Teaching with Artificial Intelligence in Architecture: Embedding Technical Skills and Ethical Reflection in a Core Design Studio

Abstract

This case study examines the integration of artificial intelligence (AI) into undergraduate architectural education through a 2024–25 core studio teaching experiment at Zhejiang University. A dual-module framework was implemented, comprising a 20 h AI skills training module and in-class ethics discussions, without altering the existing studio structure. The AI skills module introduced deep learning models, LLMs, AIGC image models, LoRA fine-tuning, and ComfyUI, supported by a dedicated technical instructor. Student feedback indicated phase-dependent and tool-sensitive engagement, and students expressed a preference for embedded ethical discussion within the design studio rather than separate formal instruction. The experiment demonstrated that modular AI education is both scalable and practical, highlighting the importance of phase-sensitive guidance, balanced technical and ethical framing, and institutional support such as cloud platforms and research-based AI tools. The integration enhanced students’ digital adaptability and strategic thinking while prompting reflection on issues such as authorship, algorithmic bias, and accountability in human–AI collaboration. These findings offer a replicable model for AI-integrated design pedagogy that balances technical training with critical awareness.

1. Introduction

Since the early twentieth century, architectural design has undergone a profound technological transformation—from hand-drawn, intuition-driven practices to computer-aided drafting, parametric modeling, and, most recently, the integration of artificial intelligence (AI) into design workflows [1,2]. AI’s data-driven generation, model-based reasoning, and real-time feedback mechanisms have redefined both the cognitive and operational dimensions of design, shifting the paradigm from human intuition to algorithmic logic [3]. The development of AI-assisted design workflows, knowledge graph databases, and augmented creative platforms has become a critical strategy for advancing industry competitiveness [4,5] and reflects a broader transition from expert systems toward hybridized human–AI collaboration [6]. This shift signals not only a technological upgrade but also a deeper transformation in how design knowledge is produced, moving from hypothesis-driven deduction and intuition to data-driven pattern recognition and computational prediction [7]. In architectural education, this paradigm shift manifests in the reconfiguration of teaching logic: traditional studio models centered on conventional drafting and representational skills are increasingly replaced by new paradigms that emphasize modeling systems, generative processes, and iterative feedback structures [8]. AI-driven educational transformation, therefore, is not merely a matter of tool substitution, but it is a structural shift from inference-based knowledge construction to data-driven discovery.

1.1. AI in Architectural Practice

In architectural practice, the application of AI has progressed from computational assistance and visualization to active involvement in generative design processes, especially through AI-generated content (AIGC). Leading firms have begun developing customized AI platforms to enhance early-stage design capacity and improve decision-making efficiency. For instance, Foster + Partners developed and deployed the intelligent tool Hydra between 2019 and 2022, generating over 240,000 design iterations and running more than 1.3 million simulations, which significantly accelerated the design process [9]. Since 2022, Zaha Hadid Architects has incorporated tools such as Midjourney, Gendo, and in-house-developed systems, reportedly doubling or tripling design productivity during competition phases while reducing rendering time by 80% [10].

1.2. AI in Architectural Education

This shift in practice has triggered corresponding changes in architectural education, prompting curriculum reform at leading institutions worldwide. Harvard Graduate School of Design (GSD) has integrated AI tools into standard design support systems, established an internal digital platform, and issued evolving guidelines on AI usage, pedagogy, and ethics [11]. In early 2025, GSD also launched a short course titled AI, Machine Learning, and the Built Environment aimed at non-technical learners, focusing on applications of AI and machine learning in real estate, architecture, landscape, and urbanism [12]. Columbia Graduate School of Architecture, Planning and Preservation (GSAPP) has offered electives such as Spatial AI and AI for Existing Buildings since 2022, while TU Delft introduced a global-access MOOC titled AI in Architectural Design: Introduction [13,14,15].

In China, top-ranked universities—those comparable to Zhejiang University within the QS Top 50 for Architecture & Built Environment—are also introducing AI-related coursework. Tongji University launched a graduate course on AI-assisted architectural programming in 2023, incorporating tools like Midjourney, ChatGPT, and Stable Diffusion. Student feedback indicated remarkable improvements in creativity and efficiency, though concerns emerged regarding fragmented workflows and limited control over generated outcomes [16]. In the same year, Tsinghua University offered an interdisciplinary undergraduate course titled AI-Generated Imagery, guiding students through end-to-end short film production using tools like Midjourney and Runway [17]. In 2024, Southeast University introduced a senior undergraduate studio on Adaptable Housing, which integrated generative algorithms, Processing, and Multi-Objective Evolutionary Algorithms (MOEAs) to automate spatial layout design [18].

