Governing AI innovation in local government: how trust and risk perceptions shape public support

Recent technological advances in AI have accelerated its adoption in local government operations, particularly within cities where AI increasingly shapes everyday urban experiences (Vogl et al., 2020; Yigitcanlar et al., 2024a). A global review of real-world AI applications in public services shows that leading domains, including information management, administrative services, transportation, waste collection, infrastructure maintenance, and social welfare, are deeply embedded in urban spaces and affect how residents interact with city systems (Yigitcanlar et al., 2024b). These applications provide multidimensional benefits for municipalities and communities, including efficiency, transparency, and responsiveness, while also reshaping the spatial organisation of service delivery and citizen-state interaction (Kulal et al., 2024; Anshari et al., 2024). For example, the Hong Kong government has promoted smart governance through AI-driven civil service initiatives, including department-wide chatbots (Wong, 2023).

At the same time, the deployment of AI in urban governance raises critical ethical and political concerns related to fairness, accountability, and spatial inclusion (Yigitcanlar et al., 2020a; Poudel, 2024). Predictive policing systems and algorithmic resource allocation have been criticised for perpetuating bias against marginalised neighbourhoods and social groups, potentially reinforcing exclusionary practices (Ziosi and Pruss, 2024). Such risks amplify the calls for AI governance frameworks that are sensitive not only to ethical principles but also to the spatial contexts in which AI operates (Son et al., 2023; Lifelo et al., 2024). Moreover, AI systems remain inherently imperfect, with technical limitations related to bias, error, and opacity that can manifest unevenly across urban populations. Responsible-AI in local governance therefore requires attention to transparency, accountability, public trust, and broader social values as they play out across diverse urban settings (Díaz-Rodríguez et al., 2023).

Public perceptions and acceptance dynamics are central to the legitimacy and sustainability of AI in cities. As AI becomes more visible in streets, transport systems, service platforms, and public safety infrastructures, residents' everyday encounters with these technologies shape how they interpret risks, benefits, and governance responsibilities. Achieving responsible-AI implementation therefore depends on inclusive governance that recognises diverse perspectives and engages stakeholders ranging from technology officers and policymakers to community organisations and citizens (Rakova et al., 2021; Son et al., 2023). Without attention to spatially differentiated experiences and public controversies around bias, surveillance, and unequal impacts, AI governance risks deepening social and spatial divides within cities (Ulnicane et al., 2021).

Although research on AI in government is expanding (Reis et al., 2019; Valle-Cruz et al., 2019; Ahn and Chen, 2022), empirical studies examining ethics, public trust, and participation at the municipal level remain limited, especially from the perspective of residents' lived experiences in cities (Taeihagh, 2021; Zuiderwijk et al., 2021; Yigitcanlar et al., 2022). Understanding how urban residents perceive and evaluate AI systems is critical for designing governance mechanisms that respond to local expectations and accommodate diverse sociotechnical and spatial contexts (Lodato et al., 2021).

This study examines what drives supportive and critical perceptions of AI in local government, how these perceptions shape the prioritisation of responsible-AI principles, and how these dynamics vary across Australia, Hong Kong, and Saudi Arabia. These three contexts were selected because they represent distinct trajectories of municipal AI adoption, governance capacity, urban digitalisation, and public engagement, providing an analytically useful basis for comparison. Building on recent research on public perceptions of AI and responsible local government AI governance (Yigitcanlar et al., 2024c; Brauner et al., 2025), this study advances a municipal-level account of public sector AI innovation. Its contribution lies in showing how local governments operate as practical settings where residents encounter and evaluate AI through everyday services, rather than only through national strategies or abstract policy frameworks. By modelling relationships among experience, understanding, trust, perceived risks, perceived benefits, application-specific AI attitudes, and responsible-AI priorities, the study explains how public support for AI is formed in local service contexts. It further distinguishes routine urban-AI applications from more contested surveillance-AI applications, enabling a more nuanced understanding of how visible, service-oriented AI may shape public openness to sensitive municipal AI uses while informing responsible and inclusive AI governance (Hoang, 2024).

