 1, wi h
highe alues indica ing be e pe o mance.
53
Decoding An isemi ism
conside ed close om he pe spec i e o a gene al language
model, bu hey a e no sui able eplacemen s o each o he
in he con ex o he Middle Eas con lic . Models ha ha e
been u he ained o ine- uned on a co pus e lec ing such
nuances a e likely mo e sui able, as well as o he mo e
sophis ica ed ex augmen a ion s a egies ha we a emp o
explo e in he u u e.
Nume ous expe imen s we e conduc ed o educe he
so-called ‘easy nega i es,’ e e ing o examples om he
nega i e (i. e. non-an isemi ic) class ha a e easily co ec ly
classi ied. While a model is being ained, all da a poin s
con ibu e in he same way o he compu a ion o he e o
which guides he aining p ocess. Thus, educing such exam-
ples has he po en ial o enhance he de ec ion o he posi i e
class. While his s a egy led o a highe ecall o he posi i e
class, and hus an inc eased iden i ica ion o an isemi ic ex s,
i came a he cos o p ecision, so no subs an ial o e all
imp o emen was achie ed.
Fu he mo e, language-speci ic models we e ained o
Ge man and F ench, along wi h a mul ilingual model.
Howe e , he esul s ell sho o hose ob ained by he Eng-
lish-language monolingual model. The wo se pe o mance
o he Ge man and F ench models can be mainly a ibu ed
o he subs an ially smalle numbe o posi i e examples in
hese languages compa ed o he English co pus. A quali a-
i e analysis o he model’s ou pu wi h ega d o discou se
coun ies would be necessa y o make mo e p ecise s a e-
men s in his ega d.
In he pas six mon hs, we ha e inco po a ed new
English-language da a om bo h exis ing and eme ging
discou ses in o ou aining o new models. These new dis-
cou ses included, o example, Elon Musk’s wee abou
Geo ge So os and his nega i e ema ks abou he ADL, as
well as he eac ions o he 7 Oc obe massac e in Is ael and
he subsequen escala ion in he A ab-Is aeli con lic . In o al,
676 addi ional posi i e da a poin s we e included.
Wi h he in usion o hese new da a poin s, we p oceeded o
ain a ious models. As ou lined ea lie , ou objec i e was
no only o enhance model pe o mance bu also o ensu e
s able pe o mance when implemen ing di e se da a spli s
o aining, alida ion, and es ing pu poses. Mo e p ecisely,
we c ea ed i e spli s o he en i e da ase , always keeping
10% o he da a o es ing and assu ing ha he pe cen ages
o bo h classes a e compa able ac oss he di e en spli s.
As shown in Table 1, he o e all esul s a e less accu a e
han in he p e ious model. The ecall o he posi i e class,
which measu es he ac ion o an isemi ic ins ances
ha we e e ec i ely e ie ed by he model, was
especially low. Mo eo e , he esul s o he posi i e class
a y o a ela i ely high deg ee be ween he di e en spli s,
indica ing a high dependence o he model on da a
used o aining and e alua ion. The e a e a ious plau-
sible explana ions o his. The new da a con ains addi ional
discou se e en s in which an isemi ism migh ha e been
exp essed di e en ly (see da ase s in he epo ’s chap e s 2
and 3). Mo eo e , a quali a i e inspec ion o eco ds in he
posi i e class sugges s ha anno a o s end o s uggle wi h
ecognising he le el o ou side con ex hey d aw upon o
anno a e an an isemi ic ex as such. The pa icula na u e o
he anno a ion p ocess – whe e messages a e conside ed
wi hin he speci ic discou se e en and su ounding mes-
sages – makes i e y di icul o simula e he si ua ion whe e
his ype o con ex ual knowledge – knowledge ha is no
p o ided o he ained model – is una ailable.
