Lost in translation. Pilot research on the problem of translating human thoughts into the language of the machine

Anna Jaruga-Rozdolska, Artur Zaguła

doi:10.37190/arc250108

Abstract

The article presents the course of a pilot study conducted in early summer 2024 at the Institute of Architecture of the Lodz University of Technology. The initial study was intended to establish the scope of the research with the participation of a larger number of students, scheduled for the last quarter of the same year. Thus, both the pilot study and the subsequent follow-up research are aimed at determining and implementing the most efficient and productive way to alter the standards for educating future architects, adequate to the transformation catalysed by the technological boom tied to the development of artificial intelligence. Another focus of the study was to monitor user-script communication. Significantly, English was not the first language for the survey participants. Therefore, the study touches on the important problem of communicating thoughts sufficiently to achieve the expected results, identifying errors in communication due to both a lack of experience working with generative image models and the need to translate key words into English.

Full article view is only available on bigger screens.

2025

1(81)

Anna Jaruga-Rozdolska*, Artur Zaguła**

Lost in translation. Pilot research on the problem

of translating human thoughts into the language of the machine

DOI: 10.37190/arc250108

Published in open access. CC BY NC ND license

Abstract

The article presents the conduct of a pilot study carried out at the Institute of Architecture of the Lodz University of Technology in the summer of

2024. The subject of the research that architecture students took part in was to trace the process of creating architectural concepts using the generative

image model Midjourney by individuals with no previous exposure to AI tool-assisted design, for whom English – read by a tool capable of converting

a string of keywords into an image – is not their native language.

The authors’ goal was to capture diculties arising from translation errors and lack of script literacy. The precisely dened architectural design

problem was carried out by the participants of the study by selecting from among the images generated in the version 6.0 model on the basis of the

prompts they had previously prepared.

The results indicate the challenges arising from the need to translate thoughts into a language fully understandable by the machine. The title of

the article, referring to a movie Lost in translation directed by Soa Coppola, highlights the eeting essence of the problem: words are only a certain

signaller of meaning, which often fails in the full depiction of thought.

Key words: articial intelligence, AI architecture, Midjourney, AI for Higher Education

Introduction

The use of articial intelligence (AI)-assisted tools has

sparked the need for the development of a new user skill,

which is the ability to communicate the expectations

of a required result through properly entered key words

(prompts). While this communication is intuitive to some

extent, creating sets of prompts to produce precise, com-

pliant results still requires experience (Allen et al. 2023).

This fact conrms the existence of certain diculties in

communicating human thoughts in a way that can be pro-

perly interpreted by a machine. However, at the same

time, the increasing scope of implementation of the new

technology being a catalyst for the transformation of all

branches of science, art and industry leads us to reect on

the validity of the views of Charles Petzold, who in his

2000 publication Code: The Hidden Language of Com-

puter Hardware and Software stated that […] just as Latin

was for centuries the common language of scholars, so the

language of computers has become the common language

of our modernity (Petzold 2000)

. The neural networks of

most of the available generative scripts were trained on da-

tabases in English. Although some generative models are

now available in more language options, English remains

the dominant language for AI communication and devel-

opment.

In a pilot study conducted as an extension of the re-

search on the potential of usage of the Midjourney tool in

the future design practice of an architect (Jaruga-Rozdol-

ska 2022), an attempt was made to identify the scale of

potential diculties in communicating a precisely dened

Interestingly, similar predictions were made as early as 1993 in

a New York Times article that discussed the evolution of programming

languages, predicting that languages used for coding could eventually

become a universal form of communication (Marko 1993).

* ORCID: 0000-0003-4196-6437. Architectural Humanities Re-

search Association, Manchester, United Kingdom, e-mail: jaruga.anna@

protonmail.com

** ORCID: 0000-0002-8934-8630. Institute of Architecture, Lodz

University of Technology, Poland.

80 Anna Jaruga-Rozdolska, Artur Zaguła

design task using keywords by individuals who had no pre-

vious experience with AI-assisted scripts. Architecture stu-

dents from the Faculty of Construction, Architecture, and

Environmental Engineering at Lodz University of Tech-

nology took part in the study

. The article presents the rst

of a series of planned research, which is the starting point

for establishing new standards for the education perfor-

mance evaluation among students and an attempt to adapt

the current study program to the dynamically changing re-

quirements that the modern labour market will impose on

graduates. This is particularly important at the stage of just

forming regulations regarding the implementation of new

technologies.