This study is informed by responsible innovation, public sector innovation governance, and trust-based technology acceptance perspectives (Stilgoe et al., 2013; Li et al., 2023; Ardabili et al., 2024; Chen et al., 2025). Responsible innovation emphasises the need to anticipate potential impacts, include diverse perspectives, reflect on social and ethical consequences, and respond to emerging concerns in the governance of new technologies (Stilgoe et al., 2013). These principles are particularly relevant to local government AI because municipal technologies are implemented close to residents' everyday lives and must balance innovation benefits with accountability, legitimacy, public value, and social acceptability.

Public sector innovation research further highlights that AI adoption in government is not only a technical or efficiency-oriented process, but also an institutional process through which new technologies are tested, legitimised, and embedded in public service systems under conditions of organisational context, publicness, service delivery responsibilities, and governance capacity (Walker, 2008; Cinar et al., 2024; Tilly et al., 2025). Municipal contexts are distinctive because AI innovation becomes visible through everyday urban services, such as traffic management, administrative support, infrastructure maintenance, and public safety systems. Unlike national AI frameworks, which often operate at the level of policy direction, regulatory positioning, or economic strategy, local government AI is experienced directly by residents through service encounters and urban infrastructure systems. This makes local governments important settings for examining how responsible innovation principles are translated into practice.

Accordingly, this study moves beyond a purely instrumental view of AI acceptance by treating public support for local government AI as shaped not only by perceived usefulness, but also by trust, risk awareness, and responsible governance expectations. Building on these perspectives, the conceptual model examines how experience, understanding, trust, risk perception, application-specific AI attitudes, and responsible-AI priorities are connected in local government contexts. The study conceptualises an exposure-to-ethics pathway in which direct experience with local government AI shapes perceived understanding, which then influences trust. Trust, in turn, informs perceived benefits, perceived risks, and attitudes toward urban-AI applications. The model further distinguishes routine urban-AI applications from more contested surveillance-AI applications to examine whether support for visible, service-oriented AI can extend to more sensitive municipal AI uses. Finally, the model assesses how perceived risks and urban-AI attitudes shape the prioritisation of responsible-AI characteristics. This provides a relational framework for explaining how residents move from everyday exposure to broader expectations about responsible-AI governance.

Prior research highlights the relationship between exposure, comprehension, and trust. Experience with AI technologies can increase familiarity and improve understanding of a technology's functions and societal role (Ajitha et al., 2024). Understanding AI is a critical factor in building trust, particularly when individuals can make sense of how AI systems operate and how they are applied in public services (Scantamburlo et al., 2025). When hands-on experience aligns with perceived trust, people are more likely to develop acceptance and satisfaction toward technology (Lankton et al., 2014). Studies also show strong associations between direct experience, trust, and understanding of emerging technologies, including AI (Glikson and Woolley, 2020; Reagan et al., 2023).

Applied to local government AI, individuals who encounter AI-enabled services may develop a clearer understanding of how these technologies are used in municipal settings. This understanding can strengthen trust by making AI systems appear more familiar, explainable, and institutionally embedded (Scantamburlo et al., 2025). At the same time, understanding may also make individuals more aware of AI's limitations, risks and governance requirements, creating a more conditional form of trust.

Trust plays an important role in shaping beliefs, attitudes, and behaviours toward technological innovation (McKnight et al., 2009; Lankton et al., 2014). It can reduce perceived risks and amplify perceived benefits, thereby encouraging more favourable views of technological applications (Seo and Lee, 2021). However, public attitudes toward AI often differ from attitudes toward other technologies because AI raises distinctive concerns about accountability, opacity, and social consequences (Crockett et al., 2020). In local government contexts, trust is therefore expected to shape both favourable and critical perceptions of AI.

Trust in local government AI may increase perceived benefits by making AI appear more capable of improving service efficiency, infrastructure management, and administrative responsiveness. It may also reduce perceived risks where residents believe that local governments have the competence and responsibility to use AI appropriately. However, trust does not necessarily mean uncritical acceptance. Individuals may trust institutions while still recognising ethical, legal, and technical uncertainties surrounding AI. This is especially relevant in local government, where AI systems are embedded in everyday services and can directly affect residents' interactions with public authorities.

