Wo kshop
Recommended ci a ion: La Scala, J., & Gille , D. (2025). Explo ing Collabo a ion
Models o Suppo ing Idea ion wi h AI Agen s. In Kangaslampi, R., Langie, G.,
Jä inen, H.-M., & Nagy, B. (Eds.), SEFI 53 d Annual Con e ence. Eu opean
Socie y o Enginee ing Educa ion (SEFI), Tampe e, Finland. DOI:
10.5281/zenodo.17631593.
This Con e ence Pape is b ough o you o open access by he 53 d Annual Con e ence
o he Eu opean Socie y o Enginee ing Educa ion (SEFI) a Tampe e Uni e si y in
Tampe e, Finland. This wo k is licensed unde a C ea i e Commons
A ibu ion-NonComme cial-Sha e Alike 4.0 In e na ional License.
Explo ing Collabo a ion Models o Suppo ing Idea ion wi h AI
Agen s
J La Scala a,1, D Gille b
a École Poly echnique Fédé ale de Lausanne,
Lausanne, Swi ze land,
0000-0002-8057-7787
b École Poly echnique Fédé ale de Lausanne,
Lausanne, Swi ze land,
0000-0002-2570-929X
Con e ence Key A eas: Digi al ools and AI in enginee ing educa ion
Keywo ds: Gene a i e Language Models, Human-AI Collabo a ion, Idea ion,
Inno a ion, Design-based Lea ning
ABSTRACT
This wo kshop explo es Human-AI collabo a ion in idea ion wi hin design-based
lea ning, a pedagogical app oach widely used in enginee ing educa ion. Wi h he ise
o Gene a i e AI and gene a i e language models (GLMs), new possibili ies ha e
eme ged o suppo bo h di e gen and con e gen hinking in c ea i e p oblem-
sol ing. Howe e , while AI agen s ha e shown p omise in gene a ing ideas and
guiding he idea ion p ocess, hei in eg a ion in o educa ional con ex s emains
unde explo ed, pa icula ly om he pe spec i e o educa o s.
Pa icipan s engaged in a s uc u ed, hands-on explo a ion o AI-assis ed idea ion.
The wo kshop began wi h an in oduc ion ollowed by an in e ac i e demons a ion,
whe e pa icipan s expe ienced an AI-suppo ed idea ion ac i i y om a s uden ’s
pe spec i e. This was ollowed by a guided u o ial on implemen ing AI-gene a ed
eedback in a collabo a i e online en i onmen (G aasp.o g), allowing pa icipan s o
ailo he beha io o he AI agen o hei own educa ional scena io. The inal
segmen consis ed o a s uc u ed discussion, whe e pa icipan s e lec ed on he
bene i s, limi a ions, and pedagogical implica ions o AI eedback o idea ion, as well
1 Co esponding Au ho
J La Scala
je emy.la[email p o ec ed]
as sha ed hei pe spec i es and expec a ions ega ding human-AI collabo a ion o
idea ion ac i i ies.
By he end o he wo kshop, pa icipan s had i s hand expe ience designing and
in eg a ing AI eedback in idea ion ac i i ies, and con ibu ed o a collec i e epo
cap u ing educa o insigh s—in o ming ongoing esea ch and p ac ice a he
in e sec ion o AI, design-based lea ning, and enginee ing educa ion.
1 BACKGROUND AND RATIONALE
1.1 Idea ion in Design-based Lea ning
Design-based lea ning is an app oach used in enginee ing educa ion o each
s uden s he in eg a ion o hei disciplina y knowledge and skills owa ds he design
o an a i ac . Wi hin his app oach, design hinking is he co e me acogni i e and
manage ial p ocess ha s uden s p ac ice. A i s co e, he c ea i i y o he s uden s is
solici ed o combine hei knowledge o de ise po en ial solu ions o hei p oblem
s a emen h ough idea ion. Idea ion comp ises wo complemen a y cogni i e
p ocesses: di e gen hinking, which gene a es a b oad a ay o ideas, and
con e gen hinking, which e ines and selec s he mos sui able ones (F ich e al.,
2018, 2019).
1.2 Human-AI Collabo a ion o Idea ion
Wi h ad ancemen s in Gene a i e AI and mo e speci ically, gene a i e language
models (GLMs), new possibili ies ha e eme ged o suppo ing he idea ion p ocess.