OpenAI’s ecen s ides in he ield o gene a i e
p e- ained ans o me s (GPT), pa icula ly wi h he
in oduc ion o GPT-3 and he mul i-modal GPT-4,
ha e a ac ed conside able a en ion owing o hei
no ewo hy p o iciency in gene a ing esponses
ha closely esemble human-like exp essions. These
models, accessible h ough pla o ms such as
Cha GPT, also p esen no el oppo uni ies o ex
and image classi ica ion. Meanwhile, openly acces-
sible models such as Llama 2, de eloped by Me a,
ha e been p o ided, achie ing simila pe o mance
as GPT-3 models in a ious asks.
These models can be applied in bo h ze o-sho
and ew-sho se ings. In he o me , he model is
p esen ed solely wi h a ask desc ip ion, whe eas he
la e pe mi s he model o glean knowledge om a
limi ed se o so-called in-con ex examples, whose
ole is hus compa able o ha o aining da a o
supe ised models. The o mula ion o ex ual ins uc-
ions, e e ed o as p omp ing, signi ican ly in lu-
ences he model’s beha iou . I se es as a c i ical
ask, as i shapes he model’s ocus and de e mines
i s esponses, unde sco ing he impo ance o p eci-
sion and nuance in he design o p omp s.
Ini ial in es iga ions highligh he po en ial o hese
models in ex classi ica ion applica ions, e. g. in
de ec ing ha e ul o insul ing language.64 How-
e e , ecen esea ch also sheds ligh on he a ied
e ec i eness o GPT-3 in iden i ying “coded exp es-
sions ha simul aneously con ey one meaning o a
b oad audience and a second one, o en ha e ul o
p o oca i e, o a na ow in-g oup”65 – so-called
dog whis les – including an isemi ic ones. Fu he
in es iga ion in o he capabili ies and limi a ions o
gene a i e language models in ackling in ica e
classi ica ion asks a e needed, pa icula ly in he
con ex o an isemi ic ex s. The expe imen s we
p esen in he ollowing sec ion ep esen an ini ial
s ide in his di ec ion.
P ecision Recall F1-sco e Numbe
o eco ds Accu acy
class 0 (AS) 0.96 ± 0.00
(0.97 ± 0.00)
0.98 ± 0.01 0.97 ± 0.00
(0.97 ± 0.01)
3.533
0.94 ± 0.01
(0.95 ± 0.01)
class 1 (non-AS) 0.71 ± 0.05
(0.73 ± 0.04)
0.52 ± 0.04
(0.63 ± 0.03)
0.6 ± 0.02
(0.68 ± 0.02)
309
Table 1: Mean ± s anda d de ia ion o he
bes -pe o ming models pe spli on es da a
(and alida ion da a in b acke s i di e en ).
5.2 Expe imen s wi h p omp -based models
5. Au oma ed de ec ion o an isemi ic ex s: is con ex all we need?
64 – Chiu, Ke-Li/Collins, Annie/Alexande , Rohan, 2021. De ec ing Ha e Speech
wi h GPT-3. P ep in . a Xi :2103.12407; Wang, Yau-Shian/Chang, Yingshan, 2022.
Toxici y De ec ion wi h Gene a i e P omp -based In e ence. P ep in .
a Xi :2205.12390; Huang, Fan/Kwak, Haewoon/An, Jisun, 2023. Is Cha GPT
be e han Human Anno a o s? Po en ial and Limi a ions o Cha GPT in Explaining
Implici Ha e Speech. In Companion P oceedings o he ACM Web Con e ence
2023. ACM, Aus in TX USA, 294–297, doi:10.1145/3543873.3587368.
65 – Mendelsohn, Julia/Le B as, Ronan/Choi, Yejin/Sap, Maa en, 2023.
F om Dogwhis les o Bullho ns: Un eiling Coded Rhe o ic wi h Language Models.
In: P oceedings o he 61s Annual Mee ing o he Associa ion o Compu a ional
Linguis ics (Volume 1: Long Pape s). doi:10.18653/ 1/2023.acl-long.845.
55
Decoding An isemi ism
In ou expe imen s, we kep o he ze o-sho
se ing, p ima ily o wo easons. P io esea ch
indica es ha he pe o mance o models in ew-sho
scena ios may exhibi a decline compa ed o ze o-
sho se ings.66 The incu ed cos is con ingen on he
p omp leng h, which ends o be signi ican ly longe
when inco po a ing in-con ex examples, pa icula ly
when coupled wi h a cus om de ini ion. This is espe-
cially p onounced o GPT-4, which is al eady wice
as expensi e as GPT-3.5.