Attitudes toward specific AI applications are highly relevant in local governance because residents encounter AI through different service settings and levels of sensitivity. Local government AI tools such as traffic management systems or crime hotspot detection are often perceived as service-oriented and beneficial. In contrast, surveillance-oriented AI, particularly facial recognition, is more contested because it raises concerns about privacy, discrimination, and institutional power. Critical sociotechnical scholarship on surveillance shows that algorithmic monitoring can normalise unequal visibility and reinforce existing social hierarchies, particularly when data-driven systems are deployed in public spaces without sufficient transparency and oversight (Lyon, 2007; Browne, 2015; Eubanks, 2018). Therefore, public openness to surveillance-AI should not be interpreted only through familiarity or perceived usefulness, but also through broader ethical and governance concerns.

It is therefore essential to understand how AI is embedded in cities and how perceptions vary across application scenarios (Zhu and Liu, 2025). Generally, the public recognises the potential benefits of local government AI; however, perceptions differ according to the type and purpose of implementation (Dorotic et al., 2024; Yigitcanlar et al., 2024c). Research suggests that positive perceptions of AI in urban services can extend to broader AI acceptance, as functional value shapes public attitude toward technology (Vongvit et al., 2024). For instance, familiarity and perceived benefits of surveillance can influence acceptance of facial recognition technologies (Choung et al., 2024). Furthermore, community perspectives are essential in contested domains such as AI-driven surveillance, where perceptions differ and often involve concerns of trust and privacy (Ardabili et al., 2024).

This study, therefore, examines whether positive attitudes toward routine urban-AI applications extend to perceptions of more contested surveillance-AI applications and to broader perceived benefits of local government AI. This provides insights into the conditions under which optimism toward familiar AI tools in cities may contribute to broader support for, or caution toward, more sensitive AI uses.

The model examines how public perceptions shape the prioritisation of responsible-AI characteristics. Responsible-AI in local government involves more than technical performance; it requires attention to fairness, transparency, privacy, accountability, security, affordability, sustainability, and stakeholder participation. These characteristics are consistent with responsible innovation principles and with the Responsible Innovation Technology (RIT) framework used to operationalise the responsible-AI construct in this study (Li et al., 2023).

Research suggests that public perceptions of AI play an important role in shaping ethical adoption and policy development (Brauner et al., 2025). While the public may recognise the benefits of AI, preferences for AI governance and regulation are often strongly shaped by perceived risks (O'Shaughnessy et al., 2023; Bullock et al., 2025). This is particularly relevant in local government, where AI systems can affect service access, public safety, privacy, and trust in public institutions. Attitudes toward specific AI applications also influence broader AI opinions, suggesting that residents' evaluations of urban-AI may shape how strongly they prioritise responsible-AI requirements (Yigitcanlar et al., 2022). At the same time, the importance placed on ethical principles is not uniform across groups or contexts, indicating that AI governance must balance democratic values with local expectations and institutional realities (Kieslich et al., 2022; Ajitha et al., 2024).

Accordingly, this study examines whether perceived risks, positive attitudes toward urban-AI, and perceived benefits influence the importance residents place on responsible-AI characteristics in local government. This allows the model to assess whether responsible-AI priorities are driven mainly by caution, optimism, or perceived public value.

This study employs a quantitative survey design, drawing on data collected through an online questionnaire administered in Australia, Hong Kong, and Saudi Arabia. Structural equation modelling (SEM) was used to examine the hypothesised relationships among the study constructs.

Australia, Hong Kong, and Saudi Arabia were selected to capture contrasting contexts of local government AI adoption. While they differ in governance traditions, AI maturity, and public engagement, all three illustrate how AI is increasingly deployed through city-scale services that shape residents' everyday interactions. Their selection was also informed by research scope, feasibility of cross-context data collection, and relevance to ongoing debates on responsible and inclusive municipal AI. Each context has made notable advances in local government AI, providing an analytically useful basis for comparing how public perceptions are shaped by different institutional, demographic, and urban conditions (Yigitcanlar et al., 2024c).

Saudi Arabia represents rapid, state-led digital transformation, with AI positioned as a central tool in urban modernisation agendas. Government initiatives emphasise efficiency, sustainability, and service quality, reflecting broader national ambitions for digital transformation and smart urban development (UNESCO, 2023; Digital Government Authority, 2025; Helou, 2025). These initiatives both reflect and reshape the spatial and institutional configuration of cities, making public perspectives important for ensuring responsible and socially acceptable implementation.

