AI-Driven Content Quality Beyond Technological Convenience: A Dual-Track Model of Sustainable Architectural Heritage Engagement

Abstract

Sustainable building research increasingly incorporates AI technologies to enhance efficiency and decision-making, yet little is known about how algorithmic mediation shapes the cultural identity processes essential for heritage sustainability. This study proposes and validates a Content-Driven Dual-Track (CDDT) Model examining the relationships among AI content quality (AIQ), technology acceptance, and Architectural Cultural Identity (ACI). Based on a survey of 631 architecture and design students, structural equation modeling identified three patterns. First, AIQ strongly predicts perceived usefulness, perceived ease of use (PEOU), and perceived enjoyment, supporting a content-driven formation of system evaluations. Second, an “ease-of-use paradox” is observed: PEOU negatively relates to ACI (β = −0.18, p = 0.005), suggesting that frictionless browsing may hinder the cognitive effort required for deeper heritage value internalization. Third, ACI independently predicts continuous engagement intention (β = 0.110, p < 0.05) and correlates strongly with perceived content quality (r = 0.719, p < 0.001). Together, these findings suggest that while operational convenience serves as an essential entry point, sustainable digital heritage engagement requires moving beyond interface usability to prioritize the cultural depth of content assets, a principle applicable to BIM-driven cultural heritage systems and AI-based educational platforms.

1. Introduction

1.1. Background and Research Questions

AI technology has long been a pillar of the “hard technology” domains in architecture, optimizing Building Information Modeling (BIM), energy performance [1,2,3,4], and smart city planning [5,6]. However, a transformative shift is occurring in sociocultural communication, where AI content recommendation mechanisms have redefined the dissemination of architectural knowledge [1,7]. The 2024 reopening of Notre-Dame de Paris exemplifies this; global audiences engaged with the restoration via live streams, acquiring granular knowledge of Gothic ornamentation, stone masonry, and architectural symbolism despite lacking formal training [7,8,9,10]. Similarly, in China, the Palace Museum has used short video platforms to disseminate traditional architectural culture, with individual videos garnering tens of millions of views, while the Dunhuang Research Institute has significantly enhanced public awareness of Dunhuang murals and architectural art through social media [2,7,11,12,13]. These cases highlight a broader trajectory in virtual heritage: digital technologies no longer merely display architecture but actively mediate the emotional and cognitive connection between the public and the built environment [2,7,8,14,15]. Consequently, social media has emerged as a data source for understanding human-centric use patterns of landscape and heritage assets [16].

In this evolving digital landscape, AI recommendation systems function as decision infrastructures that shape how users encounter architectural heritage. These systems operate by analyzing user preferences, behavioral patterns, and social networks to achieve content filtering and personalized delivery [17,18]. On social media platforms, such algorithms do not simply dictate visibility; they fundamentally reconfigure users’ cultural exposure and cognitive frameworks [18,19,20,21]. Rather than passively consuming architectural content, users now continuously encounter and engage with it through browsing, liking, and commenting, forming what can be described as “algorithmic learning paths” [10,20]. Mapping these pathways offers insights into architectural education, cultural dissemination, and design pedagogy [2,7,22,23].

While traditional Technology Acceptance Models (TAMs) focus on system-oriented functionality and usability [18,19,20,21,24,25,26], they struggle to explain the indirect engagement inherent in AI-driven environments. In these “black-box” environments, users no longer evaluate a system through direct operation; instead, they infer utility via the quality of the curated content [17,21]. This transition is particularly significant for architectural heritage engagement. Unlike purely functional software where utility can be objectively assessed, architectural heritage content conveys both technical information and intangible cultural values [26,27]. When algorithms mediate this content, they not only influence what users see but also shape how users understand and identify with architectural culture.

This shift is especially evident in architectural dissemination. Architectural content conveys not only technical information but also historical meaning, regional symbolism, and cultural value [8,28,29]. Architectural heritage scholarship has long emphasized that the built environment functions not merely as a physical or aesthetic object, but as a cultural medium through which memory, belonging, and collective meaning are continuously negotiated [8,14,28,29]. Designers’ continued engagement with architectural cultural content depends not only on system efficiency or accessibility, but also on whether the content resonates with their cultural understanding and value [7,8,30]. However, it remains unclear how content perception and cultural meaning together influence designers’ behavior in AI recommendation environments [2,19,20]. Therefore, this study asks whether AI-driven architectural content can foster cultural identity among designers, and if so, through which psychological mechanisms.

Accordingly, the following research questions are addressed:

Q1: How does design learners’ perception of AI-recommended architectural content influence their technology-related cognitive evaluations in algorithmic environments?

Q2: Does technology-related cognition contribute to the formation of architectural cultural identity under algorithmic recommendation mechanisms?

Q3: Does architectural cultural identity promote learning intention beyond technological factors?

1.2. Literature Review and Theoretical Foundation

1.2.1. Extensions of the Technology Acceptance Model (TAM)

The TAM was proposed by Davis [24], suggesting that Perceived Usefulness (PU) and Perceived Ease of Use (PEOU) jointly determine Behavioral Intention (BI). Later developments introduced Perceived Enjoyment (PE) to capture affective dimensions [25], with the TAM and its variants widely applied in educational technology, social media, and artificial intelligence systems [31].

However, the classic TAM emphasizes a system-oriented logic, in which users form behavioral intentions based primarily on system functionality and operational convenience. This assumption becomes less adequate in AI-driven recommendation environments, where users rarely engage with the underlying system mechanics directly.

Instead, users typically encounter architectural and built-heritage content through platform-based recommendation features (e.g., the “Explore” feed, “Trending” lists, and “You Might Like” suggestions). In the present study, this process is referred to as algorithmic mediation, namely, the exposure to culturally relevant content through algorithmically curated interfaces rather than through direct interaction with the recommendation mechanism itself.

At the same time, recommendation systems often operate in a black-box environment, in which their internal logic remains opaque to ordinary users. Under such conditions, users are less likely to evaluate the system through knowledge of algorithmic procedures and more likely to infer its value indirectly from the quality of the content they receive. Consequently, system evaluation shifts from direct assessment of technical mechanics to content-mediated experience [17,18,19,21]. These concepts serve as theoretical context rather than measured constructs in this study model.

In the present study, both algorithmic mediation and the black-box environment are treated not as a directly measured construct, but as the interpretive context under which users evaluate recommendation systems indirectly through content cues rather than through knowledge of algorithmic procedures. Prior research further suggests that algorithmic awareness and media literacy shape how users interpret recommended information [32,33,34,35]. Therefore, in AI-mediated cultural learning contexts, users’ judgments are formed not only through usability-related cognition but also through their perceptions of the originality, credibility, and personalization of recommended content [17,26].