The title of the article, with reference to Soa Coppola’s

lm work (Coppola 2003), points to the ephemeral essence

of the problem: a word is only a certain signaller of mean-

ing, which often fails in the full depiction of thought. An

additional level of diculty is the necessity for translation

– Midjourney’s script reads the keywords solely in English,

which is not the mother tongue for the survey participants.

Thus, multiple translations are evident here: human

thoughts into words, then those into English, and then

a sequence of keywords into an image. Although the script

uses a model that allows natural language interpretation

(NLP), it interprets the keyword sequence according to

the learned context of its meaning. The outcome depends

thus on three variables: the ability to describe an architec-

tural concept with words, the prociency in nding their

equivalents in a foreign language, and the degree of train-

ing of the model’s neural network capable of reading them

against the overall meaning.

State of research

The literature on the applications of generative articial

intelligence and its integration into architectural design is

undergoing extraordinary growth in recent years. Evidence

in this regard is the March 2024 review article Generative

AI for Architectural Design: A Literature Review (Li et al.

2025). Among other topics, it contains both background in-

formation on the principles of generative AI and a review

of the recent literature on the subject, citing 191 items on

the topic published between 2018 and 2024. According to

an article, its authors state that a clear trend of increasing

research indicates a growing inclination among the archi-

tectural design community to use generative articial in-

telligence. Such interest is reected both in mainstream ar-

chitecture and design portals such as Dezeen (Wiles 2022),

ArchDaily (Valencia 2023) and Designboom (Bhatia 2022;

Petridou 2022), as well as in the books that are being writ-

ten on the subject (Leach 2022; Bernstein 2022; Husam,

Ahmed 2023) and surveys by a respected institution – in-

cluding the Royal Institute of British Architects, which is-

sued a report on architects’ use of AI in 2024 (RIBA 2024)

Studies in this eld are conducted in Polish.

Answering the question, How do you generally assess your

knowledge of articial intelligence? 9% of respondents said they had

none, 51% said they had basic knowledge, 32% described their knowl-

edge as practical, only 6% had advanced knowledge, and 2% considered

It shows that knowledge on AI and its use in architecture re-

mains at an early stage. Research on generative AI, particu-

larly Midjourney, deals with aspects ranging from technical

(Ploe nnings, Berger 2023), educational, aesthetic (Rad-

hakrishnan 2023), cultural (Sukkar et al. 2024; Tanugraha

2023; Kamaoğlu 2023) or ethical (Paraman, Anamalah

2023), and even relating to case applications (Mousavi,

Dinçer 2024). Papers concerning communication between

the user and AI (Stanusch 2023; Allen et al. 2023; Mayers

2024; Zierock, Jungblut 2024), including prompt writing,

as well as those dealing with the application of AI in educa-

tion, including research conducted with students as future

users of generative tools (Tsidylo, Sendra 2023), were rele-

vant to the issues raised in the article.

Purpose of the research

The article presents the course of a pilot study conduct-

ed in early summer 2024 at the Institute of Architecture of

the Lodz University of Technology. The initial study was

intended to establish the scope of the research with the par-

ticipation of a larger number of students, scheduled for the

last quarter of the same year.

Thus, both the pilot study and the subsequent follow-up

research are aimed at determining and implementing the

most ecient and productive way to alter the standards for

educating future architects, adequate to the transformation

catalysed by the technological boom tied to the develop-

ment of articial intelligence. Another focus of the study

was to monitor user-script communication. Signicantly,

English was not the rst language for the survey partici-

pants. Therefore, the study touches on the important prob-

lem of communicating thoughts suciently to achieve the

expected results, identifying errors in communication due

to both a lack of experience working with generative image

models and the need to translate key words into English.

Methods

Fourteen students of the master’s degree program in

architecture pursued at the Institute of Architecture at the

Technical University of Lodz participated in the survey.

The participants were asked to prepare individual sets of

keywords referring to architectural form in in any context.