Hong Kong illustrates a dense, infrastructure-rich urban context with strong technical capacity and widespread deployment of smart city technologies. Its ambition to become an AI hub is supported by advanced digital infrastructure and research capacity, although governance arrangements continue to evolve, including debates around the need for dedicated AI regulatory frameworks (Chan and Tsoi, 2025). In this context, AI is highly visible in urban services, making public opinion especially relevant for legitimacy, trust, and the politics of urban inclusion.

Australia reflects a democratic urban governance context where AI is increasingly integrated into local public services, including planning, transport, and service delivery. While AI's transformative potential is recognised, regulatory frameworks and public awareness remain uneven, and trust issues continue to influence adoption and acceptance. These conditions highlight the importance of participatory and community-aligned approaches to support safe, trusted, and responsible-AI within Australian cities (Yigitcanlar et al., 2024c).

Taken together, these three contexts provide analytically useful contrasts in how AI is embedded in urban governance systems. Saudi Arabia exemplifies state-led acceleration, Hong Kong represents advanced infrastructure with evolving governance arrangements, and Australia represents democratic adoption shaped by trust, awareness, and accountability expectations. This variation provides a basis for examining how public perceptions of local government AI are shaped by different political, institutional, and spatial configurations.

Ethics clearance was granted by the University's Human Research Ethics Committee. Data were collected online using a Qualtrics survey. The online format enabled efficient cross-context data collection, ensured standardised delivery, and provided anonymity and flexibility, making it suitable for AI perception and technology adoption research (Deutskens et al., 2006; Gerlich, 2023).

A minimum sample size of 385 respondents per country was estimated using a 95% confidence level (Krejcie and Morgan, 1970; Fritsch et al., 2022). The final dataset comprised 1,183 respondents: 388 from Australia, 407 from Saudi Arabia, and 388 from Hong Kong. Only individuals aged 18 years or above were eligible to participate.

In Australia, participants were recruited through Prolific (May–June 2024), a commercial online panel platform, with demographic distributions cross-checked against census data to improve sample balance. In Hong Kong and Saudi Arabia, participants were recruited through partner networks using convenience sampling. While this approach enabled access to respondents in different contexts, it may have introduced sampling bias. Accordingly, the findings are interpreted as structural perception patterns observed within the collected samples, rather than as nationally representative estimates of public opinion.

The survey was structured for ease of completion and required approximately 15–20 min to complete. A pilot study involving 50 participants was conducted to identify and address potential issues related to question order and clarity (Memon et al., 2017). Participation was voluntary, informed consent was obtained, and respondents could withdraw at any stage. While no prior experience with AI was required, respondents' familiarity and exposure to AI-enabled urban services were assessed through targeted questions.

The survey instrument was designed to test the study hypotheses, with latent constructs representing unobservable theoretical dimensions and indicators serving as measurable variables of the constructs.

The experience constructs captured respondents' direct interactions with AI-enabled municipal services encountered in everyday urban contexts, such as smart parking and traffic management, ensuring that exposure reflected diverse city-based service environments (Ostrom et al., 2019; Keng et al., 2025). Perceived understanding was evaluated through participants' self-reported clarity regarding specific local government AI applications (Brauner et al., 2023). Trust in local government AI reflected beliefs about institutional competence, responsible use, and stewardship of data and privacy within urban governance systems (Crockett et al., 2020; Brauner et al., 2023).

Perceived benefits captured respondents' beliefs about AI's contribution to efficiency, infrastructure management, and public safety (Ahn and Chen, 2022). Perceived risks captured concerns related to the use of municipal AI, including potential social, ethical, and governance risks (Bullock et al., 2025). Positive attitudes toward urban-AI were assessed through perceived usefulness of applications such as traffic management and crime hotspot prediction, which are connected to city environments and everyday urban services (Yigitcanlar et al., 2020b; Lehtiö et al., 2023; Wolniak and Stecuła, 2024). Positive attitudes toward surveillance-AI were measured through indicators reflecting comfort with and perceived usefulness of surveillance technologies that raise contested spatial and ethical questions in urban settings (Ardabili et al., 2024; Choung et al., 2024). The perceived importance of responsible-AI was measured using 15 characteristics, including transparency, fairness, security, affordability, and accountability based on Li et al.’s (2023) RIT framework.