This shift has catalyzed various “content-centric” expansions of the TAM framework. For instance, research by Tran et al. [36] demonstrates that in AI-driven music curation, perceived originality and emotional resonance are stronger predictors of continued usage than mere technical functionality. Similarly, within short-video learning environments, Wang et al. [31] found that content credibility exerts a more profound influence on learning motivation than operational ease. These insights underscore that, in curated recommendation environments, content perception functions as a critical antecedent to users’ subsequent technology-related evaluations [17,37]. This provides the theoretical basis for extending the TAM beyond traditional usability metrics in the present study.

1.2.2. Conceptualization of AI Content Quality (AIQ)

The Algorithmic Media Content Awareness (AMCA) scale developed by Zarouali et al. [32] provides a systematic framework for understanding how people perceive algorithmically recommended content. AMCA includes four dimensions: Content Filtering (FIL), Automated Decision-Making (ADM), Human–Algorithm Interplay (HAI), and Ethical Considerations (ETH) [32]. This model shows how algorithmic cognition influences both trust levels and evaluations of content authenticity and diversity [20,32]. Related research has further demonstrated that personality-tailored algorithmic recommendations can significantly shape user perceptions and behavioral responses [38].

However, AMCA emphasizes users’ awareness of algorithmic processes rather than their evaluative experience of content attributes. Based on this, this paper adapts AMCA and proposes AIQ, focusing specifically on evaluative content experience relevant to learning and meaning construction [10], comprising three dimensions:

• Perceived Originality (PO): perceptions of novelty and creativity in recommended architectural content [39,40];
• Perceived Credibility (PC): trust in the sources, expertise, and authenticity of architectural information [41,42];
• Perceived Personalization (PP): perceived alignment between content and individual interests, learning needs, or cultural backgrounds [43,44,45].

In sum, AIQ represents a shift toward content-level perception in digital learning [26]. In algorithmic ecosystems, design learners prioritize curated content over system mechanics [17,20]. Empirical evidence suggests that personalization enhances engagement [17], while originality and credibility are foundational for building trust [35]. In the present study, AIQ is not proposed as a substitute for TAM, but as a content-level antecedent that shapes users’ subsequent technology-related evaluations in recommendation-driven environments. Thus, this study conceptualizes AIQ as a critical antecedent within the TAM framework, capturing the cognitive foundation required for technology acceptance in recommendation-driven environments.

1.2.3. Conceptual Introduction of Architectural Cultural Identity (ACI)

ACI refers to the emotional belonging and value identification that design learners develop toward architectural values, regional characteristics, and historical symbolism through their engagement with architectural culture [46]. This perspective is consistent with architectural heritage scholarship, which views heritage not merely as preserved built form but as a cultural medium through which collective memory, identity, and shared meaning are constructed and negotiated over time [8,28,29]. In this sense, architecture is not only a spatial or formal object, but also a material expression of cultural memory and social meaning [28,29]. In digital heritage contexts, such cultural engagement is further shaped by users’ sense of presence and immersion [47].

Cultural identity theory identifies two psychological stages in cultural learning: Exploration and Commitment [48,49,50]. Adapting this framework to architectural contexts, our research defines ACI as a two-dimensional construct:

• Exploration (ACI_E): Tendencies to actively search for, follow, like, share, and actively learn about architectural cultural content;
• Commitment (ACI_C): Emotional identification with architectural cultural values and willingness to preserve them.

In the present study, the exploration dimension of ACI is interpreted not as generic social media activity but as identity-oriented exploratory behavior directed toward architectural cultural meaning, value, and belonging. Although some ACI_E items involve actions such as searching, following, liking, or sharing, these behaviors are treated here as culturally oriented exploration only when they reflect an effort to engage with architectural cultural significance rather than mere platform activity.

In AI recommendation environments, ACI functions as a psychological outcome of architectural learning. It captures how designers develop value identification through two stages: Exploration (active engagement with content) and Commitment (emotional belonging to cultural values) [7,10,13]. By modeling ACI alongside TAM, this study examines how algorithmically delivered content facilitates or potentially hinders, the deep cognitive commitment necessary for heritage preservation.

1.3. Research Framework and Hypotheses

1.3.1. Theoretical Integration Framework

In this study, the term “dual-track” refers to two theoretically distinct cognitive processes rather than merely two parallel predictors of behavioral intention. The first is a technology-evaluation process, in which users infer the usefulness, ease of use, and enjoyment of AI-recommended environments through their perceptions of recommended content. The second is a culture-identity process, in which engagement with architectural cultural content involves exploration, value internalization, and identity-based motivation.

On this basis, the present study integrates AIQ and ACI into the TAM, while retaining the TAM as a baseline cognitive structure rather than the sole explanatory mechanism. In this framework, AIQ functions as a content-level antecedent that shapes technology-related evaluations, whereas ACI represents a value-oriented psychological pathway linked to cultural engagement. Together, these constructs form a content–technology–culture framework (Figure 1).

The framework proposes four core pathways:

1.

Path 1: AIQ → TAM → BI (Content-Technology Path)

Users first perceive the originality, credibility, and personalization of recommended content, which shapes their technology cognition of AI systems (PU, PEOU, PE), ultimately influencing continuous learning intention [24,25].

2.

Path 2: TAM → ACI (Technology-Culture Path, Exploratory)

Whether technology cognition promotes cultural identity remains an open question. While positive experiences may enhance learning engagement, excessive convenience may lead to shallow learning and diminish cultural understanding [7,10].

3.

Path 3: ACI → BI (Identity-Driven Path)

Cultural identity serves as a key psychological mechanism driving continuous learning [48,49,50]. This framework transcends traditional system-centered TAM paradigms by exploring the complex relationships among content, technology, and culture in algorithm-mediated cultural learning contexts.

4.

Path 4: AIQ → ACI (Content-Culture Path, Exploratory)

High-quality architectural content may directly evoke emotional resonance without mediation by technology cognition [17,28].

Accordingly, the proposed framework conceptualizes behavioral intention as shaped by both instrumental cognition and identity-based meaning construction, and the following hypotheses are developed to test the relationships among these constructs empirically.

1.3.2. Research Hypotheses

We propose the following hypotheses. While the classical TAM pathway [24] is retained as a control structure, this study specifically investigates the distinct mechanisms of content-driven acceptance and cultural identity.

1.

H1a–H1c: AI Content Quality and Technology Cognition

• H1a: AIQ positively influences PU.
• H1b: AIQ positively influences PEOU.
• H1c: AIQ positively influences PE.

2.

H2a–H2c: Technology Cognition and Architectural Cultural Identity (Exploratory)

In an algorithmic recommendation environment, the direction of influence between technology cognition and cultural identity is uncertain. A positive technological experience may enhance learning engagement and facilitate deeper cultural understanding, whereas excessive convenience could encourage superficial browsing and weaken cultural internalization [7,10].

Therefore, this paper adopts an exploratory approach for H2a–H2c without presupposing the direction of influence:

• H2a: PU has an effect on ACI.
• H2b: PEOU has an effect on ACI.
• H2c: PE has an effect on ACI.

3.

H3: Cultural Identity and Continuous Learning Intention

• H3: ACI positively influences BI [48,49,50].