The task included a hint: try to imagine or recall in your

memory a building, thereafter describing it with words,

including in the descriptive language relating directly to

the architectural form (function of the building, mate-

rials used, style, scale, form and dominant colour) and

those containing suggestions about the context in which

the building will be presented (environment, time of day,

perspective). No minimum or maximum number of words

used was required. The prompts prepared by the students

were then used to prepare the architectural visualization in

the Midjourney version 6.0.

themselves an authority on the subject. To the question, how often is arti-

cial intelligence used in the projects you are currently working on? 2%

said in any, 4% said in most, 15% said in some projects, 20% said they

do it occasionally, and 59% said they have never used it (RIBA 2024).

Lost in translation. Pilot research on the problem of translating human thoughts into the language of the machine 81

First four images – marked A in the form – were gener-

ated on the basis of the original prompt prepared by each of

the study participants. Afterwards, the prompt was edited,

correcting any grammatical and linguistic errors that might

cause it to be misinterpreted by the script – the images

generated on the basis of the prompt revised in this way

constituted the group of images marked with the letter B.

Further, bearing in mind that the authors of the keyword

sequences have no experience in working with a similar

tool, a kind of bridge in the communication between the

human and the tool was created: both the original and the

corrected set of prompts were entered into the generative

text model – ChatGPT version 4.0. The word sequences

expanded in such a manner

contributed to the generation

of successive sets of images, marked successively with the

letter C for the original prompt corrected by the generative

text model, and with the letter D for the previously linguis-

tically corrected prompt expanded by ChatGPT (Fig. 1).

While not without drawbacks

, the decision was made un-

der the assumption that translating the human thought tran-

script into script-generated text could improve the quality

of the resulting images, especially when the user is an in-

dividual with no prior experience in writing prompts and

for whom English is not a native language. The keyword

sequences generated using the generative text model were

signicantly longer and more descriptive (Fig. 2).

All sixteen images (four corresponding to each prompt)

were then placed in each of fourteen individually prepared

questionnaires – each participant in the study evaluated

The form of the query: please provide a Midjourney prompt to

generate… – in the place of an ellipsis, the original and revised set of

keywords were inserted sequentially.

The generative image model should interpret the keyword se-

quence without requiring external scripts to improve the user prompt.

solely the images generated from their own set of key-

words. Participants were asked to evaluate the aesthetic

and content quality

of each image. They further made

a selection of the best and worst visuals in each category

In the nal question, participants were asked to identify the

visualization they would most like to use as inspiration for

a full architectural concept

Results

During the conducted study, a total of 224 images were

generated. The selected sets of four images are shown be-

low, representing as follows: the visualizations generated

for each group in alphabetical order along with the initial

prompt prepared by the survey participant (Figs. 3, 4), ex-

amples of pairs of the best and worst rated images (Fig. 5),

and the visualizations that participants indicated they

would be most likely to use as the impetus for an architec-

tural design (Fig. 6).

The survey results in percentage form showed the fol-

lowing:

1. Participants most often pointed to images in group C

(original prompt expanded using ChatGPT’s generative

text model version 4.0) as the aesthetically best images

– they accounted for 32.14% of all answers given. In the

same category, the second most frequently selected images

were those from group D (the language-corrected prompt

Dened as compliance with the original set of keywords.

Shown in a tabular summary in the following sections of the arti-

cle as BEST E for the aesthetically best visualization, WORST E for the

lowest rated visualization in this category, and BEST M and WORST M

similarly for the substantive evaluation of the images.

According to a tabular summary in the following sections of the

article, dened by the INSPO acronym.

Fig. 1. Diagram showing how

individual groups of images

were created using the generative

Midjourney image model

(elaborated by

A. Jaruga-Rozdolska)

Il. 1. Schemat tworzenia

poszczególnych grup obrazów

za pomocą generatywnego

modelu obrazowego Midjourney

(oprac. A. Jaruga-Rozdolska)

82 Anna Jaruga-Rozdolska, Artur Zaguła

Fig. 3. A set of images generated from the keyword sequence:

Modern university building in the shape of the letter “X”, vertical. Colourful glazing, contemporary architecture, as high as the Eiffel tower.