GLMs can gene a e a wide a ie y o ideas (Haase & Hanel, 2023; Memme e al.,
2024; S e enson e al., 2022) and hus, con ibu e o di e gen hinking. In he o m
o cha bo s, hese models can aid indi iduals in he idea ion p ocess, p o iding
aluable sugges ions o guidance (Memme & Bi ne , 2024).
In his wo kshop, we in oduced and discussed new AI agen s and hei
co esponding in e ac ion models ha we designed o suppo ing s uden s du ing
collabo a i e idea ion. We classi y hese agen s in wo ca ego ies: (1) he a i icial
pee s and (2) he a i icial acili a o s. The a i icial pee s enac he ole o a
pa icipan o he idea ion ac i i y. They can p opose ideas and p o ide eedback on
hese ideas. S udies o a i icial pee s as con ibu o s o he pool o ideas ha e
epo ed a posi i e ecep ion by pa icipan s o he idea ion bu we e inconclusi e
ega ding he e ec s on he c ea i i y o he g oups (La Scala, Ba łomiejczyk, e al.,
2025; Schwabe e al., 2025).
On he o he hand, he acili a o s p o ide guidance du ing he ac i i y, wi hou
di ec ly con ibu ing o he idea ion. This ype o agen may be mo e ele an o
educa ional con ex s as hei ask do no o e lap wi h he ole o he s uden s and do
no p o ide an oppo uni y o he s uden s o o load hei hinking o he agen . This
wo kshop ocuses on he concep o a i icial acili a o and, in pa icula , AI
eedback.
To da e, ew in es iga ions ha e examined eache s’ pe cep ions in designing and
implemen ing sys ems ha acili a e human-AI collabo a i e idea ion.
2 WORKSHOP OBJECTIVES AND DESIGN
The pu pose o his wo kshop was o ini ia e he con e sa ion in he enginee ing
educa ion communi y on Human-AI collabo a ion and he co esponding in e ac ion
models o suppo c ea i i y in idea ion. Mo e speci ically, we collec ed he
pe cep ion o he eache s ega ding hese echnologies, discuss new collabo a ion
models wi h hei bene i s, isks and limi a ions. The pa icipan s had he oppo uni y
o expe imen wi h he con igu a ion and he use o a eedback agen h ough an
online collabo a i e en i onmen in eg a ed o G aasp.o g, a lea ning expe ience
pla o m (LXP) (Gille e al., 2022).
2.1 Ta ge Audience
This wo kshop has been designed o science and enginee ing educa o s in e es ed
in explo ing he bene i s o Gene a i e AI in educa ion, pa icula ly h ough i s
applica ion in p ojec -based and design-based lea ning app oaches. The wo kshop
was also in ended o be ele an o educa o s wo king in in e disciplina y con ex s,
whe e s uden s engage wi h design p ojec s ha in ol e collabo a ion wi h social
sciences, humani ies, o o he non- echnical disciplines.
2.2 S uc u e
Table 1. Time plan o he wo kshop.
Run ime
Ac i i y
15 min
In oduc ion and p esen a ion o Human-AI collabo a ion models o
idea ion wi h ocus on AI eedback.
5 min
In e ac i e demo - Pa icipan s ied a collabo a i e idea ion ac i i y wi h
AI eedback.
25 min
Agen implemen a ion u o ial - Pa icipan s we e in oduced o p omp
design o eedback gene a ion and con igu ed an agen o a scena io o
hei choice.
15 min
S uc u ed discussion – Pa icipan s sha ed hei expe ience wi h p omp
design and hei ision o in eg a ing AI eedback in collabo a i e
ac i i ies.
The wo kshop s a ed wi h he p esen a ion o possible agen oles in he con ex o
collabo a i e idea ion. Then, we opened a sho discussion ollowed by a
ques ionnai e on AI eedback and he isk and bene i s associa ed o i . Following
his in oduc ion, he pa icipan s we e p o ided wi h an example idea ion ac i i y in
which hey could y o p opose ideas and ecei e au oma ed eedback.