1.3. Educational Feedback and Pedagogical Challenges

Despite growing experimentation, integrating AI into design education remains fraught with challenges. On the cognitive level, while AI tools lower technical and representational barriers, their “black box” nature raises concerns about authorship, control, and transparency [19,20]. Students may struggle to understand the underlying logic of AI outputs, leading to reduced creative control and difficulties articulating structured design reasoning. In design education, there is still a lack of systematic development of technology-integrated competencies, and insufficient emphasis on structured multidisciplinary collaboration further exacerbates the gap between pedagogy and industry innovation [21,22,23]. Survey- and interview-based studies further highlight widespread enthusiasm among students for AI tools, but students also report frustration over technical barriers, homogenized outputs, algorithmic bias, and ethical ambiguity [16,24,25,26,27,28]. While AI integration has been widely explored in disciplines such as engineering and computer science—often through lecture-based courses and assignment-driven evaluation—the unique structure of architectural design studios, characterized by one-to-one tutoring and open-ended project development, poses distinct pedagogical challenges and opportunities. In response to these challenges, recent global discussions have called for a triadic governance approach comprising human oversight, algorithmic transparency, and social accountability [29,30], thereby underscoring the risk of AI-driven value erosion in educational contexts.

1.4. Summary

Based on the above context, this study draws on a teaching experiment conducted in the core design studio Architectural Design III/IV (2024–2025) at the School of Architecture, Zhejiang University. The studio features a dual-module framework that integrates AI skills training with AI ethics reflection, aiming to investigate how AI technologies can be systematically embedded into the core curriculum of undergraduate architectural design education. By analyzing implementation processes and student responses, this study seeks to contribute practical insights toward building a pedagogy that aligns technological literacy with cognitive development and ethical awareness. This study is guided by the following four research objectives:

• To develop a pedagogical framework for systematically embedding AI content into undergraduate core design studios while maintaining disciplinary coherence.
• To investigate how AI tools are applied across different stages of the design process and how student engagement varies accordingly.
• To evaluate students’ perceptions of AI in supporting creative development, managing expression, and fostering ethical awareness, based on feedback and in-studio observations.
• To identify key practical challenges and opportunities associated with AI integration.

The remainder of this paper is structured as follows: Section 2 introduces the course design and the dual-module pedagogical framework, along with the survey design used to collect student feedback. Section 3 presents the studio implementation and illustrates how AI tools were applied across key phases of the design process, accompanied by an overview of the survey results. Section 4 discusses student responses, pedagogical implications, and challenges identified during the experiment. Section 5 concludes this study.

2. Course Design and Pedagogical Framework

2.1. Context of Curriculum Reform

Over the past decade, the integration of emerging technologies into architectural education has become a growing priority for Zhejiang University’s School of Architecture. This trajectory was marked by the establishment of the Laboratory for Artificial Intelligence and Robotic Construction in 2018, a strategic milestone in advancing both research and pedagogy in computational and AI-assisted design [31]. Since then, the school has steadily introduced emerging technologies into the curriculum, initially through technical electives and gradually through broader pedagogical restructuring. A major curriculum reform was launched in 2024 for incoming undergraduate cohorts, aimed at systematically integrating AI and digital design technologies into core architectural education. This reform introduced a required course, Fundamentals of Artificial Intelligence (B), and expanded offerings through new electives, modular restructuring, and short-term intensive workshops.

The revised curriculum adopts a three-tiered model—foundation, integration, and innovation—to embed intelligent technologies across the undergraduate design sequence (Figure 1). These developments are framed within the program’s five-year design studio sequence, commonly referred to as a “3 + 1 + 1” structure [32] (Figure 2). Grade 1 focuses on developing digital literacy through AutoCAD drafting, SketchUp/Rhino modeling, and D5/Enscape rendering. Grade 2 advances computational design thinking with parametric generation (Grasshopper), environmental analysis (Climate Consultant), and spatial syntax techniques (Depthmap). In Grade 3, students engage in AI-driven innovation, applying deep learning models of various types: large language models (LLMs) for text-based reasoning, AIGC image models for visual exploration, and multimodal deep learning models for integrating textual, visual, and spatial inputs in design tasks. In Years 4 and 5, the use of intelligent design tools becomes more optional and diversified, allowing students to self-direct their engagement with advanced technologies.