4.

H4: AI Content Quality and Architectural Cultural Identity (Exploratory)

• H4: AIQ is positively associated with ACI [20,28].

Note: Given the conceptual proximity between AIQ (Perceived Content Quality) and ACI (cultural identity), this hypothesis will be tested through both path analysis and correlation analysis. The final analytical approach will be determined based on model diagnostics, particularly multicollinearity assessment.

1.4. Research Objectives

This study aims to examine how AI recommendation mechanisms shape architectural culture learning by integrating AIQ, TAM, and ACI into a three-dimensional “content-technology-culture” framework and employing structural equation modeling to validate the proposed pathways. Specifically, this research seeks to (1) validate whether perceived content quality serves as an antecedent to technology cognition in algorithmic environments, (2) explore the relationship between technology cognition and cultural identity formation, and (3) examine whether heritage value internalization contributes to learning behavior beyond technological factors.

This study extends AI research in sustainable building by shifting attention from technical optimization to cognitive–cultural sustainability in digital architectural systems.

2. Methods

2.1. Research Design and Procedures

A questionnaire survey was employed and structural equation modeling (SEM) [37] was used to examine the relationships among users’ perceptions of architectural cultural content, technology cognition, cultural identity, and behavioral intentions within AI recommendation contexts on social media.

The research was conducted in six stages:

• Reviewing relevant literature and developing the conceptual model and hypotheses;
• Adapting measurement scales with expert review and pilot testing;
• Administering the questionnaire and collecting data;
• Data screening and sample refinement;
• Assessing reliability and validity through EFA and CFA;
• Testing and refining the structural model using SEM.

2.2. Participants and Sampling

Participants were architectural design students recruited from architecture, design, and digital media programs. The questionnaire was distributed using purposive sampling through academic design networks to ensure respondents possessed the requisite disciplinary background. The initial sample included participants from Taiwan and Hong Kong. However, due to regional differences in architectural sustainable heritage management, social media usage patterns, and cultural backgrounds, the final analysis retained only the Taiwan sample (N = 631) to ensure sample homogeneity [37]. All participants provided informed consent, and all responses were collected anonymously. The research protocol complied with ethical review standards of the Ministry of Education and the institution.

The sample consisted of 631 respondents (58.0% female, 42.0% male). Most were undergraduates (78.9%), while 21.1% were enrolled in graduate programs. Regarding academic background, 44.4% majored in architecture, 27.1% in digital media, and 28.5% in other design-related fields. The majority (87.3%) were aged 18–24 years, with 12.7% aged 25 years or older. This age distribution aligns with Generation Z, whose architectural learning behaviors in digital environments have been shown to differ significantly from previous generations [51].

Regarding daily exposure to AI-recommended architectural or cultural content, 37.7% reported 1–3 h per day, 33.9% reported 4–6 h, 15.1% reported less than 1 h, and 13.3% reported more than 6 h. These results indicate a generally high level of engagement with AI recommendation systems within this emerging professional group in Taiwan.

2.3. Research Instruments and Measurement Scales

The questionnaire was developed through a systematic process involving literature review and item adaptation. AI recommendations specifically refer to algorithmic discovery features on social media platforms, such as the ‘Explore’ feed, ‘Trending’ lists, and ‘You Might Like’ suggestions, which actively push personalized content to users. In this study, all measurement items were explicitly contextualized to architectural and built heritage related cultural content to ensure domain specific validity. Its content validity and appropriateness for architectural culture learning contexts were established via reviews by three experts in design studies, cultural studies, and AI algorithms. The final questionnaire consisted of five sections with 31 items, each measured using a 5-point Likert scale (1 = strongly disagree; 5 = strongly agree), as shown in

Table 1. Scale Structure and Reliability.

Section	Construct	Items	Cronbach’s α	Source
Part 1	AI Content Quality (AIQ)	9		Adapted from AMCA [32,38,39,40,41,42,44,45]
	- Perceived Originality (PO)	3	0.787
	- Perceived Credibility (PC)	3	0.828
	- Perceived Personalization (PP)	3	0.761
Part 2	Technology Acceptance Model (TAM)	8		Adapted from the TAM [24,25]
	- Perceived Usefulness (PU)	3	0.812
	- Perceived Ease of Use (PEOU)	3	0.782
	- Perceived Enjoyment (PE)	2	0.830
Part 3	Behavioral Intention (BI)	2	0.926	Extended TAM adapted [24,25]
Part 4	Architectural Cultural Identity (ACI)	6		Adapted from MEIM-R and MENI [46,48,50,52]
	- Exploration (ACI_E)	3	0.769
	- Commitment (ACI_C)	3	0.805
Part 5	Demographics	6	—	Self-developed

.

Accordingly, in this study, algorithmic mediation is operationalized as respondents’ exposure to architectural cultural content through these recommendation features, while the black-box character of the recommendation environment is treated as the broader interpretive condition under which users evaluate systems through content cues rather than direct knowledge of algorithmic procedures.

The AIQ scale was adapted from the AMCA-scale developed by Zarouali et al. [32], with conceptual refinement to capture users’ content-level perceptions rather than direct algorithmic awareness [20,21]. Specifically, PO items were derived from the Content Filtering dimension (AMCA-FIL), PC items from the Ethical Considerations dimension (AMCA-ETH), and PP items from the Automated Decision-Making (AMCA-ADM) and Human–Algorithm Interplay (AMCA-HAI) dimensions. Expert reviews confirmed the questionnaire’s relevance and suitability for this research context.

TAM items were adapted from Davis [24,25] with semantic adjustments to reflect “AI-recommended browsing of architectural content”. Behavioral intention (BI) was measured using two focused items adapted from extended TAM research [24,25] to capture respondents’ intention to continue engaging with AI-recommended architectural cultural content. Although this represents a parsimonious operationalization, its adequacy was supported by the reliability and validity results reported later in the manuscript. ACI items were adapted from Phinney [48,50,52] and Maehler, Zabal et al. [46] with semantic adaptations to architectural culture contexts. A complete list of all measurement items is provided in Appendix A.

2.4. Reliability and Validity Assessment

2.4.1. Normality Testing and Estimation Method

Descriptive statistics of the 25 observed items indicated skewness and kurtosis values outside the ±1 range typically associated with normal distribution. However, given the large sample size (N = 631), the central limit theorem implies that the sampling distribution of the parameter estimates approximates normality [37]. Nevertheless, to ensure the most robust estimation, we employed the Weighted Least Squares Mean and Variance Adjusted (WLSMV) estimator, which is specifically designed to accommodate non-normal distributions and the ordinal nature of Likert-scale data [37,53].

2.4.2. KMO and Bartlett’s Tests

The Kaiser–Meyer–Olin (KMO) measure of sampling adequacy was 0.932, and individual MSA values for all items were above 0.80, confirming that each variable met the minimum criterion for factor analysis. Bartlett’s test of sphericity yielded χ² (300) = 9024.405, p < 0.001, indicating that the data were suitable for factor analysis [37].