Demographic variables, including age, gender, and education, were incorporated as observed covariates to assess their association with experience, understanding, and trust (Ardabili et al., 2024). All constructs were measured using at least three indicators on a five-point Likert scale to enhance measurement robustness (Si et al., 2022). To mitigate common method bias, several procedural remedies were applied, including anonymity assurance, separation of items measuring predictors and outcomes, and variation in wording and response anchors (Al-Sharafi et al., 2023). The item-level indicators and standardised factor loadings are reported in Table 2, while the complete survey wording for each item is provided in Supplementary Material Table A1.

Data were analysed in IBM AMOS using a two-step process. First, confirmatory factor analysis (CFA) was conducted to validate the measurement model. Second, the structural model was estimated to examine the hypothesised relationships among latent constructs.

Reliability and validity were assessed through factor loadings, Cronbach's alpha, Composite Reliability (CR), and Average Variance Extracted (AVE). Internal consistency was assessed through Cronbach's alpha and CR, while convergent validity was assessed using AVE. Discriminant validity and measurement invariance were also examined to ensure that the constructs provided an appropriate basis for cross-context SEM analysis.

The structural model was estimated using Covariance-Based Structural Equation Modelling (CB-SEM). Multi-group SEM was then employed to examine cross-context differences in structural relationships. Model fit was evaluated using Chi-Square/Degrees of Freedom (CMIN/DF), Comparative Fit Index (CFI), Tucker-Lewis Index (TLI), Root Mean Square Error of Approximation (RMSEA), and Standardised Root Mean Square Residual (SRMR). Following standard practice, configural, metric, and scalar invariance were tested sequentially. The measurement and structural results are reported in Section 4, including model fit indices and structural path estimates, while the complete item wording and additional diagnostics are provided in  Supplementary Tables.

Model development proceeded in two stages. First, a theory-based baseline model was estimated including all hypothesised paths. Second, the model was refined using clearly defined decision rules intended to reduce redundancy and enhance parsimony while preserving theoretical coherence and statistical adequacy: a direct path was considered for removal only when (1) its theoretical justification for a direct effect was limited, (2) the path was statistically non-significant in at least two of the three samples, and (3) its removal did not worsen model fit.

Applying these criteria led to the removal of the perceived benefit to responsible-AI path. Prior research suggests that, although perceived benefits often influence adoption, governance preferences are frequently shaped more strongly by perceived risks (O'Shaughnessy et al., 2023; Bullock et al., 2025). Empirically, this path was negligible in Australia and Hong Kong and did not improve model fit. Its removal also reduced a suppression pattern, clarifying the positive urban-AI to responsible-AI association in Australia that was obscured in the baseline model.

The refined model demonstrated equivalent performance to the baseline model across contexts, indicating no meaningful deterioration in model fit, with the change in Comparative Fit Index (ΔCFI) remaining below the recommended 0.01 threshold (Nowak and Zajkowski, 2025). Although the chi-square difference (Δχ²) was statistically significant in Saudi Arabia, which can occur with larger samples, the associated ΔCFI was 0.001, indicating negligible practical impact on model fit (Sass et al., 2014). The final model was therefore considered parsimonious, theoretically coherent, and statistically well-fitting, enabling clearer interpretation of the structural relationships. Figure 1 illustrates the research model.

The final dataset comprised 1,183 respondents, including 388 from Australia, 388 from Hong Kong, and 407 from Saudi Arabia. Overall, 59.9% of respondents identified as female, 39.4% as male, and 0.7% as other (Table 1). The sample was relatively young, with the largest age group being 18–24 years (43.9%). More than half of the respondents held a bachelor's degree (51.1%), and 67.2% reported moderate to high familiarity with AI, which is particularly relevant given the study's focus on public perceptions of AI in everyday local government.