Within this structure, the third-year Architectural Design III/IV studio serves as a critical inflection point, providing both the pedagogical continuity and technical readiness necessary for piloting AI-integrated design education.

2.2. Overview of Architectural Design III/IV Studio

Building on the technical foundations and digital literacy developed in the first two years, the Architectural Design III/IV studio represents a key stage for students to engage in complex design exploration through the application of advanced methods and emerging technologies. The year-long course follows a progressive pedagogy and traditionally includes four design tasks—ranging from building renovation to urban regeneration—each designed to address varying levels of constraint, system complexity, conceptual openness, and site specificity.

While maintaining the original course structure, this teaching experiment investigated how AI tools could be meaningfully embedded in the design studio setting. This initiative was developed in response to three key challenges:

• The integration of AI technologies into architectural pedagogy remains in its nascent stage, with few established instructional models or empirically grounded precedents to guide curriculum development.
• Students predominantly rely on conventional design thinking and procedural habits, which limits their ability to engage with AI tools in a conceptually coherent and methodologically structured manner.
• Students’ understanding of AI is often fragmented and disproportionately focused on visual outputs, highlighting the urgent need for a more structured integration throughout the full design process.

Addressing these challenges required embedding AI not as a replacement but as an enhancer within the studio’s pedagogical structure, fostering both technical proficiency and critical engagement.

2.3. Modular Integration of AI Skills and AI Ethics Components

To extend students’ engagement with AI beyond tool exposure, the studio introduced two complementary modules—AI skills and AI ethics (

Table 1. The integration of AI modules into the original Architectural Design III/IV studio curriculum.

Week	Class-Hours 1	Design Task	AI Ethics Module	AI Skills Module 2
Fall 1–8	64	Architecture & Re-build	Ethic Ⅰ: Data Privacy and Copyright	Skill Ⅰ (2 class-hours): Foundations of AI
				Skill Ⅱ (4 class-hours): Prompt Engineering and Image Generation
Fall 9–16	64	System & Synthesis	Ethic Ⅱ: Accountability in Human–AI Collaboration	Skill Ⅲ (4 class-hours): LoRA Model Training and Fine-tuning
				Skill Ⅳ-1 (3 class-hours): ComfyUI Workflow—Basics
Summer 1–4	32	X & Hypothesis	Ethic Ⅲ: Stylistic Authorship and Cultural Bias	Skill Ⅳ-2 (3 class-hours): ComfyUI Workflow—Advanced
Summer 5–16	96	Urban & Renewal	Ethic Ⅳ: Model Bias	Skill Ⅴ (4 class-hours): Multimodal Integration

¹ At Zhejiang University, 1 class hour is equivalent to 45 min. ² AI skills modules have extra class-hours.

).

The AI skills module provided 20 class-hours of structured technical instruction, covering a sequence of core technologies: LLMs, AIGC image tools, LoRA-based model fine-tuning, multimodal workflows using ComfyUI, and deep learning models. Instruction combined concise lectures with hands-on exercises, supported by the university’s e-learning platform and cloud-based computational infrastructure, enabling flexible participation online or in person. To support varying hardware conditions, students could either install Stable Diffusion and ComfyUI locally or access a university-provided cloud platform where related models and workflows were readily available. The course further maintained and updated downloadable resources in line with emerging open-source models and datasets, allowing students to work with current tools throughout the studio.

The AI ethics module was delivered as integrated discussions during regular studio sessions without additional class-hours. Key topics—including data privacy and copyright, accountability in human–AI collaboration, stylistic authorship and cultural bias, and model bias—were embedded within design phases to foster critical reflection on AI’s implications in creative and professional practice.

Rather than replacing traditional design skills, AI integration was employed to enhance existing pedagogical objectives while maintaining instructional continuity. In this sense, both modules were progressively embedded within the studio’s “research–design–representation” cycle of each design task (Figure 3). They were first introduced during the design development phase, then extended to support representation tasks, and eventually integrated into pre-design research and site analysis. This staged integration allowed students to shift from isolated tool usage to full-cycle AI engagement, fostering more strategic, reflective, and context-aware applications of AI across the design process.

