THE PROJECT KISU: AI-BASED SEARCHING AND FINDING – EASY,
INCLUSIVE, AND SELF-DETERMINED
K. Al meye , M. Hladky, S. Malone, M. Pla z, K. Reese, L. Schick, V. Wol , Saa land Uni e si y,
Saa b ücken, Ge many
T. Go smann, M. Wiesne , agFINN e.V., Be lin, Ge many
C. Plo e, Open Sea ch Founda ion, S a nbe g, Ge many
Abs ac
Many use s, especially child en, olde people, o people
wi h special needs, ind i di icul o o mula e p ecise
sea ch que ies ha lead o ele an esul s. This can lead o
us a ion, limi ed use o he in e ne , and ul ima ely, a
eeling o digi al exclusion. A lack o unde s anding o he
sea ch logic and an inapp op ia e in e p e a ion o he
sea ch esul s ha bo he isk o use s consuming un eliable
in o ma ion o e en being exposed o disin o ma ion.
In he KIsu p ojec , we use an LLM ha analyses and
in e p e s he sea ch e ms en e ed. This AI model is ained
o ecognize and speci y he use ’s ac ual sea ch in en ion
om incomple e o imp ecise que ies. We a e also de el-
oping wo kshops and in o ma ion ma e ials o p omo e he
in elligen use o sea ch engines and AI.
KIsu is a coope a i e p ojec be ween Saa land Uni e -
si y and agFINN e.V. (as well as he Ge man Resea ch
Cen e o A i icial In elligence and Open Sea ch Founda-
ion). In his pape , he cu en s a us o he p ojec is p e-
sen ed.
INTRODUCTION
Google domina es he sea ch engine ma ke , ollowed by
Bing, which also p o ides he sea ch esul s o mos so-
called al e na i e sea ch engines and is gaining popula i y
hanks o he AI cha bo Cha GPT. These dominan p o id-
e s do no disclose hei algo i hms and sea ch indices,
leading o es ic ions in anspa ency, accessibili y, da a
p o ec ion, and secu i y, among o he hings. The conse-
quences can be bias, in o ma ion asymme ies, and da a-
d i en disc imina ion (Bobic, Pla z & Gü l, 2021 [1];
Galindo & Ga cia-Ma co, 2017 [2]). Al hough sea ch en-
gines play a signi ican ole in ou e e yday li es, mos
people use hem wi hou knowing o ques ioning how hey
wo k. E ec s such as a lack o anspa ency and imma u-
i y a e inc easing massi ely due o gene a i e AI language
models. In addi ion, comme cial p o ide s ha e li le in e -
es in esea ching and o e ing low-ba ie access.
To enable all people – ega dless o hei age o gende ,
wi h o wi hou disabili ies – o use in e ne sea ches and
sea ch engines in an in o med, sel -con iden , and sel -de-
e mined way and hus acili a e hei pa icipa ion in he
digi ally pe aded wo ld, we a e de eloping lea ning and
in o ma ion ma e ials and o ganizing wo kshops in he
KIsu p ojec . We wan o suppo ci izens in o e coming
he black box e ec o sea ches and analyzing and c i iciz-
ing sea ch esul s. Addi ionally, we aimed o speci ically
iden i y he challenges encoun e ed by pa icula a ge
g oups du ing in e ne sea ches and o de elop me hods o
suppo ing hese g oups bo h pedagogically and echnolog-
ically.
The lea ning ma e ials on sea ch engine li e acy (e.g.,
Pla z e al., 2023 [3]) and he Open Sea ch Founda ion
(h ps://opensea ch ounda ion.o g/en/child en-and-in e -
ne -sea ch/) se ed as a s a ing poin . They a e co-c ea-
i ely adap ed wi h a ious g oups om ci il socie y and
u he de eloped conce ning he use o AI on ends and
cha bo s. We ocus on child en, young people and olde
people.