Located in Paris, historical surroundings. Sunny weather, some zeppelins flying high selected consecutively from the images of groups A, B, C and D

(elaborated by A. Jaruga-Rozdolska; images generated using Midjourney based on phrases provided by the study participant)

Il. 3. Zestaw obrazów wygenerowanych na podstawie ciągu słów kluczowych:

Modern university building in the shape of the letter „X”, vertical. Colorful glazing, contemporary architecture, as high as the Eiffel tower.

Located in Paris, historical surroundings. Sunny weather, some zeppelins flying high wybranych kolejno spośród obrazów z grup A, B, C i D

(oprac. A. Jaruga-Rozdolska; obrazy wygenerowane za pomocą Midjourney na podstawie fraz dostarczonych przez uczestnika badania)

Fig. 4. A set of pictures generated based on the keyword string: Skyscraper, antique, inspired by ancient Rome; stone/sandstone; natural colours

mixed with red, blue and yellow; large city; early sunny morning, summer selected consecutively from images in groups A, B, C and D

(elaborated by A. Jaruga-Rozdolska; images generated using Midjourney based on phrases provided by the study participant)

Il. 4. Zestaw obrazów wygenerowanych na podstawie ciągu słów kluczowych: Skyscraper, antique, inspired by ancient Rome; stone/sandstone;

natural colors mixed with red, blue and yellow; large city; early sunny morning, summer wybranych kolejno spośród obrazów z grup A, B, C i D

(oprac. A. Jaruga-Rozdolska; obrazy wygenerowane za pomocą Midjourney na podstawie fraz dostarczonych przez uczestnika badania)

Fig. 2. An example of

the transformation

of the original prompt

(elaborated by

A. Jaruga-Rozdolska)

Il. 2. Przykład transformacji

oryginalnego promptu

(oprac. A. Jaruga-Rozdolska)

Lost in translation. Pilot research on the problem of translating human thoughts into the language of the machine 83

enhanced using ChatGPT’s generative text model version

4.0), indicated by 28.57% of respondents. The least fre-

quent (17.86%) images in this category were indicated as

the best in terms of group A (original keyword sets provid-

ed by survey participants).

2. Most often, images from group B (32.14%) were

in dicated as the worst in terms of aesthetics; it was fol-

lowed by images from group D (28.57%), C (21.43%) and

A (17.86%) being selected most frequently.

3. The images in group C ranked best in terms of con-

tent (32.14% of responses given); the second most statisti-

cally common response was from group B (25%). Images

from groups A and D were indicated the least frequently in

this category (21.43%).

4. Remarkably, Group A images (generated from unmo-

died sets of keywords provided by survey participants) were

considered as the worst most frequently in terms of compa-

tibility with the original prompt – up to 42.89% of respon-

dents selected this answer. On the other hand, group C images

were the least often chosen by students in this category.

5. Group A images were also selected least frequently as

those stimulating a full architectural concept. The results

of the pilot study and the results in percentage terms are

presented in the table (Fig. 7).

Fig. 5. A set of top and worst rated images generated based on the keyword sequence:

Religious, brick, sheet metal, glass, neo-gothic, mustard yellow, tarnished copper, low and mid-rise buildings, night,

respectively – the best and worst image in terms of aesthetics, and the best and worst image in terms of content

(elaborated by A. Jaruga-Rozdolska; images generated using Midjourney based on phrases provided by the study participant)

Il. 5. Zestaw najlepiej i najgorzej ocenianych obrazów wygenerowanych na podstawie ciągu słów kluczowych:

Religious, brick, sheet metal, glass, neo-gothic, mustard yellow, tarnished copper, low and mid-rise buildings, night,

kolejno – najlepszy i najgorszy obraz pod względem estetycznym oraz najlepszy i najgorszy obraz pod względem merytorycznym

(oprac. A. Jaruga-Rozdolska; obrazy wygenerowane za pomocą Midjourney na podstawie fraz dostarczonych przez uczestnika badania)

Fig. 6. Example of visualizations identified as those with the greatest potential to be used as inspiration for a full architectural concept

(elaborated by A. Jaruga-Rozdolska; images generated using Midjourney based on phrases provided by the study participant)

Il. 6.