2.4. Survey Design for Student Feedback

To evaluate the pedagogical implementation of this AI-integrated teaching experiment, a structured questionnaire with 16 questions was developed, comprising three components (Appendix A): demographic information, a quantitative assessment of AI tool effectiveness, and descriptive feedback on students’ course experiences.

The first section collected background data, including participants’ age and year level and whether they had participated in AI-integrated design coursework. This information was used to verify the validity and diversity of responses.

The second section used a Likert scale to assess how effectively AI tools supported four core design competencies: information synthesis, analytical reasoning, creative thinking, and visual representation.

The third section gathered descriptive feedback on students’ actual use of AI tools throughout the studio. It asked when and how students applied AI technologies across the research, design, and representation phases. Additionally, participants ranked the perceived usefulness of different AI tools based on their relevance to each phase. The section concluded with two reflective questions: whether students intended to use AI tools in future design projects and whether they supported the inclusion of AI ethics content in architectural education, along with their preferred format for such discussions.

This multi-dimensional structure allowed for both quantitative and qualitative insights into students’ engagement with AI, providing a foundation for future pedagogical refinement and curriculum development.

3. Studio Implementation and Results

As mentioned above, AI modules were progressively integrated into the studio through a staged approach aligned with the four design tasks. Although each task followed a complete research–design–representation cycle, the outcomes are presented here in the order of design, representation, and research, corresponding to the evolving stages of AI adoption and aligned with students’ growing technical capabilities (Figure 3).

3.1. AI-Integrated Design Development

The design development phase constituted the initial stage of AI integration within the studio, providing a structured context for introducing both foundational and advanced AIGC technologies. Given that most students had no prior experience with such tools, the instructional approach followed a gradual, task-oriented structure that supported creative thinking and iterative design.

In the first design task, students were encouraged to use LLMs to articulate and refine design concepts through natural language interaction, helping to clarify design logic. In parallel, foundational applications of Stable Diffusion enabled students to enhance initial sketches via text-to-image and image-to-image generation. By adjusting prompts and model parameters, students iteratively generated and evaluated visual outputs, improving the efficiency and depth of early-stage design. From the second task onwards, students began training LoRA-based models, building discipline-specific model libraries to meet architectural representation needs (Figure 4). By the third and fourth tasks, they were able to construct customized ComfyUI workflows to test design variations, manage visual styles, and structure repeatable image generation pipelines (Figure 5).

Ethical instruction in this phase emphasized the boundaries of authorship and accountability in human–AI collaboration. A key observed outcome was students’ growing recognition that AI should serve as a design co-agent rather than a shortcut or substitute. Through iterative engagement and structured critique, students developed a more rigorous application of professional judgment through the lens of disciplinary standards, evaluating whether AI-generated texts and diagrams aligned with architectural logic and intent. They also began to question the extent to which AI output should be adopted, learning to intervene, redirect, or reject results as part of an intentional and author-driven process. This approach fostered a deeper sense of agency, positioning AI not as a dominant actor but as a responsive and negotiable component within the design process.

3.2. AI-Integrated Visual Representation

In the visual representation phase, students were engaged in more advanced uses of generative models, confronting several technical challenges: maintaining architectural consistency in output images, ensuring semantic coherence between buildings and their surroundings, and achieving proportionally accurate representations of architectural details, materials, and environmental context. Due to these domain-specific complexities, AI-integrated architecture representation was designated as an advanced instructional component and introduced during the final two design tasks.

To address these challenges, students built ComfyUI workflows incorporating specialized nodes such as ControlNet, Inpaint, Mask-based drawing, and High-Resolution Fix, allowing for precise control over image outputs (Figure 6). In parallel, they selected base models—including SDXL, Flux, and DALL·E 3—based on stylistic preferences and project-specific design needs, further enhancing image fidelity and contextual coherence.

Ethical discussions during the visualization phase centered on stylistic authorship and cultural bias, both of which are particularly pertinent to architectural rendering. A notable outcome was students’ heightened sensitivity to the esthetic assumptions embedded in pretrained AIGC image models. Through comparative image tests using identical prompts across models, students identified risks such as stylistic convergence and regional misrepresentation. These exercises illuminated how algorithms can reinforce stereotypical visual paradigms or distort materiality [33]. In response, students were encouraged to maintain creative control over the generation process by tuning parameters not to conform to AI outputs but to actively shape them in accordance with their design intent. This approach emphasized the use of parameter settings as tools of expression rather than restriction, reinforcing the role of the designer as a critical curator of visual representation in the age of generative tools.