All ma e ials de eloped in he p ojec a e published as
Open Educa ional Resou ces (OER). Design p inciples a e
de i ed and implemen ed o designing AI on ends o
sea ch engines ha a e o ien a ed owa ds he common
good based on he p e iously de eloped lea ning ma e ials
and explo a i e use wo kshops.
This pape desc ibes he objec i es and s eps o he KIsu
p ojec . Fu he mo e, he esea ch design and i s esul s
o he p ojec a e p esen ed.
OBJECTIVES
Wi h KIsu, we wan o s eng hen he digi al so e eign y
and pa icipa ion o people wi h special needs, child en,
and olde people. We aim o enable as many use g oups as
possible o use sea ch engines e icien ly and inclusi ely.
Wi h he help o AI-suppo ed language models, acces-
sible sea ch in e aces, and aining, we c ea e he condi-
ions o using he in e ne as a sou ce o in o ma ion
sa ely, con iden ly, and compe en ly. In addi ion, ou ac-
companying s udies p o ide new insigh s in o how gene -
a i e AI models can be used o educe exis ing ba ie s o
in e ne use. In summa y, KIsu pu sues he ollowing
goals:
•P omo ing he compe en and esponsible use o In e -
ne sea ches h ough de eloping and es ing lea ning
and in o ma ion ma e ials.
•The iden i ica ion o ac o s and speci ic needs in
dealing wi h sea ch engines and he de i a ion o im-
plica ions o he u he de elopmen o lea ning ma-
e ials and o designing low-ba ie AI sea ch engine
on ends.
•Fine- uning a needs-based AI language model and de-
eloping a model-agnos ic da a pipeline.
•P o ision o AI on ends o op imized sea ch ( ech-
nology ma u i y le el 5).
•P o ision o he da a collec ed in he p ojec ( he in-
e en ion and e ec i eness s udy).
h ps://doi.o g/10.5281/zenodo.17229191
PROJECT STEPS
The p ojec du a ion is 19 mon hs (01.06.2024-
31.12.2025). The p ojec consis s o ou s eps:
•S ep 1 – Ma e ial and on -end de elopmen : lea n-
ing, aining, and in o ma ion ma e ials a e designed
co-c ea i ely wi h a ious g oups om ci il socie y o
p omo e he compe en use o in e ne sea ches. The
ollowing g oup composi ions and sizes we e en is-
aged: 15 child en, including child en wi h a ious spe-
cial educa ional needs (e.g., emo ional-social de elop-
men , men al de elopmen , hea ing, physical and mo-
o de elopmen , lea ning, ision, and language), 15
young people, including young people wi h a ious
special educa ional needs, 15 olde people, including
people wi h a ious special educa ional needs. Ca e is
aken o ensu e a balanced gende composi ion. Fo he
de elopmen o he i s in ui i e on ends o he
sea ch engines, a modula design is being sough ha
enables simple ule-based in e ac ions and can be ex-
panded la e . Su ey ools a e selec ed and de eloped
o iden i y ac o s and speci ic needs in dealing wi h
sea ch engines. A kick-o wo kshop wi h all p ojec
pa icipan s ook place. Miles one: i s e sion o ma-
e ials and on end.
•S ep 2 – Ma e ial and on -end op imiza ion: ma e i-
als and on ends a e es ed wi h a con olled in e en-
ion s udy (n=45) and op imized based on he esul s.
Miles one: Op imized e sion o ma e ials and he
on end.
•S ep 3 – Dissemina ion o he ma e ials de eloped in
he p ojec and on -end de elopmen : A aining con-
cep and OER will be p epa ed in coope a ion wi h
a ious g oups om ci il socie y and published design
p inciples o he design o low-ba ie AI sea ch en-
gine on -ends a e de i ed and implemen ed in a inal
applica ion ( aining o a needs-based AI language
model). The design p inciples o de eloping AI
on ends o emo e ba ie s o use and acili a e ac-
cess o ci ic da a will be made a ailable. In addi ion, a
con e ence will be held wi h all p ojec pa icipan s. A
mul iplie ne wo k will be es ablished. Miles one:
aining concep , OER, op imized e sion o he
on end, design p inciples, on ends.