We conduc ed ze o-sho expe imen s employing
GPT-3.5, GPT-4 and Llama 2,67 whe ein we com-
pa ed pe o mance unde wo condi ions: one
whe e he p omp solely ou lines he ask, and
ano he whe e a cus om de ini ion o an isemi ic
speech is p o ided. Speci ically o GPT-3.5, we
u he in es iga ed he impac o cons ain s on
model ou pu . In he ‘Yes/No’ con igu a ion, he
model was asked wi h p o iding a bina y deci-
sion abou he p esence o an isemi ic con en in
he message. In he ‘Yes/No wi h Reason’ se ing,
an addi ional sho easoning was eques ed.
The hi d con igu a ion, ‘P obabili y Sco ing,’ man-
da ed he model o p o ide a sco e be ween 0
and 1 ha indica es he likelihood o a message
being an isemi ic.
The ollowing shows wha he p omp looked like o
he ‘Yes/No’ expe imen wi h he p o ision o an
expe de ini ion:
Conside he ollowing s a emen :
‘{STATEMENT}’
You ha e o decide whe he he
s a emen communica es an isemi ism
conside ing he ollowing de ini ion:
‘{DEFINITION}’
Gi e you answe using one o he
wo op ions:
A) Yes
B) No
Do no p o ide any o he ou pu o
explana ion in you ou pu .
E alua ions we e un on alida ion da a o he
model p esen ed in he p e ious epo , comp ising
249 posi i e and 2,061 nega i e samples. As shown
in Table 2, he con e sa ional models pe o m e en
wo se han ou bes BERT-based app oach (see
bo om ow in Table 2). Wi h an F1 sco e on class
1 o 0.55, GPT-4 is clea ly he bes -pe o ming
model, albei achie ing a signi ican ly lowe
sco e han p e ious a emp s using BERT-like
models. The F1 sco e o class 0 also emains below
he p e ious one. Bo h GPT-4 and Llama 2 pe o m
be e when equipped wi h an expe de ini ion o
an isemi ic speech. Howe e , GPT-3.5 shows di e -
en beha iou depending on he ou pu cons ain ,
wi h pe o mance depending on he p o ided de i-
ni ion only in he bina y ‘Yes/No’ se ing. This indi-
ca es ha GPT-3.5 highly depends on he chosen
con igu a ion.
5.3 Expe imen al esul s
66 – Chae, Youngjin (YJ)/Da idson, Thomas, 2023.
La ge Language Models o Tex Classi ica ion:
F om Ze o-Sho Lea ning o Fine-Tuning. P ep in ,
h ps://os .io/p ep in s/soca xi /s hwk; Pus e , Milena/
S e en, Elisabe h/Mihalje ić, Helena, 2023.
Conspi acy Na a i e De ec ion Beyond Keywo d Bias in
Ge man-language Teleg am. P ep in .
67 – The exac model e sions a e gp -3.5- u bo-0613,
gp -4-061 and llama 2 70b.
Model Ou pu cons ain De ini ion F1-class 0 F1-class 1 Accu acy
GPT-3.5
Yes/No
-0.93 0.40 0.88
X 0.95 0.19 0.90
Yes/No + Reasoning
-0.85 0.42 0.76
X 0.86 0.43 0.77
P obabili y Sco ing
( h eshold = 0.5)
-0.94 0.42 0.89
X 0.94 0.42 0.88
GPT-4 Yes/No
-0.95 0.46 0.91
X 0.95 0.55 0.92
Llama 2 Yes/No
-0.78 0.23 0.65
X 0.77 0.30 0.65
ine- uned
be -base-
uncased
0.96 0.69 0.94
Table 2: Pe o mance o he gene a i e
models on alida ion da a om UK discou ses,
comp ising 249 posi i ely and 2,061 nega i ely
labelled samples.
5. Au oma ed de ec ion o an isemi ic ex s: is con ex all we need?