3.3. AI-Integrated Pre-Design Research Stage

As a critical component of architectural education, pre-design research plays a critical role in shaping conceptual clarity, contextual relevance, and methodological rigor. However, in time-constrained design studio settings, students often struggle to conduct thorough data analysis, leading to superficial and predominantly qualitative outcomes. To address this, the AI training for pre-design research focused on improving analysis efficiency and cultivating data literacy. Given the technical complexity of the tools involved, this module was introduced during the final design task, despite its conceptual placement at the beginning of the design process.

An observed outcome of this teaching experiment was the marked improvement in students’ research capacity through the integration of multimodal AI tools. This enhancement was reflected in three key areas: (1) LLMs supported intelligent retrieval and analysis of textual data, enabling the production of more structured and analytically rich design briefs; (2) building on prior parametric design training in Year 2, students were able to further employ AIGC image tools for generative scenario testing and comparative visualization (Figure 7); and (3) pretrained deep learning models—particularly those for semantic segmentation and feature extraction—proved highly effective in quantifying site conditions and informing data-driven design decisions (Figure 8 and Figure 9).

Ethical instruction at this stage focused on model bias, a central concern in AI-assisted design. Students were guided to critically examine how limitations in training data and algorithmic design may lead to systemic distortions. For instance, in semantic segmentation, models may overlook inherent visual biases in the dataset. In text analysis, algorithms may disproportionately reinforce dominant esthetic paradigms embedded in the training corpus. Through comparative exercises juxtaposing AI-generated outputs with manually derived research findings, students developed a dual awareness: recognizing both the operational advantages of intelligent tools and their epistemological constraints. This engagement cultivated a mode of research that is both data-driven and critically reflexive, empowering students to use AI not only to accelerate analysis but also to interrogate the assumptions that underpin architectural knowledge production.

3.4. Survey Results

A total of 70 valid responses were obtained from students who had participated in the 2024–25 AI teaching modules(Table S1: Survey Results). Among them, 38 were male (54.3%), and 32 were female (45.7%), indicating a balanced gender distribution. In terms of participation type, 24.3% of students attended AI skills lectures only, 45.7% were engaged in practical exercises, and 30.0% further applied AI tools in their design studio work. As for perceived benefit, 95.7% found the course at least moderately beneficial, with 35.7% rating it as very beneficial and 24.3% as extremely beneficial. Regarding the actual frequency of AI tool use during the course, 13.7% reported “Never used”, 51.0% “Briefly experimented”, 29.4% “Used multiple times for process discussions”, and 5.9% “Frequently used for multiple tasks” (

Table 2. Demographic and participation summary of valid respondents.

Category	Option	n	Percentage (%)
Gender	Male	38	54.3%
	Female	32	45.7%
Participation in AI-integrated Coursework	Attended lectures only	17	24.3%
	Attended lectures and took exercises	32	45.7%
	Practiced exercises and applied AI tools in design studio	21	30.0%
Perceived Benefit	1: Not beneficial at all	1	1.4%
	2: Slightly beneficial	2	2.9%
	3: Moderately beneficial	25	35.7%
	4: Very beneficial	25	35.7%
	5: Extremely beneficial	17	24.3%
Overall AI Tool Usage in Architectural Design Studios	1: Never used	14	13.7%
	2: Briefly experimented	52	51.0%
	3: Used multiple times for process discussions	30	29.4%
	4: Frequently used for multiple tasks	6	5.9%

).

Students evaluated AI tools across four core capabilities: information synthesis, analytical reasoning, creative thinking, and visual representation. As shown in

Table 3. Descriptive statistics of perceived AI tool capabilities.

Question	Mean (M)	Standard Deviation (SD)	Percentage Agree (4–5)
AI tools demonstrate strong information synthesis capabilities	4.28	0.84	84.1%
AI tools demonstrate strong analytical reasoning capabilities	3.86	0.94	67.1%
AI tools demonstrate strong creative thinking capabilities	3.46	1.02	44.3%
AI tools demonstrate strong visual representation capabilities	3.47	1.03	44.3%

, information synthesis received the highest evaluation, with 84.1% of respondents selecting a score of 4 or 5. Analytical reasoning followed, with 67.1% agreement. Creative thinking and visual representation were rated less positively, with only 44.3% of respondents expressing agreement in both cases. Although the mean scores for these two dimensions remained above the neutral midpoint, the higher standard deviations suggest greater variability in student perceptions, reflecting uncertainty or divergent views on AI’s creative and expressive potential in design contexts.