2.4.3. Reliability Analysis

All constructs demonstrated satisfactory internal consistency, with Cronbach’s α values ranging from 0.761 to 0.926 (as shown in Table 1). The overall reliability of the questionnaire was α = 0.914, indicating high measurement stability and internal coherence. According to Nunnally [54], values above 0.70 indicate acceptable reliability, and values above 0.80 indicate strong internal consistency. Therefore, all constructs in this paper met or exceeded the recommended threshold.

2.4.4. Validity Analysis

EFA was conducted using principal axis factoring (PAF) with oblimin rotation [37]. The results indicated that all item factor loadings exceeded 0.70 with no significant cross-loadings.

A confirmatory factor analysis (CFA) was then performed to evaluate the measurement model. CFA with WLSMV estimation showed good model fit: χ² (255) = 1156.99, p < 0.001, χ²/df = 4.54, RMSEA = 0.075 (90% CI [0.071, 0.079]), CFI = 0.960, TLI = 0.953, SRMR = 0.043. These fit indices indicate an acceptable fit of the measurement model [53], supporting the convergent validity of the latent constructs. All standardized factor loadings ranged from 0.71 to 0.92 and were statistically significant (p < 0.001), meeting the convergent validity criterion (λ > 0.6) [37].

2.4.5. Common Method Bias Assessment

Since all data were collected via a single self-report instrument, common method bias (CMB) was assessed using Harman’s single-factor test [55]. An unrotated exploratory factor analysis (EFA) including all 25 observed items showed that the first factor accounted for 41.81% of the total variance, below the 50% threshold typically used to indicate serious CMB concern. In addition, five factors with eigenvalues greater than 1.0 were extracted, cumulatively explaining 64.24% of the variance. These results suggest that common method bias is unlikely to be a serious concern in the present study.

2.5. Data Analysis Methods

Data analysis was conducted in four stages:

• Descriptive statistics: Means, standard deviations, skewness, and kurtosis were computed using IBM SPSS Statistics 29 (IBM Corp., Armonk, NY, USA).
• Reliability and validity testing: Cronbach’s α, composite reliability (CR), and average variance extracted (AVE) were assessed to evaluate measurement quality. Discriminant validity was further examined using the heterotrait–monotrait (HTMT) ratio of correlations.
• Confirmatory factor analysis: Measurement models were tested in Mplus 8.3 (Muthén & Muthén, Los Angeles, CA, USA) using WLSMV estimation to examine the structure of observed items for latent constructs and assess model fit.
• Structural equation modeling: The final retained structural model was specified to examine the effects of AIQ on TAM-related constructs, the association between AIQ and ACI, and the effects of the TAM and ACI on BI. Model fit was evaluated using multiple indices (χ²/df, RMSEA, CFI, TLI, SRMR) [37]. Path significance was assessed at p < 0.05. Variance inflation factor (VIF) values were computed to assess potential multicollinearity among structural predictors. In addition, indirect effects were examined using ML-based bias-corrected bootstrap resampling with 5000 iterations as a supplementary robustness check, because bootstrap estimation is not available under WLSMV in Mplus 8.3 [56].

2.6. Model Adjustments

The final retained model was derived from the initial conceptual specification through theory-guided refinement. Although the initial model included directional paths from AIQ to ACI and from TAM-related constructs to ACI, preliminary testing did not support retaining these paths as stable directional relationships in the final model. Given the cross-sectional design and the strong but non-directional empirical association between AIQ and ACI, the relationship was re-specified as a correlation. This refinement was intended to preserve theoretical relevance while aligning the final model more closely with the empirical evidence.

1.

Removal of the AIQ → ACI Direct Path

Although the composite correlation between AIQ and ACI was moderate (r = 0.576, p < 0.001;

Table 2. Means, Standard Deviations, and Correlations.

Construct	Mean	SD	AIQ	PU	PEOU	PE	ACI	BI
AIQ	3.53	0.63	1
PU	3.63	0.81	0.772 ***	1
PEOU	3.69	0.8	0.695 ***	0.639 ***	1
PE	3.78	0.85	0.591 ***	0.610 ***	0.536 ***	1
ACI	3.59	0.73	0.576 ***	0.536 ***	0.466 ***	0.400 ***	1
BI	3.43	0.9	0.615 ***	0.629 ***	0.497 ***	0.538 ***	0.476 ***	1

Note: N = 631. *** p < 0.001. Correlations computed using composite scores in SPSS 29. Latent variable correlation between AIQ and ACI from SEM analysis was 0.719 *** in Mplus 8.3.

), the SEM results revealed a high latent-level association between the two constructs (r = 0.719, p < 0.001), which raised identification and interpretability concerns when specifying a direct causal path. Conceptually, AIQ captures users’ evaluative perceptions of recommended content, whereas ACI reflects value-oriented cultural identification. Given the absence of clear theoretical directionality and the interpretive difficulty of specifying this relationship as a causal path, the direct path from AIQ to ACI was removed. Instead, their relationship was modeled as a correlation and examined under H4.

2.

Treatment of the TAM → ACI Paths

Hypotheses H2a–H2c were exploratory, intended to test whether technology cognition contributes to the development of cultural identity. However, the initial model showed that none of the three paths (PU → ACI, PEOU → ACI, PE → ACI) provided theoretically interpretable support for technology-driven cultural identification, and one showed a significant negative coefficient. The overall model fit was also unsatisfactory. Theoretically, this divergence underscores the fundamental distinction between TAM—which measures instrumental evaluations of functional systems—and ACI, which reflects affective identification with cultural values [19,48,49]. These findings suggest that in algorithm-mediated environments, perceived content quality (AIQ) likely acts as a primary exogenous driver that independently shapes both technology acceptance and cultural identity, rather than following the traditional serial mediation via TAM.

3.

Final Model Revision

Based on these theoretical and statistical considerations, the final structural model was revised as follows:

Paths retained:

• AIQ → TAM
• TAM → BI
• ACI → BI

Paths removed:

• AIQ → ACI (replaced with correlational analysis)
• TAM → ACI

Additional analysis:

• Correlation between AIQ and ACI (H4)

After these revisions, model fit improved substantially: Compared to the initial structural model, the revised framework demonstrated a significantly better fit, as evidenced by the marked optimization of indices. This improvement indicates a closer alignment between the conceptual model and the observed data. The specific statistical parameters and the definitive comparison between the initial structural model and the final model are detailed in Section 3.3.

2.7. Methods Summary

This chapter synthesized the research design, measurement validation, and analytical framework. By employing WLSMV robust estimation and rigorous reliability and validity assessments, this study ensured the statistical integrity of the measurement instruments. Furthermore, the strategic model refinements detailed in Section 2.6 established a parsimonious and theoretically grounded structure, providing a robust empirical foundation for the hypothesis testing and path analysis presented in Section 3.