Country-level differences were evident. The Australian sample had the most balanced demographic profile across age, gender, and education. The Hong Kong sample showed variation in education levels, while the Saudi Arabian sample was younger, more female-skewed, and highly educated, with over 70% aged 18–24 years, 77.4% female, and 61.2% holding a bachelor's degree. These demographic differences reinforce the need to interpret cross-context comparisons cautiously as patterns observed within the collected samples, rather than as nationally representative estimates of public opinion.

To account for demographic variation, age, gender, and education were included as observed covariates in the SEM analysis (O'Shaughnessy et al., 2023; Ardabili et al., 2024). This enabled the analysis to assess whether demographic characteristics were associated with experience, understanding, and trust in local government AI, while keeping the primary focus on the structural relationships among the latent constructs. Rather than aiming for full population generalisability, the analysis focuses on structural patterns of perception within and across country samples, recognising the influence of contextual realities and demographic composition.

Reliability and validity of the measurement models were evaluated using CFA. Common method bias was first evaluated using Harman's single-factor test, which showed that the first factor accounted for 25% of the total variance, well below the 50% threshold, indicating that common method bias is unlikely to be a dominant concern (Al-Sharafi et al., 2023). As an additional robustness check, an unmeasured latent method factor test was conducted in AMOS using the pooled sample. The inclusion of the method factor did not improve model fit, with CFI decreasing from 0.928 to 0.920, indicating that common method bias is unlikely to substantially affect the estimated relationships.

Internal consistency was confirmed, with all constructs exceeding the recommended 0.7 thresholds for Cronbach's alpha and CR (Shrestha, 2021). Convergent validity was generally supported, as AVE values were above 0.50 or acceptable when CR exceeded 0.60 (Huang et al., 2013; Shrestha, 2021). Some constructs, particularly experience, responsible-AI, and urban-AI, showed slightly lower AVE values in country-specific samples, which is expected for broad constructs with heterogeneous indicators (Reise et al., 2013; Henseler et al., 2015). Experience captures multiple forms of AI-enabled local government services, urban-AI spans different application scenarios, and responsible-AI encompasses a wide set of governance principles, including fairness, transparency, privacy, accountability, security, affordability, and sustainability. Treating responsible-AI as a single factor therefore involves a modelling trade-off, as it captures overall prioritisation of responsible-AI characteristics rather than potential subdimensions. Nonetheless, convergent validity remains acceptable because higher CR values confirmed the constructs' reliability and theoretical relevance. Accordingly, these constructs were retained as broad constructs rather than as narrow technical subdimensions, consistent with their theoretical framing.

Discriminant validity was assessed using the Fornell-Larcker criterion (Fornell and Larcker, 1981). All constructs met the criterion except the urban-AI and surveillance-AI pair in the Australian and Hong Kong samples, where the inter-construct correlation marginally exceeded the square root of AVE for urban-AI. This likely reflects practical overlap between urban and surveillance-related AI applications, as technologies such as traffic monitoring and surveillance systems can share similar spatial and technological foundations. Given their conceptual distinction, policy salience and the study's aim to estimate differential pathways, the two constructs were retained as distinct constructs, and results are interpreted with a focus on structural path estimates rather than comparisons of latent means.

Factor loadings exceeded 0.50 for all retained indicators, demonstrating acceptable construct representation (Cheung et al., 2024). Multi-group CFA supported configural and metric invariance, indicating that the factor structure and loadings were comparable across contexts. However, full scalar invariance was not established (ΔCFI >0.01). As the analysis focuses on relationships rather than mean differences, configural and metric invariance were sufficient for multi-group SEM. Accordingly, cross-context comparisons are interpreted in terms of relative structural associations within the collected samples, not as absolute comparisons of latent construct mean. Table 2 reports the item-level indicators and standardised factor loadings for each construct across the three samples, while the complete survey wording, reliability and validity results, covariate effects, and indirect effects are provided in the  supplementary material. Overall, the models provided an acceptable basis for structural model testing.

The structural models assessed the hypothesised pathways and mediating dynamics among constructs. As shown in Table 3, model fit indices indicated satisfactory fit across Australia, Hong Kong, and Saudi Arabia (CMIN/DF < 3; CFI >0.90; TLI >0.90; RMSEA <0.08; SRMR <0.08). These results supported the adequacy of the final structural model (Chen et al., 2025).