In terms of actual application, 62.9% of respondents reported using AI tools during the design development stage, followed by 55.7% in early research and 31.4% in final representation. This indicates broader adoption in upstream phases of the design process (

Table 4. AI tool usage across design phases.

Design Stage	n	Percentage (%)
Pre-design Research	39	55.7
Design Development	44	62.9
Final Representation	22	31.4

).

Respondents who reported using AI tools at each phase were asked to rank the usefulness of four tool types: LLMs, multimodal models, AIGC image tools, and custom AI workflows. As shown in

Table 5. Ranking statistics of AI tools across design phases.

Tool Type	Mean Rank (Research)	Top-1 Count (Research)	Mean Rank (Design)	Top-1 Count (Design)	Mean Rank (Representation)	Top-1 Count (Representation)
Large Language Models	1.35	25	2.07	15	2.47	4
Multimodal Models	1.74	14	2.16	14	2.64	3
AIGC Image Tools	3.04	0	2.41	9	1.80	10
Custom AI Workflows	3.59	0	2.82	6	2.14	5

, LLMs were rated the most practical on average in the research phase and received the highest number of first-place selections. This indicates strong consensus among students regarding their utility for information synthesis and early conceptual exploration. In the design phase, all four tools were more evenly rated. While LLMs still led in both mean rank and top-1 count, multimodal models and custom AI workflows also showed growing relevance, suggesting that more complex tools were becoming increasingly useful as design tasks progressed. In the representation phase, AIGC image tools received the highest number of top-1 selections, but their mean ranking reflected greater variability in perception, indicating that while some students highly favored them, others rated them lower. Notably, custom AI workflows outperformed LLMs in this phase, pointing to their potential value in producing refined or tailored visual outputs.

Overall, the results suggest that LLMs were particularly useful in the early stages, whereas multimodal models and custom AI workflows gained importance in later phases. The perception of AIGC image tools appeared more polarized, reflecting individual variation in familiarity or workflow integration (Figure 10).

Regarding their future use of AI tools (

Table 6. Respondents’ attitudes toward AI literacy.

Category	Option	n	Percentage (%)
Future Use of AI Tools	Yes, I am very interested and plan to increase my use	18	25.7%
	Yes, but I believe current tools are not yet mature and I will observe how they evolve	40	57.1%
	Not for now, but I will stay informed	10	14.3%
	No, I do not plan to use them due to high technical barriers	1	1.4%
	No, I do not find AI tools relevant or useful	1	1.4%
AI Ethics Teaching Attitude	Yes, formal teaching and structured classroom discussions are necessary	20	28.6%
	Somewhat necessary, but informal discussion or self-reflection may be sufficient	46	65.7%
	Not necessary	4	5.7%
	No opinion or not sure	0	0

), the majority of respondents (82.8%) expressed a positive attitude. Among them, 25.7% reported strong interest and planned to expand their use in future design work, while 57.1% indicated conditional willingness, acknowledging that current technologies were still immature and would require further observation. An additional 14.3% stated that they had no immediate intention to adopt AI tools but intended to stay informed. Only 2.8% reported no intention to use such tools, either due to technical barriers or a lack of relevance.

In terms of AI ethics education, 28.6% of respondents considered formal instruction and structured classroom discussion necessary, while a larger group (65.7%) favored more flexible approaches such as informal discussion or individual reflection. A small minority (5.7%) deemed such instruction unnecessary, and no respondent reported unfamiliarity with the topic.

4. Discussions

4.1. Modular Integration Enables Scalable and Responsive Curriculum Reform

The 2024–25 AI-integrated teaching experiment in the Architectural Design III/IV studio was deliberately designed to be structurally restrained yet pedagogically impactful.

The structural restraint lies in minimal disruption to the existing studio framework—including teaching sequences, studio group structure, and existing design tasks—while embedding AI modules through a lightweight, modular format. This was achieved by appointing one dedicated technical instructor and introducing no more than 20 additional class-hours of instruction, primarily delivered via e-learning platforms. Given that architectural studios operate with a typical faculty–student ratio of 1:9, requiring every tutor to master and teach the latest AI tools would pose a substantial challenge under limited training and time resources. The designated technical instructor model mitigated discrepancies in faculty expertise, ensured continuity in AI-related instruction, and offered sustained technical support throughout the high-intensity studio cycle. For a high-credit core studio course, this strategy maintained a careful balance between instructional efficiency and curricular innovation, providing a replicable model for scalable AI integration in design education, addressing concerns that AI may undermine rather than support the role of teachers [34].