3. Results

3.1. Descriptive Statistics and Correlations

Descriptive statistics revealed that construct means ranged from 3.43 to 3.78 (SDs = 0.63–0.90), indicating moderately high levels overall. Pearson correlations among composite scores revealed significant positive associations among all constructs (p < 0.001), as shown in Table 2. AIQ correlated 0.591–0.772 with the three TAM dimensions and 0.576 with ACI, medium to strong effects by Cohen’s [57] standards. These patterns provide preliminary support for the hypothesized structural model. Latent variable correlations from structural equation modeling are reported in Section 3.3.

3.2. Measurement Model

We tested the measurement model via confirmatory factor analysis in Mplus 8.3 using WLSMV estimation. The model comprised nine first-order constructs (PO, PC, PP, PU, PEOU, PE, ACI_E, ACI_C, BI) and two second-order factors (AIQ, ACI), measured by 25 observed items.

Fit indices were acceptable: χ² (255) = 1156.99, p < 0.001; χ²/df = 4.54; RMSEA = 0.075 (90% CI [0.071, 0.079]); CFI = 0.960; TLI = 0.953; SRMR = 0.043. Although the chi-square test was significant—typical with large samples—incremental fit indices (CFI, TLI) and error metrics (RMSEA, SRMR) met standard cutoffs [37,53]), confirming good model fit.

All standardized factor loadings (λ) ranged from 0.675 to 0.962 (p < 0.001). Composite reliability (CR) values spanned 0.847–0.945, and average variance extracted (AVE) ranged from 0.649 to 0.895, satisfying convergent validity criteria. For discriminant validity, each construct’s √AVE exceeded its correlations with other constructs, supporting scale distinctiveness [58] (see

Table 3. Confirmatory Factor Analysis Results Summary.

Construct Type	Construct	Indicators	Factor Loadings	CR	AVE	Cronbach’s α
First-Order Constructs	PO	PO1	0.812	0.869	0.689	0.787
		PO2	0.852
		PO3	0.791
	PC	PC1	0.835	0.897	0.745	0.828
		PC2	0.881
		PC3	0.879
	PP	PP1	0.675	0.847	0.652	0.761
		PP2	0.753
		PP3	0.891
	PU	PU1	0.82	0.895	0.742	0.812
		PU2	0.806
		PU3	0.814
	PEOU	PEOU1	0.843	0.875	0.701	0.782
		PEOU2	0.776
		PEOU3	0.737
	PE	PE1	0.85	0.901	0.822	0.83
		PE2	0.917
	ACI_E	ACI_E1	0.751	0.847	0.649	0.769
		ACI_E2	0.813
		ACI_E3	0.728
	ACI_C	ACI_C1	0.794	0.886	0.721	0.805
		ACI_C2	0.865
		ACI_C3	0.784
	BI	BI1	0.955	0.945	0.895	0.926
		BI2	0.937
Second-Order Constructs	AIQ	PO	0.789	0.889	0.668	0.854
		PC	0.851
		PP	0.821
	ACI	ACI_E	0.878	0.916	0.785	0.872
		ACI_C	0.962

Note: CR = Composite Reliability; AVE = Average Variance Extracted. All factor loadings are significant at p < 0.001. AIQ is a second-order factor comprising PO, PC, and PP. ACI is a second-order factor comprising ACI_E and ACI_C.

).

To further assess discriminant validity, HTMT ratios were calculated for both first-order and higher-order construct pairs [59], as shown in

Table 4. Heterotrait–Monotrait Ratios for the Study Constructs.

Panel A. First-order construct HTMT matrix
Construct	PO	PC	PP	PU	PEOU	PE	BI	ACI_E	ACI_C
PO	—
PC	0.689	—
PP	0.585	0.604	—
PU	0.781	0.798	0.771	—
PEOU	0.676	0.694	0.824	0.804	—
PE	0.577	0.632	0.577	0.743	0.669	—
BI	0.515	0.663	0.572	0.725	0.584	0.615	—
ACI_E	0.441	0.544	0.522	0.602	0.531	0.434	0.488	—
ACI_C	0.525	0.662	0.561	0.626	0.56	0.476	0.531	0.856	—
Note: Most HTMT values were below the conservative threshold of 0.85. The HTMT value between ACI_E and ACI_C was 0.856, slightly exceeding the strict 0.85 criterion but remaining below the more liberal 0.90 threshold.
Panel B. Second-order and key construct pairs
Construct 1			Construct 2			HTMT
AIQ			PU			0.924
AIQ			PEOU			0.859
AIQ			PE			0.702
AIQ			ACI			0.671
PU			PEOU			0.804
PU			PE			0.743
PU			BI			0.725
PEOU			BI			0.584
PE			BI			0.615
ACI			BI			0.533
ACI			PU			0.642
ACI			PEOU			0.571
Note: the HTMT value between AIQ and PEOU was 0.859, and that between AIQ and PU was 0.924. These results indicate conceptual proximity among certain construct pairs and are therefore interpreted together with CFA evidence and VIF diagnostics in the text.

. Most HTMT values were below the conservative threshold of 0.85. Three pairs slightly exceeded the strict criterion: ACI_E–ACI_C (0.856), AIQ–PEOU (0.859), and AIQ–PU (0.924). Given the conceptual proximity among these constructs, these values were interpreted together with CFA and VIF results (see

Table 5. Variance Inflation Factor Values for Structural Predictors.

Dependent Variable	Predictor	VIF	Tolerance
BI	PU	2.236	0.447
BI	PEOU	1.849	0.541
BI	PE	1.697	0.589
BI	ACI	1.461	0.684
PU	AIQ	2.223	0.45
PU	PEOU	2.032	0.492
PU	PE	1.611	0.621
PEOU	AIQ	2.62	0.382
PEOU	PU	2.716	0.368
PEOU	PE	1.688	0.592

Note: All VIF values were below the recommended threshold of 3.0, indicating that multicollinearity was not a serious concern in the structural model.

) rather than as evidence of construct redundancy.

3.3. Structural Model and Hypothesis Testing

3.3.1. Model Adjustment and Fit

The structural equation model was established following the measurement model validation. As detailed in Section 2.6, the initial model underwent refinement by excluding the non-significant and theoretically inconsistent TAM → ACI paths and converting the AIQ → ACI path into a correlation to resolve collinearity issues (see

Table 6. Structural Equation Modeling Path Coefficients and Hypothesis Testing.

Hypothesis	Path Relationship	Std. β	S.E.	t-Value	p-Value	Support
H1a	AIQ → PU	0.964	0.011	84.729	<0.001	Supported
H1b	AIQ → PEOU	0.878	0.015	59.103	<0.001	Supported
H1c	AIQ → PE	0.765	0.018	43.017	<0.001	Supported
H2a	PU → ACI	−0.017	0.015	1.127	0.260	Not Supported
H2b	PEOU → ACI	−0.18	0.064	2.807	0.005	⚠ Not Supported (Negative)
H2c	PE → ACI	−0.004	0.029	0.138	0.890	Not Supported
H3	ACI → BI	0.110	0.045	2.281	<0.05	Supported
H4 *	AIQ ↔ ACI	0.719	0.023	30.623	<0.001	Correlation

Note: H1a–H1c, H3, and H4 are reported from the final structural model. H2a–H2c were tested in the preliminary model; these paths were subsequently removed due to non-significance or negative direction (see Section 2.6 for model adjustment rationale). * H4 was tested using correlation analysis due to multicollinearity concerns.