The final model supported the hypothesised pathways across the three samples, although the direction and strength of some relationships varied by context (Table 4). Experience with local government AI was positively associated with perceived understanding, supporting H1. Perceived understanding was positively associated with trust, supporting H2. Trust was positively associated with perceived benefits and urban-AI attitudes across the three samples, supporting H3(a) and H3(c). The trust to risk pathway, H3(b), was negative in Australia but positive in Hong Kong and Saudi Arabia. This indicates a cross-context difference in the direction of the trust-risk relationship, which is interpreted further in Section 5.

Positive attitudes toward urban-AI were positively associated with surveillance-AI attitudes and perceived benefits across the three samples, supporting H4(a) and H4(b). Perceived risk was positively associated with responsible-AI prioritisation, supporting H5(a). Urban-AI attitudes were also positively associated with responsible-AI prioritisation, supporting H5(b). The perceived benefits to responsible-AI pathway corresponding to H5(c) was tested in the baseline model but removed during model refinement under the stated decision rules because it was negligible in two samples and did not improve model fit.

Indirect effects showed that trust operated as an important mediating construct, transmitting effects toward perceived benefits, surveillance-AI attitudes, and responsible-AI prioritisation through the model pathways. Perceived understanding also produced indirect effects through trust, while experience influenced downstream constructs mainly through perceived understanding and trust. These indirect effects are reported in Supplementary Table A3.

Demographic covariates showed weaker and less consistent associations than the main perception pathways. Age was the only covariate with a consistent pattern, with younger respondents reporting higher experience with local government AI across all three samples. Education was positively associated with perceived understanding in Hong Kong and Saudi Arabia, while gender effects were limited and context-specific. These covariate results are reported in Supplementary Table A2 and should be interpreted cautiously given the sample compositions.

Section 5 discusses the implications of these pathways for institutional trust, critical risk awareness, urban-AI as a gateway to more contested applications, and context-sensitive, responsible-AI governance. Figure 2 shows the path diagram of the models for Australia, Hong Kong and Saudi Arabia.

This section interprets the empirical results through a comparative and contextualised lens, integrating structural relationships, indirect effects, and demographic influences. It offers a multidimensional account of how residents experience, perceive, and prioritise AI in everyday local governance, where public support depends not only on perceived service value and technical functionality, but also on trust, accountability, risk awareness, and social acceptability.

Across the three samples, direct experience with local government AI consistently enhanced perceived understanding, which in turn reinforced trust. This experience-to-trust pathway suggests that exposure to AI in everyday urban services does more than familiarise residents with technology. It helps people to contextualise AI's role in public service delivery, recognise its operational logic, and evaluate the capacity of local governments to use AI responsibly. This pattern is consistent with prior work showing that familiarity and understanding are important precursors to trust and acceptance of AI (Glikson and Woolley, 2020; Ajitha et al., 2024).

The indirect effects reinforce this interpretation. Experience shaped trust through perceived understanding, indicating that trust is formed not only through general confidence in public institutions but also through situated encounters with AI-enabled services. In local government contexts, residents rarely evaluate AI as an abstract technology. They encounter it through traffic systems, service platforms, public safety tools, automated scheduling, or administrative interfaces. These encounters can translate technological innovation into institutional trust when residents are able to observe practical value and understand how AI is being used.

However, this pathway is not evenly distributed. Demographic covariates showed that younger respondents reported higher exposure to AI across all three samples, reflecting generational differences in engagement with digitally mediated urban services. This is consistent with prior work showing greater openness to emerging technologies among younger cohorts (Kassens-Noor et al., 2024). In Hong Kong and Saudi Arabia, education was positively associated with perceived understanding, while age was negatively associated with understanding. In Hong Kong, age also indirectly reduced trust via diminished understanding, suggesting that generational gaps in comprehension can shape trust. Gender effects were minimal and context-specific.

Experience also had modest and context-dependent indirect effects on risk perceptions. In Australia, exposure slightly reduced perceived risks, whereas in Hong Kong and Saudi Arabia, the effects were negligible and positive. This does not necessarily contradict the experience-to-trust pathway. Rather, it suggests that familiarity with AI can coexist with risk awareness, particularly in samples that are younger, more educated, or more exposed to digitally mediated public services. These patterns indicate that experience may support understanding and trust, but its effects are filtered through digital literacy, demographic composition, and the maturity of AI implementation in local services.