•S ep 4 – P o ision o da a and publica ion o esea ch
esul s: The da a collec ed will be p ocessed and
passed on o a esea ch da a eposi o y. The esea ch
wo k and s udy esul s will be published o open ac-
cess by he ele an specialis audience. Miles one:
Open da a, jou nal publica ions.
RESEARCH DESIGN AND RESULTS
An Ac ion Design Science Resea ch (ADSR) app oach
is pu sued (Mulla key & He ne , 2019 [4]). Design Sci-
ence Resea ch (DSR) is a pa adigm oo ed in he philoso-
phy o p agma ism. DSR in ol es p oblem-sol ing e-
sea ch o answe esea ch ques ions ela ed o human p ob-
lems and p oduces aluable a e ac s. ADSR cen e s on co-
c ea i e collabo a ion be ween scien is s and use s. The
goals o he i s phase (diagnos ic phase) a e o analyze
he p oblem space and he solu ion space (he e: he iden i-
ica ion o ac o s and speci ic needs in dealing wi h sea ch
engines and he de i a ion o implica ions o he de elop-
men o lea ning ma e ials and o he design o low-ba ie
AI sea ch engine on -ends) o esea ch and p ac ice and
hei ele ance in mu ual ag eemen be ween he e-
sea che -use eam. A mixed me hods app oach (e.g.,
Kucka z, 2014 [5]) is pu sued in which quan i a i e da a
collec ion me hods such as ques ionnai es a e combined
wi h quali a i e da a collec ion me hods such as in e iews
and obse a ions in co-c ea i e wo kshops. The sample
comp ises child en, young people, and olde people (see
subsec ions below). The quan i a i e da a is analyzed de-
sc ip i ely and in e en ially, he quali a i e da a is ana-
lyzed using Design hinking me hods, such as de eloping
Pe sonas (Uebe nickel e al., 2015 [6]) combined wi h
quali a i e con en analysis (May ing, 2015 [7]), and he
esul s a e co ela ed.
Then ollows he design phase, in which he a e ac is
iden i ied and concep ualized (he e: lea ning ma e ials and
low-ba ie AI sea ch engine on ends). Design p inciples
a e ( u he ) de eloped h ough se e al i e a i e cycles
wi hin he design phase. Collabo a i e ac i i ies wi h co-
c ea i e ac i i ies a e essen ial he e, as he esea che -use
eam aims o c ea e a e ac s ha inco po a e inno a i e
ideas o sol ing he gi en p oblems. In he implemen a-
ion phase, concep s a e de eloped o use he a e ac . An
ac ual applica ion o e s he oppo uni y o e alua e he e -
iciency and e ec i eness o he p oposed design in p ac-
ice.
We a e cu en ly in he design phase (S ep 2 o he p o-
jec ). We ha e al eady ou lined de ailed lea ning ma e ials
and selec ed speci ic ools o es he e ec i eness o hese.
We will soon be es ing hese on an ini ial sample. A he
same ime, we a e wo king on implemen ing he AI-sup-
po ed sea ch on ends and op imizing hei unc ionali ies
based on use eedback. A cen al p oblem is ha he e has
been e y li le sound esea ch in o he he e ogeneous a -
ge g oups and hei beha io when sea ching he in e ne .
We mus , he e o e, i s c ea e a solid empi ical basis o
u he de elopmen s eps. The p ojec elemen s, co-c ea-
i e wo kshops, a con olled in e en ion s udy, and he AI
on end a e desc ibed below.
Co-c ea i e Wo kshops
In line wi h Ind & Coa es (2013 [8]), end-use s a e in-
ol ed, which leads o mo e ele an and usable p oduc s
and se ices while educing isk. Pa icipa o y design is
used o de elop i e a i e p o o ypes o es use eac ions.