Przykład wizualizacji wskazanych jako te o największym potencjale do wykorzystania jako inspiracja do stworzenia pełnej koncepcji architektonicznej

(oprac. A. Jaruga-Rozdolska; obrazy wygenerowane za pomocą Midjourney na podstawie fraz dostarczonych przez uczestnika badania)

84 Anna Jaruga-Rozdolska, Artur Zaguła

Additional research

To supplement the above pilot research, one of the con-

tributors to the study, Artur Zaguła, conducted his own in-

vestigation. It was intended to address not only the language

of prompt generation, but also the non-English-speaking

cultural area. In order to achieve this, a prompt was writ-

ten in English giving the command to generate a building

based on a well-known example of Polish building culture

– the Wilanów Palace. The rst prompt was: Baroque resi

dence in central Poland based on Wilanów Palace; the next

prompt specied the construction age: 17

century baroque

residence in central Poland based on Wilanów Palace. Both

resulted in poor results clearly demonstrating the decien-

cies in the Mid journey script’s tutoring of Polish examples

of historical architecture (Fig. 8). The next step involved

writing a prompt in Polish to see how the script would be-

have when receiving a request in this language. The prompt

Highest score:

image aesthetics

criterion [%]

Lowest score:

image aesthetics

criterion [%]

Highest score:

keyword sequence

matching criterion [%]

Lowest score:

keyword sequence

matching criterion [%]

Images most frequently

selected as most inspiring

among others [%]

A = 17.86 A = 17.86 A = 21.43

A = 42.86

A = 14.29

B = 21.43

B = 32.14

B = 25.00 B = 21.43

B = 28.57

C = 32.14

C = 21.43

C = 32.14

C = 7.14

C = 28.57

D = 28.57 D = 28.57 D = 21.43 D = 28.57

D = 28.57

Fig. 7. Tabular representation of the results of the conducted surveys.

The highest percentage score in a given category is highlighted in bold (elaborated by A. Jaruga-Rozdolska)

Il. 7. Tabelaryczne przedstawienie wyników przeprowadzonych badań.

Pogrubionym pismem wyróżniono najwyższy procentowy wynik w danej kategorii (oprac. A. Jaruga-Rozdolska)

Fig. 8. Images generated on the basis of the prompt: 17

century baroque residence in central Poland based on Wilanów Palace

(elaborated by A. Zaguła with the use of Midjourney tool)

Il. 8. Obrazy wygenerowane na podstawie promptu: 17

century baroque residence in central Poland based on Wilanów Palace

(oprac. A. Zaguła z użyciem narzędzia Midjourney)

Lost in translation. Pilot research on the problem of translating human thoughts into the language of the machine 85

ran as follows: siedemnastowiecznarezydencjawśrodkowej

Polsce oparta na Pałacu w Wilanowie, and the generated

images completely puzzled the researcher (Fig. 9). It de-

mon strated a complete misunderstanding of the command

and proved that being situated outside the English-speak-

ing world, one cannot count on anything even close to the

correct results. The nal step involved using the describe

command for the creation of a prompt from a photograph of

the Wilanów Palace

(Fig. 10); four prompts generated had

factual errors, in none of which the Wilanów Palace name

appeared. However, after corrections made by the research-

er, clearly the best results were achieved using such method.

Conclusions

It must rst be stated that when using the Midjourney

script to generate images from the eld of architecture, ex-

tensive and linguistically well-constructed sets of keywords

Using Midjourney’s generative script, the describe module al-

lows the process to be reversed: the user uploads an image of their choice

and, based on that image, forces the describe command to generate four

options for a textual description of the image. Each of the prompts gen-

erated thus can then be used to generate a new set of images.

– prompts – are essential. In addition to subject matter

knowledge, it is important to know how to properly con-

struct word prompts and prociency in working with gen-

erative image models. According to the study, additional

diculties may arise from the need to translate the expected

outcome into English. The pilot study ndings show that im-

ages generated from prompts created by non-English-speak-

ing users were evaluated the lowest by the users themselves.