).

This adjustment led to a significant improvement in fit indices: the final model achieved χ²/df = 2.34, RMSEA = 0.073, CFI = 0.962, TLI = 0.956, SRMR = 0.043. All fit indices met or exceeded the recommended standards [37,53], indicating that the adjusted model adequately explains the sample data. The transition from the initial to the final model resulted in a significant improvement in fit, with RMSEA declining from 0.086 and CFI increasing from 0.943.

To assess potential multicollinearity among structural predictors, VIF values were calculated for the retained structural paths. As shown in Table 5, all VIF values ranged from 1.461 to 2.716, well below the recommended threshold of 3.0 [37], indicating that multicollinearity was not a serious concern in the structural model. These results further suggest that, although AIQ and PU were strongly associated, their relationship did not reflect problematic statistical redundancy.

3.3.2. Hypothesis Testing Results

Hypothesis testing results are summarized in Figure 2 and Table 6.

Supported Paths: AIQ exhibited powerful positive effects on PU (β = 0.964), PEOU (β = 0.878), and PE (β = 0.765), supporting H1a, H1b, and H1c (p < 0.001). ACI significantly influenced BI (β = 0.110, p < 0.05), supporting H3.

Exploratory Paths: The H2 series was not supported. Specifically, while PU and PE showed no significant associations with ACI, PEOU exhibited a significant negative association (β = −0.18, p = 0.005), as shown in Table 6. Due to model fit issues and theoretical considerations, these paths were removed from the final model.

TAM Controls: PU remains the strongest predictor of intention (β = 0.667), while PEOU had no significant direct effect on BI (β = −0.156, p > 0.05).

The final structural model includes the following paths:

• AIQ → PU/PEOU/PE (content-driven technology acceptance);
• PU/PEOU/PE → BI (technology acceptance influencing behavioral intention);
• ACI → BI (cultural identity driving continuous learning);
• AIQ ↔ ACI (correlation relationship).

Based on the TAM paths:

• PU → BI: β = 0.667, p < 0.001;
• PEOU → BI: β = −0.156, p > 0.05;
• PE → BI: β = 0.208, p < 0.05.

The fit indices of the measurement model, initial model, and final model are compared in

Table 7. Model Fit Comparison.

Model	χ2/df	RMSEA	CFI	TLI	Notes
Measurement Model (CFA)	4.54	0.075	0.96	0.953	Reported in Section 2.4
Initial Model	3.89	0.086	0.943	0.954	TAM→ACI included
Final Model	2.34	0.073	0.962	0.956	TAM→ACI removed

.

3.3.3. Explained Variance of Endogenous Variables

The structural model’s explained variance (R²) for each endogenous variable is as follows:

1.

Technology acceptance dimensions:

• PU: R² = 0.929, with AIQ explaining 92.9% of its variance;
• PEOU: R² = 0.771, with AIQ explaining 77.1% of its variance;
• PE: R² = 0.585, with AIQ explaining 58.5% of its variance.

2.

Behavioral intention:

• BI: R² = 0.613, with the TAM (PU, PEOU, PE) and ACI jointly explaining 61.3% of its variance.

According to conventional benchmarks for structural models, these R² values indicate substantial explanatory power [57]. Note that ACI is specified as an exogenous variable in the final structural model. Because its relationship with AIQ is modeled as a correlation rather than a directional path, no R² value is reported for ACI. At the measurement level, the second-order ACI factor demonstrated strong loadings on its first-order dimensions (ACI_E: λ = 0.878; ACI_C: λ = 0.962), indicating robust construct coherence.

As a supplementary robustness check, an ML-based bias-corrected bootstrap analysis with 5000 resamples was conducted to examine indirect effects. The results indicated a significant total indirect effect of AIQ on BI (β = 0.375, 95% BC CI [0.303, 0.442]), which was consistent with the broader pattern observed in the WLSMV model. Because the final retained model specified the relationship between AIQ and ACI as a correlation rather than a directional path, and removed the TAM → ACI paths, the bootstrap decomposition of specific indirect effects is not interpreted as a main result.

3.3.4. Results Summary

This chapter presented the empirical results of hypothesis testing through SEM analysis. AIQ demonstrated strong effects on all TAM dimensions (β = 0.765–0.964). ACI showed a significant effect on BI (β = 0.110, p < 0.05). The TAM→ACI paths were not supported and were removed following preliminary testing. The final model demonstrated good fit (χ²/df = 2.34, RMSEA = 0.073, CFI = 0.962).

4. Discussion

4.1. Findings Consistent with Expectations

We validated the proposed three-dimensional framework integrating content, technology, and culture using SEM. The final model demonstrated good fit, supporting the proposed “content-technology-culture” framework.

4.1.1. Content-Driven Technology Cognition

As shown in Table 6, AIQ had significant positive effects on all three technology cognition dimensions. These findings corroborate the “content-driven” pathway, where design learners’ evaluations are primarily structured by their experience with recommended content rather than direct system manipulation. In such “black-box” algorithmic contexts, content serves as a “psychological interface for technology,” functioning as the primary tangible interface for performance evaluation [27].

The exceptionally high explanatory power of AIQ for PU (R² = 0.929) merits careful interpretation. This strong association does not indicate construct redundancy; rather, it reflects a fundamental shift in how system evaluations are formed in algorithm-driven environments. In contemporary recommendation systems, designers rarely interact with underlying technical mechanisms directly; instead, their evaluation of system performance is largely inferred from the quality of the content delivered. In this sense, recommended content functions as a perceptible interface through which technological utility is cognitively constructed.

Thus, the high explanatory power suggests that perceived usefulness is formed mainly through content evaluation, rather than through direct experience of system features. This result, however, should be interpreted cautiously. The HTMT ratio between AIQ and PU reached 0.924, exceeding both the strict 0.85 criterion and the more liberal 0.90 threshold, indicating substantial conceptual proximity between the two constructs. Nevertheless, several considerations suggest that this pattern does not simply reflect construct redundancy. First, the VIF value for PU as a structural predictor was 2.223, well below the recommended threshold of 3.0, indicating that multicollinearity was not a serious concern. Second, the CFA results supported a measurement structure in which AIQ and PU loaded on theoretically distinct latent constructs rather than collapsing into a single factor. Third, the two constructs occupy different positions in the nomological network: AIQ captures evaluative judgments about the originality, credibility, and personalization of recommended content, whereas PU reflects a system-level judgment of whether the recommendation environment is useful for architectural cultural learning. In AI-mediated recommendation contexts, where users rarely access the underlying technical logic directly, it is theoretically plausible that perceived usefulness is inferred largely from content quality. The strong AIQ–PU association is therefore better interpreted as evidence of content-mediated technology cognition than as a mere measurement artifact. At the same time, future research should test this distinction more rigorously using designs less vulnerable to conceptual overlap.