The wo kshops we e concep ualized using Design hinking
(e.g., Uebe nickel e al., 2015 [6]). In he wo kshop he pa -
icipan s design hei own digi al assis an , ha can help
hem o ind wha hey sea ch on he in e ne . In o de no
o emp he wo kshop pa icipan s o ep oduce exis ing
solu ions, bu o become c ea i e hemsel es, he wo d
‘digi al assis an ’ was used ins ead o ‘sea ch engine’.
The ollowing key- ques ions guided he wo kshops:
•Wha a e he ea u es o you digi al assis an ?
h ps://doi.o g/10.5281/zenodo.17229191
oWha does you digi al assis an look
like?
oWha should you digi al assis an be
able o do?
•Inpu me hod:
oHow do you wan o ell you digi al
assis an wha you a e sea ching o ?
oWha elemen s do you need on he
sc een o s a you sea ch?
•Ou pu me hod:
oHow should you digi al assis an p e-
sen he sea ch esul s?
oHow should i ell you wha i has
ound?
oHow should he sea ch esul s be p e-
sen ed?
•Wha happens i you ha e (no ) ound wha you we e
looking o ?
oHow do you ell you digi al assis an ?
oHow does he digi al assis an eac o
his?
oWha should he digi al assis an do?
The pa icipan s designed he in e ace o hei assis an
in small g oups o 3-4 indi iduals, using small whi eboa ds
and whi eboa d ma ke s as well as icons ha could be s uck
o he boa d (see Figu e 1).
Figu e 1: Design o he digi al assis an and inpu
me hod by a child
Fou co-c ea i e wo kshops we e o ganized wi h he ol-
lowing use g oups:
•Child en a ending he 2nd g ade in p ima y school (ca.
7 yea s old)
o8 gi ls, 5 o whom speak Ge man as a
second language
o7 boys, 3 wi h Ge man as a second lan-
guage and 1 wi h Ge man as a o eign
language
•Ma hema ical in e es ed child en a ending he 3 d o
4 h g ade in p ima y school (be ween 9 and10 yea s
old)
o5 gi ls
o9 boys
•Young people a ending he 7 h g ade in g amma
school (ca. 13 yea s old)
o9 emale
o10 male
•Olde people (be ween 64 and 87 yea s old)
o10 emale
o4 male
Du ing he wo kshops, quali a i e in e iews we e pe -
o med wi h he pa icipan s. The wo kshops and in e -
iews we e ideog aphed, and key scenes we e an-
sc ibed. Fo analysis, pe sonas a e de i ed. Pe sonas a e
desc ip i e models o use s. They a e a che ypes wi h a se
o p ope ies o di e en bu – conce ning de ined aspec s
–compa able pe sons (Uebe nickel e al., 2015 [6]).
Ini ial esul s indica e ha child en p e e ha ing social
and iendly in e ac ions wi h hei digi al assis an , while
us wo hiness is pa icula ly impo an o olde use s.
Speech inpu and ou pu seem o be sui able ac oss all use
g oups. Adolescen s exp essed a clea p e e ence o ai-
lo ed assis ance, meaning he digi al assis an should p e-
cisely ma ch he complexi y o he p o ided in o ma ion o
hei speci ic needs and sensi i ely adjus i s con e sa-
ional one – such as adop ing a humo ous s yle when
sea ching o en e ainmen con en .
Con olled In e en ion S udy
In he co-c ea i e wo kshops, we obse ed ha he
g oup comp ising sligh ly olde child en and eenage s
demons a ed he highes le el o p io knowledge ega d-
ing in e ne sea ch s a egies, use o in o ma ion echnol-
ogy, and AI. Consequen ly, his g oup also p o ided he
mos subs an ial inpu o he co-c ea ion o in o ma ional
ma e ials and he design o a sui able on end o in e ne
sea ches.
Based on hese insigh s, we will conduc an ini ial in e -
en ion s udy speci ically a ge ing his age g oup, using
ailo ed lea ning ma e ials and a cus omized on end de-
signed o in e ne sea ch ac i i ies. The s udy will add ess
h ee esea ch ques ions:
1. Does he use o speci ically de eloped lea ning
ma e ials and he cus omized on end lead o
measu able imp o emen s in aspec s o sea ch en-
gine li e acy and p o iciency in AI-suppo ed in-
e ne sea ches?