Both in terms of aesthetics, consistency with the prompt,

and a potential as an inspiration to create a full architectural

concept, they received the lowest ratings. Both in terms of

aesthetics, consistency with the prompt, and inspiration to

create a full architectural concept, those images received the

lowest ratings. The second important nding of the study is

that the images generated according to the prompt correct-

ed by the generative text model were rated best. Prompts

enhanced with chat GPT were much more elaborate; being

based not on human linguistic structures, but those devel-

oped in the course of articial intelligence evolution, they

produced more appealing results. What this suggests per-

haps is that a language tailored to the way machine neural

networks read words is a new kind of language that we, as

humanity, must begin to learn if we want to communicate

with AI, which, given the pace and scope of the changes

Fig. 9. Images generated from the prompt: siedemnastowiecznarezydencjawśrodkowejPolsceopartanaPałacuwWilanowie

(elaborated by A. Zaguła with the use of Midjourney tool)

Il. 9. Obrazy wygenerowane na podstawie promptu: siedemnastowiecznarezydencjawśrodkowejPolsceopartanaPałacuwWilanowie

(oprac. A. Zaguła z użyciem narzędzia Midjourney)

Fig. 10. Images generated based on the prompt: Aerial view of Wilanów Palace, richly decorated in yellow and green with intricate details,

surrounded by lush gardens under clear blue skies, showcasing its grandeur as one of Poland’s most iconic landmarks.

The palace stands alone against an expansive background, embodying elegance and historical significance, wide shot, ultra realistic photography

(elaborated by A. Zaguła with the use of Midjourney tool)

Il. 10. Obrazy wygenerowane na podstawie promptu: Aerial view of Wilanów Palace, richly decorated in yellow and green with intricate details,

surrounded by lush gardens under clear blue skies, showcasing its grandeur as one of Poland’s most iconic landmarks.

The palace stands alone against an expansive background, embodying elegance and historical significance, wide shot, ultra realistic photography

(oprac. A. Zaguła z użyciem narzędzia Midjourney)

86 Anna Jaruga-Rozdolska, Artur Zaguła

unfolding, seems inevitable. Notably, prompt C won all

three categories in the respondents’ assessment, although as

inspiration it received the same score as prompts B and D.

The equal frequency of identifying images from groups B,

C and D as the most appealing in terms of their use as the

basis for a full architectural concept leads to the collateral

conclusion that the elimination of linguistic errors should

be considered a sucient correction, and that a good pro-

ciency in English should be considered a necessary skill in

working with the Midjourney model. The expanded prompt,

achieved through the use of ChatGPT in the process, im-

proved the aesthetics of the whole image, inuencing factors

such as lighting, context and camera angle, but for respon-

dents with architectural background, the presentation itself

appeared to be secondary to the form and aesthetics of the

building and the materials used in the design

. Due to the

small number of respondents, the results may not be entirely

conclusive, but we made the decision to publish them since

research in the use of AI, including image generation in the

eld of architecture, is critically needed today. Collecting

data to accelerate the ability to assess the usefulness and use

of AI is the imperative of the present time.

This data was contained in original prompts prepared by the par-

ticipants.

Future research

The initial pilot study presented in the article provides

a basis for the formulation of strategies for experiment rep-

etition with a greater group of respondents. The analysis of

the process of creating an architectural concept using gen-

erative image models, on the other hand, forms the basis

of a scenario for a summer workshop for students aimed at

demonstrating eective selection of words to get the result

closest to the intended one.

Recognizing, that students – and, consequently, future

architects – will use AI more frequently in the design pro-

cess, eorts must be made to develop new teaching strate-

gies and criteria for evaluating the projects carried out by

students in the course of their education. The curriculum

should undergo necessary updates in the face of transfor-

mations.

The ongoing technological revolution obliges us as re-

searchers to take responsibility, especially since we are

also operating in the eld of education. We envision the

university as an institution responsible for responding to

today’s challenges and striving to keep abreast of and stra-

tegically respond to the newest advancements.

Translated by

Anna Jaruga-Rozdolska

References

Allen, Laura D., Wengian Xu, Mariko Nishikitani, Vaishnavi Atul Patil,

and Dana Bradley. “Age Bias in Articial Intelligence (AI): A Visual

Properties Analysis of AI Images of Older versus Younger People.”

Innovation in Aging 7, no. S1 (2023): 986. https://doi.org/10.1093/

geroni/igad104.3168.