4.1.2. The Additional Contribution of Architectural Cultural Identity

Supporting H3, ACI had a significant positive effect on BI (β = 0.110, p < 0.05). Notably this effect remained significant after controlling for TAM variables, confirming a Dual-Track Mechanism. While PU remains the dominant driver of behavioral intention (β = 0.667, p < 0.001), ACI provides a distinct, and distinct motivational force grounded in value identification rather than instrumental utility. Furthermore, AIQ and ACI (r = 0.719) were strongly associated. This pattern suggests that cultural identity not only independently supports learning intention, but is also closely aligned with content perception.

4.1.3. A Three-Dimensional Explanatory Pathway

Taken together, the findings support a three-dimensional “content-technology-culture” pathway for architectural cultural learning in algorithmic environments. Perceived content quality (AIQ) serves as the starting point for technology cognition and is strongly associated with cultural identification; technology cognition predominantly influences behavioral intention through instrumental evaluation, while heritage value internalization contributes a distinct, identity-based motivational pathway to learning intention.

4.2. Exploratory Hypotheses Not Supported

4.2.1. Independence Between Technology Cognition and Cultural Identity

As shown in Table 6, neither PU nor PE significantly predicted ACI, while PEOU exhibited a counter-intuitive negative association in the preliminary testing. Considering the statistical redundancy and the poor model fit when specifying TAM → ACI paths, these results suggest that technology cognition and cultural identity are fundamentally distinct psychological processes.

This independence can be understood through three explanations:

1.

Distinct Psychological Processes

ACI involves meaning construction and value internalization, whereas technology cognition concerns instrumental evaluation of functional systems [24].

2.

The Invisibility of Algorithms

In recommendation-based environments, the actual algorithmic mechanisms remain opaque and “invisible” to the design learner [20,21,60]. As system mechanics recede into the background, the formation of attachment to heritage values relies almost entirely on the cultural value conveyed by the content, rather than on perceptions of functionality or convenience.

3.

Content Value as the Primary Driver

The strong AIQ and ACI correlation, suggesting that identity resonates more with perceived cultural value in content than with usability perceptions [8]. To further clarify their theoretical distinctiveness,

Table 8. Conceptual Distinction Between AIQ and ACI.

Dimension	AIQ	ACI
Object of evaluation	External stimulus (recommended content)	Internal psychology (self-culture relationship)
Psychological nature	Cognitive appraisal of content quality	Affective identification with cultural values
Temporal characteristic	Immediate perception of current content	Developmental identity formed over time
Theoretical tradition	Information systems, HCI research	Social psychology, cultural identity theory

distinguishes AIQ from ACI across four dimensions.

While ACI and AIQ are strongly correlated, cultural identity likely has developmental origins extending beyond immediate content experiences, including prior education, family background, and accumulated cultural exposure.

The current cross-sectional design captures their association but cannot fully model the antecedents of cultural identity formation.

4.2.2. The “Ease-of-Use Paradox” and Cognitive Friction

The significant negative effect of PEOU on ACI, observed during preliminary model testing (β = −0.18, p = 0.005), reveals an “ease-of-use paradox” in AI-mediated architectural cultural learning. This phenomenon suggests that beyond baseline usability, which contemporary platforms universally provide, sustainable heritage engagement depends primarily on content cultural value.

First, excessive ease of use may lead users into a state of frictionless browsing, where minimal cognitive effort is required, fostering a tendency toward “cognitive miserliness” or cognitive heuristic processing. As Byung-Chul Han argues [61], digital systems in the contemporary media environment increasingly pursue “smoothness,” aiming to eliminate resistance, difficulty, and negativity in order to maximize convenience. However, this very smoothness constitutes a cognitive trap: while it lowers access barriers, it also suppresses the cognitive friction necessary for critical reflection, interpretive effort, and sense making [62]. In algorithmic environments dominated by short-form and fragmented content, users may therefore engage less with the evaluative processes required to understand architectural history, spatial logic, and cultural symbolism [30,48,62,63,64,65,66].

Second, a structural tension exists between the fragmented nature of algorithmically recommended content and the integrative structure of architectural knowledge [2,7]. While architectural culture requires the synthesis of historical, spatial, technical, and symbolic dimensions [22,29], fragmented browsing may undermine the formation of coherent knowledge frameworks, resulting in a more superficial engagement with cultural value [67].

Third, cultural identity development presupposes active exploration and value commitment as core psychological stages [48]. Overreliance on algorithmic convenience may reduce users’ motivation for active exploration, limiting the depth of sense making required for cultural identification [30,41,68].

Collectively, the “ease-of-use paradox” provides negative evidence that cultural identity formation operates largely independently of technological convenience. The strong association between AIQ and ACI further suggests that content quality and cultural value, rather than usability-related system characteristics, constitute the primary forces driving deep ACI in algorithmic environments.

4.3. Theoretical Contributions

4.3.1. Content-Driven TAM

This study provides empirical support for a content-driven extension of the TAM [24]. As detailed in Section 4.1.1, AIQ’s exceptional explanatory power (R² = 0.929 for PU) suggests that in black-box algorithmic environments [21], content effectively replaces direct system interaction as the primary interface for technology cognition [17,26,69]. This finding extends the theoretical scope of the TAM by demonstrating that users evaluate system performance through mediated content cues rather than technical mechanisms, thereby establishing a content-mediated pathway to technology acceptance.”

4.3.2. Independence of Technology Cognition and Cultural Identity

The lack of support for technology cognition’s effects on cultural identity reveals critical boundary conditions for the TAM. When systems function primarily as content curators [20] rather than operational tools, and when usage goals involve cultural learning [29] rather than task completion, the traditional TAM’s predictive power substantially diminishes.

This finding challenges assumptions about the TAM’s universal applicability, demonstrating that technology cognition (instrumental evaluation) and cultural identity (value-based sense-making) [48] operate at distinct psychological levels, a distinction particularly salient in algorithmic cultural communication contexts [7,62]. As Carpo [70] argues, the “second digital turn” fundamentally transforms how architectural knowledge is produced and disseminated, shifting from human-centered design logic to algorithm-mediated processes.

4.3.3. Cultural Identity as an Independent Driver

The significant effect of ACI on BI, even after controlling for technological factors, suggests cultural identity [30,48,50] as an independent predictor of continuous learning. Although its effect was smaller than that of perceived usefulness, it reflects a value-based form of motivation distinct from utility-driven responses.

The strong AIQ–ACI association (r = 0.719) further suggests that cultural identity operates as a complementary pathway to technological acceptance in sustaining engagement with algorithmic cultural content [8].