2. How do child en e alua e he usabili y and use-
ulness o he de eloped on end?
3. How do he designed in o ma ional ma e ials im-
pac child en's in e ac ion wi h he on end du -
ing in e ne sea ches?
The s udy will in ol e a minimum o 45 pa icipan s
om g ades 5 o 8 a ending a Mon esso i school. The pa -
icipan s will wo k indi idually bu will be o ganized in o
small g oups o p ac ical pu poses. The s udy begins wi h
all child en comple ing p e- es s assessing hei knowledge
and a i udes owa d in e ne sea ching and AI. Subse-
quen ly, child en will be andomly assigned o one o wo
in e en ion g oups:
•In e en ion G oup 1 (IG 1) will i s engage wi h
he lea ning ma e ials co e ing gene al in o -
ma ion as well as inpu and ou pu p ocesses o
AI-suppo ed in e ne sea ches. They will hen
comple e a s uc u ed in e ne sea ch ask using
he newly de eloped on end. Following his,
pa icipan s will e alua e he on end’s usabili y
h ps://doi.o g/10.5281/zenodo.17229191
and use ulness. Thei in e ac ions wi h he
on end will be eco ded. Finally, he ini ial es s
on sea ch engines and AI will be epea ed o
measu e lea ning ou comes and a i ude changes.
•In e en ion G oup 2 (IG 2) di e s only in he se-
quence o asks: child en in his g oup will i s
wo k wi h and e alua e he on end, ollowed by
he s udy o in o ma ional ma e ials.
O e all, we expec bo h g oups o bene i om ou in e -
en ion by gaining essen ial knowledge ela ed o c i ical
aspec s o AI-suppo ed in e ne sea ches and, i p esen ,
co ec ing unc i ical a i udes owa ds AI. Compa ing he
in e ac ions be ween IG 1 and IG 2 will e eal he ex en
o which pa icipan s bene i om he lea ning ma e ials
du ing ac ual in e ne sea ches. I is hypo hesized ha pa -
icipan s in IG 1 will apply mo e o he p inciples co e ed
in he ma e ials compa ed o IG 2, leading hem o pe cei e
he de eloped on end as mo e use ul and usable.
AI on end
A cen al aspec o ou p ojec is he de elopmen o in-
no a i e, low-ba ie use on ends. The use in e ace
should be able o ecognize and unde s and he use ’s ques-
ions, especially hose o people wi h special needs. By in-
eg a ing an LLM, he use que y is analyzed, op imized,
and con e ed in o a sui able sea ch que y. This imp o ed
que y is hen o wa ded o a sea ch engine, e.g., F ag-
FINN.de. The e a e also plans o include o he (al e na i e)
sea ch engines (such as Ecosia). Despi e he inhe en opac-
i y o la ge language models (LLMs), we would, like o use
hem speci ically o dialogue wi h he use o be e unde -
s and hei ac ual in o ma ional needs and gene a e a mo e
p ecise and ele an sea ch eques as pa o ou p ojec .
This p ocess should enhance he unde s anding o use s'
ac ual in o ma ional needs o sea ch esul s and accessibil-
i y and use - iendliness o all use g oups. The abili y o
ine- une LLMs o inclusi e language o o adap hem o
he unique sea ch que ies o child en, olde adul s, o peo-
ple wi h disabili ies is cen al o us. Open-sou ce models
like Llama 3 o Mis al enable lexible adap a ion h ough
me hods such as LoRA o QLoRA, while p op ie a y mod-
els like GPT-4-Tu bo allow ine- uning ia API. Open-
sou ce models also ha e he ad an age ha hey a e mo e
cos -e ec i e, he e is no di ec dependency on indi idual
companies, he dynamics o he models can be ully con-
olled, and da a p o ec ion mechanisms a e easie o im-
plemen and e iew. Th ough wo kshops and educa ional
p og ams, we p o ide in o ma ion abou he da a p o ec-
ion p ac ices o sea ch engines and he gene a i e AI sys-
ems we use, which p omo e a mo e conscious and secu e
handling o pe sonal da a.