A e wa ds, hey we e guided in he p epa a ion o hei own idea ion ac i i y on a
opic o hei choice. In his s ep, he pa icipan used a web applica ion ha we
de eloped o in eg a e AI agen s in collabo a i e scena ios. The applica ion has
been designed o suppo he eache s in he con igu a ion o he agen s and he
ac i i y h ough an in ui i e in e ace, wi h he aim o conside ably educe he ime
and skills equi ed o design and implemen such ac i i ies. The main ask was o
design a p omp ha was au oma ically used by he sys em o gene a e eedback
when a new idea was submi ed. The pa icipan s we e gi en wo empla es hey had
o discuss and ailo . The app elies on a empla e engine1 o build he eques s ha
a e submi ed o he AI model. When gene a ing eedback, special ma kup is
eplaced by he p oblem s a emen , he idea (labeled esponse in he sys em), and
p e ious ideas wi h hei co esponding eedback. See P omp 1 o a de ailed
example and see (La Scala, Ba łomiejczyk, e al., 2025; La Scala, Sahli, e al.,
2025) o a de ailed desc ip ion o he whole applica ion.
A e his u o ial, we opened a g oup discussion on he p omp design, AI eedback,
and mo e gene ally he concep o AI eedback in collabo a i e idea ion. We
collec ed he answe s on an online discussion boa d whe e pa icipan s could submi
hei opinion in pa allel o he discussion. We inally opened o discussion o he
expec a ions o he pa icipan s ega ding human-AI collabo a ion in he con ex o
collabo a i e idea ion, and, mo e gene ally, collabo a i e lea ning.
3 OUTCOME
The wo kshop in ol ed eigh pa icipan s who we e o ganized in o h ee g oups o
he hands-on ac i i ies. I was acili a ed by he i s au ho . Da a collec ed, ield
no es and addi ional ma e ial a e a ailable in he supplemen a y ma e ial:
h ps://doi.o g/10.5281/zenodo.17200231.
Following he in oduc ion, we b ie ly discussed he bene i s, limi a ions, and isks
associa ed wi h human-AI collabo a ion, pa icula ly ega ding AI-gene a ed
eedback. In g oups, pa icipan s comple ed a su ey explo ing hei pe cep ions o
1 LiquidJS: h ps://liquidjs.com/ [a chi e]
Figu e 1: Example o an idea p oposed by one pa icipan wi h he AI-gene a ed eedback.
he main bene i s and isks o
inco po a ing AI-gene a ed
eedback in collabo a i e idea ion.
The p ima y bene i ha eme ged
was he po en ial o ob ain
eedback mo e quickly and a a
la ge scale. One g oup no ed ha
AI canno become a igued by
p o iding epe i i e eedback,
unlike eache s o eaching
assis an s. They also ecognized
ha using AI eedback c ea es
oppo uni ies o discuss AI and
human-AI collabo a ion in he
class oom while os e ing c i ical
hinking abou AI-gene a ed
esponses. Ano he g oup
emphasized he g ea e objec i i y
o AI, which is no in luenced by
s uden cha ac e is ics o
in e pe sonal ela ionships.
Rega ding d awbacks and isks,
wo g oups iden i ied he dange o
o e - eliance on AI eedback,
which p esen s a wo old conce n.
Fi s , he eedback may be
inaccu a e o comple ely inco ec ,
po en ially due o hallucina ions.
Second, s uden s may use AI
eedback o a oid meaning ul
in e ac ions wi h eache s.
Pa icipan s men ioned sycophan ic
AI (Sha ma e al., 2025) as a
po en ial d i e o his isk, which could impai eedback quali y by making i
insu icien ly c i ical and mo e appealing o s uden s han eache o pee eedback.
A e he in e ac i e demons a ion, we solici ed pa icipan s’ opinions on he AI-
gene a ed eedback. All pa icipan s ag eed on wo c i ical issues: he eedback was
excessi ely long and e bose, and i was o e ly posi i e, ein o cing he p e iously
iden i ied p oblem o sycophan ic AI.
Wi h hese conce ns in mind, we p oceeded o he u o ial on p omp design and AI
agen con igu a ion. Du ing he session, all g oups success ully con igu ed hei
ac i i ies and es ed hem wi hin he alloca ed ime ame. One g oup inqui ed abou
he possibili y o ha ing he AI gene a e and includes images in i s eedback, which
was no echnically easible a he ime.
By examining he p omp s de eloped by pa icipan s, we made wo key
obse a ions. Fi s , e y ew modi ica ions we e made o he empla es, which may
P oblem s a emen : {{ p oblem_s a emen }}
Pa icipan 's idea: {{ cu en _ esponse }}
{% i p e ious_ esponses.size > 0 %}
P e ious ideas and eedback:
{% o i em in p e ious_ esponses %}
- Idea {{ o loop.index }}: {{ i em. esponse }}
Feedback: {{ i em. eedback }}
{% end o %}
{% else %}
No p e ious ideas ha e been submi ed ye .