4.4. Revised Theoretical Model

Building on the empirical results, we propose a revised theoretical framework, the CDDT of Architectural Cultural Learning in Algorithmic Recommendation Environments (see Figure 3). The model illustrates two parallel pathways that reflect two theoretically distinct cognitive processes through which AI-recommended architectural content influences learning intention.

The model highlights that:

• As shown on the left side of the figure, the Baseline Cognitive Track (Instrumental Track) operates through content-technology cognition. AIQ drives technology acceptance (PU, PEOU, and PE), which facilitates short-term behavioral intention through instrumental rationality. This pathway explains immediate engagement driven by system utility perceptions.
• As shown on the right side of the figure, the Core Motivational Track (Value Track) operates through content-culture integration. AIQ is associated with architectural cultural identity (commitment and exploration), which independently predicts behavioral intention. This pathway reflects value-based motivation associated with distinct from instrumental technology evaluation;

The two pathways operate in parallel rather than sequentially. The baseline cognitive track (PU → BI) reflects instrumental evaluation of system utility, while the core motivational track (ACI → BI) reflects identity-based engagement with architectural cultural values. Together, they explain how high-quality AI recommendation cultural content supports both behavioral intention through both technology acceptance and cultural identification.

4.5. Implications for Cultural Sustainability in the Built Environment

Our findings extend the discourse on AI’s role in the cultural dimension of sustainable building. Sustainable building research increasingly recognizes that heritage preservation depends not only on physical conservation but also on continued public engagement and cultural identification [8,14]. The present study contributes to this understanding by examining how AI recommendation systems on social media shape users’ cultural connection with architectural heritage.

Our results suggest that AI recommendation systems can contribute to cultural sustainability, but not through technological convenience alone. The observed “Ease-of-Use Paradox” reveals a critical insight: when algorithmic browsing becomes too smooth, it may actually prevent the kind of deep thinking needed for people to connect with and identify meaningfully with cultural content. Instead, content quality, underpinned by originality, credibility, and cultural relevance, emerges as the cornerstone of meaningful heritage engagement.

These findings offer practical implications for heritage institutions and digital platform designers. If recommendation algorithms prioritize short-term engagement metrics over cultural depth, they risk undermining the public’s deeper connection with architectural heritage. Conversely, algorithms designed to balance accessibility with cultural richness can serve as effective tools for sustainable heritage dissemination [15,71]. The CDDT provides a psychological framework for understanding how AI technologies can support cultural sustainability in the built environment.

5. Conclusions and Prospects

5.1. Research Conclusions

This study investigated how AI recommendation mechanisms influence architectural cultural learning through a CDDT framework using SEM.

Three major findings emerged:

First, a content-driven pathway of technology cognition was confirmed. AIQ exhibited strong predictive effects on all three TAM dimensions, confirming that designers infer system value primarily through content experience.

Second, the independence between technology acceptance and cultural identity was empirically demonstrated. Technology cognition did not significantly predict cultural identity. Instead, the “Ease-of-Use Paradox” (PEOU → ACI negative) suggests that excessive convenience impedes the cognitive friction necessary for deep identity formation.

Third, ACI emerged as a significant predictor of behavioral intention even after controlling for technological factors. This finding highlights cultural identity as a distinct motivational force associated with sustained engagement beyond short-term system acceptance.

Together, these findings support the CDDT: perceived usefulness strongly predicts behavioral responses, while ACI provides an additional value-based pathway theoretically linked to long-term engagement.

More importantly, the findings indicate that in algorithmic environments, architectural cultural identity appears to be more closely associated with perceived content value than with technological convenience. This reframing clarifies that sustainable digital heritage engagement depends less on interface optimization and more on the cultural depth embedded within algorithmically delivered architectural content.

5.2. Practical Implications

The results offer practical implications for architectural education, cultural communication, and heritage management in the AI era.

5.2.1. Professional Development in Architecture: Avoiding the “Technological Convenience Trap”

Educators should combat shallow learning by designing “cognitive friction” tasks (e.g., reverse recommendation, cross-platform comparison) and developing courses on algorithmic literacy to promote critical evaluation of AI-recommended content, particularly in architectural history and structural knowledge. This aligns with calls for design thinking approaches [72], digital pedagogy integration [73], and experiential learning methods [74] into contemporary architectural education.

5.2.2. Cultural Communication: From “Traffic Logic” to “Identity Logic”

Prioritize content quality over technology by ensuring originality, credibility, and personalization. Balance algorithmic personalization with cultural diversity to avoid echo chambers. Shift evaluation metrics from short-term traffic to long-term cultural identity indicators, drawing on emerging methods for assessing cultural heritage value perception through social media data [71].

5.2.3. AI System Design: Cultural Semantics and Diversity

Although this lies beyond the scope of the present empirical findings, future AI system design may benefit from moving beyond collaborative filtering toward a deeper understanding of cultural semantics. NLP and computer vision technologies could support this shift by helping construct cultural knowledge graphs. In addition, ranking algorithms could incorporate diversity constraints, while system interfaces could offer dual modes for quick browsing and deeper learning.

5.2.4. Connecting Identity to Action

It is observed that ACI, particularly the commitment dimension, provides a psychological foundation for heritage protection participation. Cultural institutions should embed action entry points (e.g., volunteering, guided tours, community practices) in content to transform online identity into offline participation, forming a “learning–identity–participation” chain.

5.3. Research Limitations and Future Directions

First, the cross-sectional design restricts causal inferences regarding long-term behavior. Although theory suggests that ACI may support enduring motivation, our data only confirm its association with behavioral intention. Longitudinal research is therefore needed to examine whether such intention translates into sustained engagement over time [75].

Second, the sample consisted of architecture and design major students in Taiwan. As domain learners, their preference for professional depth over accessibility likely explains the observed “Ease-of-Use Paradox.” Future studies should apply this framework to non-expert populations and more diverse cultural contexts. Such studies would clarify whether the observed “content-over-ease” pattern persists across different levels of domain expertise.

Third, a further limitation concerns the conceptual proximity between AIQ and PU. Despite acceptable VIF values and distinct CFA factor structures, the HTMT ratio between AIQ and PU reached 0.924, exceeding conventional thresholds [59]. This suggests that the two constructs may be closely intertwined in algorithm-mediated contexts. Although this pattern may reflect the content-mediated nature of usefulness judgments, some degree of conceptual overlap may also have inflated the strength of the AIQ → PU relationship. In addition, while Harman’s single-factor test did not indicate that common method bias was a serious concern, shared method inflation cannot be fully excluded in a single-source self-report design. Future research should examine this boundary more rigorously through experimental, longitudinal, or multi-method designs.

Finally, the explanatory scope of the model may be improved by incorporating additional moderators, such as digital literacy, prior cultural knowledge, or familiarity with AI-generated content. Including such variables may help clarify for whom and under what conditions content quality, usability, and cultural identity interact most strongly in AI-mediated architectural learning.