The explainabili y and in e p e abili y o he esponses
gene a ed by LLMs a e cu en ly essen ial esea ch opics.
Al hough he p esen ly a ailable LLMs a e no ye ully
in e p e able, we in end o ac i ely ollow he la es ad-
ances in his a ea and, whe e possible, in eg a e hem in o
ou p ojec . This includes he e alua ion and po en ial
implemen a ion o me hods o inc ease he anspa ency
and aceabili y o AI-suppo ed p ocesses.
CONCLUSION
The p ojec KIsu aims o make digi al in o ma ion se -
ices accessible and unde s andable o e e yone and hus
s eng hen digi al pa icipa ion in he long e m. The p o-
jec un olds i s impac h ough
• he use o he aining modules in aining cou ses and
ain- he- aine cou ses,
•u ilizing and making a ailable he design p inciples
and s udy esul s o he de elopmen o AI on -ends
o educe ba ie s o use and acili a e access o ci ic
da a,
•building a mul iplie ne wo k h ough aining and
OER,
• aising awa eness h ough public ela ions wo k.
In he long e m, we hope ha AI will ecognize use s’
sea ch in en ions and p oac i ely suppo hem in speci y-
ing hei que ies, c i ically e alua ing in o ma ion, and
making in o med decisions. We also hope ha he aining
concep s and echnologies de eloped in ou we ecognize
hei po en ial o imp o e sea ch que ies signi ican ly p o-
jec will be in eg a ed and used by educa ional ins i u ions,
social ins i u ions, and ad ice cen e s o p omo e digi al
skills h oughou socie y in he long e m.
ACKNOWLEDGMENTS
We hank all he pa icipan s in he co-c ea i e wo k-
shops and su eys, ou s uden assis an s and he Ge man
Fede al Minis y o Family A ai s, Senio Ci izens,
Women, and You h o unding he p ojec , unding code
3924406K05.
REFERENCES
[1] Bobic, A., Pla z, M. & Gü l, C. (2021). Towa ds Open Sea ch
Applica ions o he b oade Communi y. In P oceedings o
he 3 d In e na ional Symposium on Open Sea ch Technol-
ogy, 11-13 Oc obe 2021, CERN, Gene a, Swi ze land.
[2] Galindo, F. & Ga cia-Ma co, J. (2017). F eedom and he in-
e ne : empowe ing ci izens and add essing he anspa ency
gap in sea ch engines. Eu opean jou nal o law and echnol-
ogy, 8(2),1–18.
[3] Pla z, M., Decke , A. & Plo e, C.(2023). Fö de ung eines
e s ändigen Umgangs mi Suchmaschinen im Elemen a -
be eich. M. R. Tex o & A. Bos elmann (H sg.), Das Ki a-
Handbuch.
[4] Mulla key, M. T., & He ne , A. R. (2019). An elabo a ed ac-
ion design esea ch p ocess model. Eu opean Jou nal o In-
o ma ion Sys ems, 28(1), 6–20
[5]Kucka z, U. (2014). Mixed me hods: me hodologie, Fo -
schungsdesigns und Analyse e ah en. Sp inge -Ve lag.
[6]May ing, P. (2015). Quali a i e Inhal sanalyse: G undlagen
und Techniken. Bel z Pädagogik
[7] Ind, N., & Coa es, N. (2013). The meanings o co‐c ea ion.
Eu opean business e iew, 25(1), 86-95.
[8]Uebe nickel, F., B enne , W., Pukall, B., Nae , T., &
Schindlholze , B. (2015). Design Thinking: Das Handbuch.
F ank u e Allgemeine Buch.
h ps://doi.o g/10.5281/zenodo.17229191