Bernstein, Phil. Machine Learning.ArchitectureintheAgeofArticial

Intelligence. RIBA Publishing, 2022.

Bhatia, Manas. “Midjourney Envisions a Futuristic Sustainable City with

Air-Purifying Biophilic Skyscrapers.” Published August 22, 2022.

Accessed November 20, 2024. https://www.designboom.com/ar-

chitecture/ai-futuristic-sustainable-city-air-purifying-biophilic-sky-

scrapers-manas-bhatia-08-22-2022/.

Coppola, Soa, director. “Lost in Translation.” 1 h 42 min. Universal,

2003.

Husam, Husain, and Nia Hourahksh Ahmed, eds. Beyond Blueprint: Ad-

vencements in Architectural Education and Innovations. Yayinlari

Cinius, 2023.

Jaruga-Rozdolska, Anna. “Articial Intelligence as Part of Future Prac-

tices in the Architect’s Work: MidJourney Generative Tool as Part of

a Process of Creating an Architectural Form.” Architectus, no. 3(71),

(2022): 95–104. https://doi.org/10.37190/arc220310.

Kamaoğlu, Melih. “The Idea of Evolution in Digital Architecture: To-

ward United Ontologies?” International Journal of Architectur-

al Computing 21, no. 4 (2023): 622–34. https://doi.org/10.1177

/ 14780771231174890.

Leach, Neil. ArchitectureintheAgeofArticial Intelligence.AnIntro-

duction to AI for Architects. Bloomsbury Academic, 2022.

Li, Chengyuan, Tianyu Zhang, Xusheng Du, Ye Zhang, and Haoran Xie.

“Generative AI for Architectural Design: A Literature Review.”

Frontiers of Architectural Research 14, no. 3 (2025): 759–83. https://

doi.org/10.1016/j.foar.2024.10.001.

Marko, John. “Will Computer Code Be the New English?” The New

York Times. Published December 8, 1993. Accessed November 15,

2024, at https://www.nytimes.com/1993/12/08/business/business- tech -

no lo gy -a-free-and-simple-computer-link.html.

Mayers, Jerey. “The Importance of Semantics and Syntax in Creating

Text-to-Image Prompts Using Midjourney AI.” Published June,

2024. Accessed November 23, 2024. https://doi.org/10.13140/RG. -

2. 2.21628.14723.

Mousavi, Seyed Mohammad Amin Mostafavi, and Ebru Erdönmez

Dinçer. “Revitalising and Renovating the Visitor Experience in the

Historic Piazza San Marco; Based on the Installation Project Gener-

ated by Midjourney AI.” Journal of Technology in Architecture, De-

sign and Planning 2, no. 1 (2024): 44–51. https://doi.org/ 10. 26650/

JTADP.24.005.

Paraman, Pradeep, and Sanmugam Anamalah. “Ethical Articial Intel-

ligence Framework for a Good AI Society: Principles, Opportu-

nities and Perils.” AI & Society 38 (2023): 595–611. https://doi.org/

10.1007/s00146-022-01458-3.

Petridou, Christina. “Hassan Ragan questions the future of architecture

in the hands of emerging AI tools.” Published October 9, 2022.

Ac cessed November 20, 2024. https://www.designboom.com/ -

ar chi tec ture/ hassan-raga-parametric-architecture-ai-tools-mid-

journey-10-09-2022/.

Acknowledgements

Authors would like to express their sincere acknowledgment to the Archi-

tecture Students of the Lodz University of Technology for participating

in the study.

Lost in translation. Pilot research on the problem of translating human thoughts into the language of the machine 87

Streszczenie

Utracone w tłumaczeniu. Pilotażowe badania nad problemem tłumaczenia ludzkich myśli na język maszyn