The pedagogical advancement was driven by the curriculum’s responsiveness to the rapidly evolving AI landscape. Its flexible modular structure enabled continuous adaptation to emerging developments in the AI ecosystem. Between the course’s planning in late 2023 and its completion in mid-2025, a series of major AI tools was released, including Stable Diffusion XL (July 2023), DALL·E 3 (October 2023), Stable Diffusion 3 (February 2024), Flux (August 2024), a major ComfyUI update (December 2024), Grok 3 (February 2025), DeepSeek R1 (January 2025), OpenAI’s o3 and o4 series (January–June 2025), and Grok 4 (July 2025). In response, the AI skills module was continuously revised throughout the academic year, allowing students to engage with state-of-the-art tools in real time. While the dedicated technical instructor ensured the depth, consistency, and currency of AI content delivery, studio instructors remained responsible for facilitating architectural learning and guiding ethical reflection in design. This division of pedagogical roles into technical and disciplinary domains supported a dual-track instructional model that preserved design integrity while embracing technological progress. Collectively, these strategies ensured that the studio remained pedagogically relevant, technologically responsive, and structurally sustainable.

Equally critical to the effectiveness of this modular integration was institutional infrastructure support. University-supported e-learning platforms facilitated differentiated instruction and flexible, asynchronous access to AI learning materials. Cloud-based or inter-institutional computing services mitigated the hardware constraints commonly encountered by students, ensuring reliable model training and inference performance. Moreover, the incorporation of internally developed research outputs such as custom datasets, pretrained deep learning models, and procedural workflows effectively bridged research and pedagogy, enhancing both the conceptual depth and applied relevance of the studio. Student works, including LoRA model packages, ComfyUI workflows, and AI-enhanced datasets, further validated the studio’s capacity to cultivate research-informed, design-driven innovation.

4.2. Student Engagement and the Challenge of Deep AI Integration

While modular integration enhanced studio flexibility, achieving sustained AI engagement across design phases remained a pedagogical challenge. Students’ receptiveness to AI tools varied significantly depending on tool type, project stage, and cognitive demand. The processes of designing prompts or constructing AI workflows required forms of computational thinking that were unfamiliar to many architecture students, particularly those whose training emphasized physical form and materiality, and this often constrained deeper levels of integration.

The patterns of engagement also diverged across the design process. Unlike commercial practice, where AI is predominantly applied in final representation, students in this course were more active during pre-design research and design development. These phases allowed for greater exploratory freedom and aligned with the generative affordances of current tools. In contrast, application during final representation was constrained by low output precision, insufficient integration with production workflows, and challenges balancing tool experimentation with high-fidelity deliverables under tight deadlines.

The perceptions of AI tools were similarly varied. Students widely appreciated the organizational and analytical capacities of LLMs and multimodal tools. However, the creative and expressive potential of AIGC image tools elicited more polarized responses, reflecting differing levels of digital fluency, user confidence, and perceived authorship. Despite these limitations, most students expressed interest in the continued use of AI tools. Nevertheless, many regarded current tools as still immature and approached them with cautious optimism, treating them more as subjects of observation than as instruments of active production.

These challenges and limitations underscored a central pedagogical aim at this stage of AI-integrated education: not to pursue full automation but to cultivate students’ technological awareness, strategic curiosity, and capacity for critical experimentation. The emphasis lies in shifting design cognition from conventional author-driven workflows toward more adaptive, tool-mediated modes of thinking. This incremental shift provides a necessary foundation for future human–AI co-agency in architectural practice.

4.3. Embedding Ethical Thinking Through AI-Integrated Design Practice

Throughout the studio, both students and instructors expressed a shared concern over the risk of esthetic convergence posed by AI tools. The templated logic and recurrent stylistic patterns embedded in many generative models were seen as potentially constraining creative development, normalizing default visual tropes, and flattening architectural imagination. These concerns exemplify a broader tendency in design education to prioritize stylistic replication and technical exploration over human-centered values [35]. Yet this anxiety also resonates with a recurring historical pattern in which each rupture in visual culture, from linear perspective to photography, has challenged and reshaped dominant conceptions of form, perception, and beauty [36]. AI, likewise, is reconfiguring the visual language of design not only through stylistic standardization but also by enabling iterative variation and speculative composition [37] and by supporting perceptual calibration beyond conventional esthetic methods [38].