{% endi %}
Please p o ide cons uc i e eedback on he pa icipan 's
idea wi h a ocus on i s ele ance o ene gy usage and
sus ainabili y. Use he ollowing s uc u e:
1. S eng hs: [<DESCRIBE WHAT MAKES STRONG IDEAS>]
2. Oppo uni ies and ques ions: [sugges one o wo ways
he idea could be u he de eloped, and pose open-
ended ques ions ha help he pa icipan elabo a e,
cla i y assump ions, o explo e new angles]
{% i p e ious_ esponses.size > 0 %}
3. Rela ion o p e ious ideas: [explain how his idea
complemen s, imp o es, o di e s om ea lie ideas and
eedback]
4. Make su e he eedback is sho , wi h a maximum o
hi y wo ds.
5. Gi e a pic u e ela ed o he eedback.
6. I s uden 's inpu is sho e han en wo ds, eply:
Please use a ull sen ence o desc ibe you idea.
7. Gi e all ou pu in Du ch.
{% endi %}
P omp 1: P omp designed by g oup 1. Addi ions o
he empla e a e in bold. Double b acke s (
{{ }}
) and
pe cen age b acke s (
{% %}
) a e used o a iable
ende ing and con ol s uc u es espec i ely.
indica e ha o e ing eady- o-use empla es signi ican ly educes he ba ie o en y
o educa o s. Second, one g oup ailed o ecognize ha mos o hei p omp would
be igno ed due o he condi ional ende ing sys em used o p omp gene a ion (see
P omp 1). Condi ional ende ing condi ions pa s o he p omp based on he
p esence o p e ious esponses in he discussion h ead.
Du ing he wo kshop’s concluding discussion, pa icipan s e lec ed on he ac i i y
and sha ed hei expe iences bo h o ally and h ough he online collabo a ion boa d.
The i s obse a ion was ha p epa ing he ac i i y and designing he p omp p o ed
easie han expec ed. One pa icipan speci ically highligh ed he bene i o ha ing a
unc ional empla e as a s a ing poin . Howe e , ano he pa icipan coun e ed ha
he sys em’s use expe ience was ini ially unclea and equi ed guidance.
All pa icipan s ag eed ha he eache ’s abili y o p ecisely con igu e AI beha io
ep esen s signi ican alue. Ne e heless, hey emphasized he di icul y o
designing e ec i e p omp s and s essed ha guidance and aining would be
essen ial o educa o s. One pa icipan sugges ed ha p ede ined empla es, simila
o hose p o ided in his wo kshop, combined wi h clea guidelines in eg a ed in o he
use in e ace, could se e as simple and e ec i e s a egies o suppo eache s.
Building on his poin , ano he pa icipan no ed ha subs an ial li e a u e on
eedback in educa ional con ex s al eady exis s, om which design p inciples o
e ec i ely p omp ing AI models o deli e eedback could likely be de i ed.
Beyond he ou comes discussed abo e, we aimed o pa icipan s o gain
unde s anding o mul iple concep ual models o human-AI collabo a ion in
collabo a i e idea ion. Addi ionally, pa icipan s acqui ed expe ience in implemen ing
AI eedback using ou online collabo a i e idea ion applica ion, hos ed wi hin he LXP
G aasp.o g.
4 CONCLUSION
This wo kshop demons a ed he po en ial o in eg a ing con igu able AI eedback
wi hin collabo a i e idea ion pla o ms o enginee ing educa ion. Pa icipan s
ecognized AI’s capaci y o p o ide scalable, objec i e eedback while iden i ying
c i ical implemen a ion challenges.
The p ima y ba ie o adop ion is he complexi y o e ec i e p omp design. While
pa icipan s ound he con igu a ion p ocess manageable wi h unc ional empla es,
success ul implemen a ion equi es subs an ial suppo in as uc u e, including
eady- o-use empla es and comp ehensi e educa o aining.
A signi ican limi a ion eme ged h ough he sycophan ic beha io o cu en AI
models, which p oduced o e ly posi i e and e bose eedback. This inding
highligh s a undamen al ension: while pa icipan s alued AI’s objec i i y, cu en
models end owa d unhelp ully accommoda ing esponses ha may impede genuine
lea ning. This ein o ces he need o AI models wi h enhanced capaci y o
p oduc i e c i icism, such as an agonis ic AI (Cai e al., 2024).
These indings unde sco e he impo ance o de eloping con igu able human-AI
collabo a ion sys ems ha can e ec i ely enhance acili a ion p ac ices in
enginee ing educa ion and beyond.