W artykule przedstawiono przebieg badań pilotażowych przeprowadzonych w Instytucie Architektury Politechniki Łódzkiej latem 2024 r. Wzięli

w nich udział studenci kierunku architektura. Tematem badań było prześledzenie procesu tworzenia koncepcji architektonicznych za pomocą genera-

tywnego modelu obrazowego Midjourney przez osoby niemające wcześniej styczności z projektowaniem wspomaganym narzędziami AI, dla których

to osób język angielski – odczytywany przez narzędzie zdolne przetwarzać ciąg słów kluczowych na obraz – nie jest językiem ojczystym. Celem

autorów było uchwycenie trudności wynikających z błędów translacyjnych oraz braku umiejętności posługiwania się skryptem. Precyzyjnie określony

architektoniczny problem projektowy uczestnicy badania realizowali, wybierając spośród obrazów wygenerowanych w modelu w wersji 6.0 na podsta-

wie uprzednio przygotowanych przez nich promptów. Uzyskane rezultaty wskazują na trudności wynikające z konieczności przełożenia myśli na język

w pełni zrozumiały dla maszyny. Tytuł artykułu, nawiązujący do lmu Międzysłowami w reżyserii Soi Coppoli, podkreśla ulotną istotę problemu:

słowa są jedynie pewnym sygnalizatorem znaczenia, który często zawodzi w pełnym zobrazowaniu myśli.

Słowa kluczowe: sztuczna inteligencja, AI, sztuczna inteligencja w architekturze, Midjourney, AI w szkolnictwie wyższym

Petzold, Charles. Code: The Hidden Language of Computer Hardware

and Software. Microsoft Press, 2000.

Ploennings, Joern, Markus Berger. “AI Art in Architecture.” AI Civil Engi-

neering 2, art. 8 (2023). https://doi.org/10.1007/s43503-023-00018-y.

Radhakrishnan, Mohesh. “Is Midjourney-Ai the New Anti-Hero of Archi-

tectural Imagery & Creativity?” GlobalScienticJournal 11, no. 1

(2023): 94–104.

RIBA. “RIBA AI REPORT 2024”. Published February 29, 2024. Accessed

November 15, 2024, at https://www.architecture.com/knowledge-

and-resources/resources-landing-page/riba-ai-report-2024?srslti-

d=AfmBOoqCZFdMzbtK2KbNQu-5ep6TziTrUlKzivthXLToMQ-

127FYaT9tV#available-resources.

Stanusch, Natalia. “Wor(l)d-Image Trans-formation: Looking through

DALL-E 2 and Midjourney.” La Valle dell’Eden, no. 41–42 (2023):

85–94. https://ojs.unito.it/index.php/vde/article/view/10831.

Sukkar, Ahmad W., Mohamed W. Fareed, Moohammed Wasim Yahia, Sa-

lem Buhashima Abdalla, Iman Ibrahim, and Khaldoun Abdul Karim

Senjab. “Analytical Evaluation of Midjourney Architectural Virtual

Lab: Dening Major Current Limits in AI-Generated Representa-

tions of Islamic Architectural Heritage.” Buildings 14, 3 (2024): 786.

https://doi.org/10.3390/buildings14030786.

Tanugraha, Stephen. “A Review Using Articial Intelligence-Generating

Images: Exploring Material Ideas from MidJourney to Improve Ver-

nacular Designs.” JournalofArticialIntelligenceinArchitecture2,

no. 2 (2023): 48–57. https://doi.org/10.24002/jarina.v2i2.7537.

Tsidylo, Ivan M., and Chele Esteve Sendra. “Articial Intelligence as

a Methodological Innovation in the Training of Future Designers: Mi-

djourney Tools.” Information Technologies and Learning Tools 97,

no. 5 (2023): 203–214. https://doi.org/10.33407/itlt.v97i5.5338.

Valencia, Nicolás. “The Impact of AI Tools on Architecture in 2024 (and

Beyond).” Published December 6, 2023. Accessed November 15,

2024, at https://www.archdaily.com/1009545/the-impact-of-ai-tools-

on-architecture-in-2024-and-beyond.

Wiles, Will. “Architects Can Rest Easy That AI Isn’t Coming for Their

Jobs Just Yet.” Published November 16, 2022. Accessed Novem ber

12, 2024, at https://www.dezeen.com/2022/11/16/architects-ai- dall-

e-midjourney-opinion/.

Zierock, Benjamin, and Asmar Jungblut. “Leveraging Prompts for Im-

proving AI-Powered Customer Service Platforms: A Case Study of

Chat GPT and Midjourney.” Published March, 2023. Accessed No-

vember 26, 2024. https://doi.org/10.13140/RG.2.2.17800.29441.