Although AI ethics were incorporated through in-class discussions, inconsistencies in instructors’ understanding and emphasis occasionally resulted in fragmented or overly subjective guidance, limiting students’ access to coherent and diverse ethical perspectives and reflecting broader pedagogical fragmentation and a lack of systematic governance in AI-integrated education [39,40]. To address this, future iterations of the studio will establish a shared repository of AI ethics teaching materials grounded in current academic research and implement regular faculty workshops to align ethical approaches and instructional strategies across the teaching team. In parallel, establishing a dual mechanism of algorithmic transparency and instructor intervention can help mitigate algorithmic bias and the “echo chamber” effect by ensuring that teachers retain the right to guide at critical stages [41]. These measures aim to enhance the consistency and pedagogical integration of ethical instruction within the AI-augmented studio framework.

The survey findings indicate that while students broadly recognized the importance of AI ethics, most preferred informal discussion or self-reflection over formal instruction. This suggests a preference for embedding ethical reasoning within the design process itself, rather than isolating it in a separate curriculum component. Ultimately, ethics instruction in architecture must go beyond identifying risks to cultivate independent judgment, critical reflexivity, and an accountable relationship with intelligent tools [42,43,44].

5. Conclusions

This teaching experiment, grounded in the 2024–25 implementation of the AI-integrated Architectural Design III/IV studio, validated a modular and pedagogically coherent approach to embedding AI into architectural education. While preserving core disciplinary values, the course enabled students to acquire essential technical skills and develop a critical awareness of human–AI collaboration. The following key conclusions were drawn from the experiment:

• Modular integration is both practical and scalable: Embedding a 20 h AI skills module alongside in-class ethics discussions proved effective without altering the original studio structure, offering a replicable model for AI-integrated design education.
• A dedicated technical instructor is essential: The rapid evolution of AI tools requires real-time updates and hands-on support beyond the capacity of traditional studio instructors. Assigning a technical lead enabled the consistent, up-to-date delivery of AI instruction throughout the course cycle.
• Student engagement is phase-dependent and tool-sensitive: Survey responses indicated higher engagement during the pre-design research and design development phases. Students demonstrated relatively consistent familiarity with large language models, while their perceptions of more complex systems such as image generation tools and custom workflows varied significantly, reflecting differences in technical confidence and cognitive readiness.
• AI integration promotes both cognitive growth and ethical orientation: The pedagogical value of AI lies in strengthening students’ strategic thinking, digital adaptability, and critical design judgment while also encouraging ethical reflection on issues such as authorship, bias, accountability, and the broader implications of human–AI collaboration. Rather than treating AI as a black box generator, students are positioned to engage with it as a co-author requiring intentional guidance and responsibility.
• Supportive infrastructure underpins successful implementation: Robust institutional systems, such as e-learning platforms, cloud computing resources, and access to research-based tools including pretrained models and custom workflows, were essential to enabling meaningful experimentation and sustained technical engagement.

Nonetheless, several limitations warrant further consideration. First, this study reflects only a single academic year and a limited number of undergraduate students, which restricts the generalizability of the findings. The relatively small sample size also limits the statistical representativeness of the survey data. Second, although student feedback indicates strong interest and participation, it may be subject to self-selection bias or novelty effects due to the short implementation period and the emerging nature of AI tools. These factors may affect the stability and objectivity of the observed engagement patterns. Finally, it remains an open pedagogical question whether AI skill modules are more effective when embedded within core design studios or offered as standalone electives. To address these limitations, future implementations will expand the sample base, adopt longitudinal tracking, and incorporate more diversified and structured evaluation frameworks to assess student learning outcomes across technical proficiency, creative capacity, and ethical reasoning.

Fundamentally, the use of AI in design entails selection, optimization, and judgment. Students must interpret and discern among multiple algorithmic outputs—an exercise that deepens critical reflection and active engagement. By framing AI as a collaborator rather than an autonomous agent, the curriculum cultivated a mindset grounded in responsible and informed design authorship. The studio model developed through this experiment proved adaptable across content, structure, and staffing, offering a replicable reference for architectural curriculum reform that supports pedagogical resilience and interdisciplinary advancement in the evolving “AI+” era.