NOTE
This a icle's language was e ined wi h he assis ance o gene a i e language
models.
REFERENCES
Cai, A., A awjo, I., & Glassman, E. L. (2024). An agonis ic AI (a Xi :2402.07350).
a Xi . h ps://doi.o g/10.48550/a Xi .2402.07350
F ich, J., MacDonald Ve meulen, L., Remy, C., Biskjae , M. M., & Dalsgaa d, P.
(2019). Mapping he Landscape o C ea i i y Suppo Tools in HCI.
P oceedings o he 2019 CHI Con e ence on Human Fac o s in Compu ing
Sys ems, 1–18. h ps://doi.o g/10.1145/3290605.3300619
F ich, J., Mose Biskjae , M., & Dalsgaa d, P. (2018). Twen y Yea s o C ea i i y
Resea ch in Human-Compu e In e ac ion: Cu en S a e and Fu u e
Di ec ions. P oceedings o he 2018 Designing In e ac i e Sys ems
Con e ence, 1235–1257. h ps://doi.o g/10.1145/3196709.3196732
Gille , D., Vonèche-Ca dia, I., Fa ah, J. C., Hoang, K. L. P., & Rod íguez-T iana, M.
J. (2022). In eg a ed Model o Comp ehensi e Digi al Educa ion Pla o ms.
2022 IEEE Global Enginee ing Educa ion Con e ence (EDUCON), 1587–
1593. h ps://doi.o g/10.1109/EDUCON52537.2022.9766795
Haase, J., & Hanel, P. H. P. (2023). A i icial muses: Gene a i e A i icial In elligence
Cha bo s Ha e Risen o Human-Le el C ea i i y. Jou nal o C ea i i y, 33(3),
100066. h ps://doi.o g/10.1016/j.yjoc.2023.100066
La Scala, J., Ba łomiejczyk, N., Gille , D., & Holze , A. (2025). Fos e ing Inno a ion
wi h Gene a i e AI: A S udy on Human-AI Collabo a i e Idea ion and Use
Anonymi y. P oceedings o he 58 h Hawaii In e na ional Con e ence on
Sys em Sciences, 337–351. h ps://doi.o g/10.24251/HICSS.2025.041
La Scala, J., Sahli, S., & Gille , D. (2025, Ap il). S imula ing B ains o ming Ac i i ies
wi h Gene a i e AI in Highe Educa ion. 2025 IEEE Global Enginee ing
Educa ion Con e ence, London, UK.
h ps://doi.o g/10.1109/EDUCON62633.2025.11016340
Memme , L., & Bi ne , E. (2024). Human-AI Collabo a ion o B ains o ming: E ec
o he P esence o AI Ideas on B ead h o Explo a ion. P oceedings o he
57 h Hawaii In e na ional Con e ence on Sys em Sciences. Hawaii
In e na ional Con e ence on Sys em Sciences, Hawaii.
h ps://hdl.handle.ne /10125/106425
Memme , L., C e ko ic, I., & Bi ne , E. (2024). The Mo e Is No he Me ie : E ec s
o P omp Enginee ing on he Quali y o Ideas Gene a ed By GPT-3. Hawaii
In e na ional Con e ence on Sys em Sciences.
h ps://doi.o g/10.24251/HICSS.2023.903
Schwabe, G., Ka siuba, D., Specke , R., & Dola a, M. (2025). AI-ThinkLe s o
B ains o ming. Hawaii In e na ional Con e ence on Sys em Sciences.
h ps://doi.o g/10.24251/HICSS.2025.043
Sha ma, M., Tong, M., Ko bak, T., Du enaud, D., Askell, A., Bowman, S. R., Cheng,
N., Du mus, E., Ha ield-Dodds, Z., Johns on, S. R., K a ec, S., Maxwell, T.,
McCandlish, S., Ndousse, K., Rausch, O., Schie e , N., Yan, D., Zhang, M., &
Pe ez, E. (2025). Towa ds Unde s anding Sycophancy in Language Models
(a Xi :2310.13548). a Xi . h ps://doi.o g/10.48550/a Xi .2310.13548
S e enson, C., Smal, I., Baas, M., & G asman, R. (2022). Pu ing GPT-3’s C ea i i y
o he (Al e na i e Uses) Tes (a Xi :2206.08932). a Xi .
h p://dx.doi.o g/10.48550/a Xi .2206.08932