For local governments, the implication is that trust cannot be built only through policy statements or general claims about innovation. It requires visible, user-facing initiatives that make AI understandable and open to public evaluation. Demonstration projects, participatory trials, service-specific explanations, and interactive dashboards can help bridge the gap between abstract policy and lived experience. Such initiatives are not merely communication tools; they are mechanisms through which public sector AI becomes socially legible and institutionally credible.

Trust consistently emerged as a key driver shaping both optimism and caution toward AI in local governance. Across the three samples, higher trust increased perceived benefits and supported more positive attitudes toward urban-AI applications such as traffic management and crime prediction. The stronger estimates for the trust to urban-AI pathway indicate that trust most strongly influences positive perceptions when benefits are tangible, place-based, visible, and closely tied to everyday urban services. This suggests that trust is not only an abstract confidence in government but also a practical endorsement of AI tools that promise efficiency, reliability, and public value.

Prior cross-national research similarly finds that higher trust is associated with more favourable AI attitudes (Scantamburlo et al., 2025). The indirect effects further emphasise trust's wider influence, as trust shaped surveillance-AI openness and responsible-AI priorities through urban-AI attitudes. This indicates that confidence in local government AI can cascade into broader perceptions when mediated by visible and functional urban applications. However, the magnitude and direction of these effects remain contextually contingent.

The relationship between trust and perceived risks revealed more nuanced patterns. In the Australian sample, trust significantly reduced risk perception, aligning with prior research showing that institutional confidence can mitigate public concern about AI (Seo and Lee, 2021). By contrast, in the Hong Kong and Saudi Arabian samples, trust's effect on risk was weaker and positive. This pattern is consistent with a critical-trust interpretation, where trust in institutional or technical capacity can coexist with heightened awareness of ethical, legal, and privacy-related risks. This is especially relevant in samples where education and perceived understanding are associated with greater awareness of AI use, suggesting that trust may sometimes enable more informed scrutiny rather than simple acceptance.

This interpretation should nevertheless be treated cautiously. Examination of indirect and covariate effects suggests that suppression effects may partly explain this pattern, although response-style differences, sample composition, and partial non-invariance may also have contributed to the observed sign reversal. Accordingly, these findings suggest that trust in local government AI does not necessarily remove concern. In some contexts, more trusting or more informed respondents may still recognise the need for safeguards, particularly when AI is associated with data use, surveillance, or automated decision-making. This aligns with prior evidence showing that technical familiarity and confidence in AI systems can coexist with reservations about specific risks (Crockett et al., 2020).

These findings refine the role of trust in public sector AI adoption. Trust is not simply a driver of acceptance, nor should it be treated as a tool for reducing public concern. Instead, trust functions as a conditional governance resource. It can support acceptance by making AI-enabled services appear useful, reliable, and institutionally legitimate, while also creating conditions for more critical evaluation. For local governments, trust-building should therefore be accompanied by transparency, accountability, and participatory communication, rather than used as a substitute for them. Trustworthy AI governance should create conditions in which residents can recognise AI's practical value while still expecting safeguards against ethical, privacy, and accountability risks.

A particularly important finding of this study is the role of urban-AI as a gateway for broader AI acceptance. Positive attitudes toward urban-AI strongly predicted openness to more controversial surveillance-AI across all three samples, with consistently high standardised effects reported in Table 3. The indirect effects further show that trust contributes to surveillance-AI openness through urban-AI attitudes, suggesting that institutional confidence operates through evaluations of visible, service-oriented applications rather than directly normalising surveillance. Urban-AI attitudes also increased perceived benefits of AI in local government. Although urban-AI and surveillance-AI are conceptually linked in real-world implementation, such as traffic monitoring, crime prediction, and facial recognition, modelling them separately revealed distinct attitudinal pathways. These findings suggest that the public may evaluate AI through connected experiences across urban service domains, where positive perceptions of visible and service-oriented applications can shape openness to more sensitive AI uses.

However, this pathway should not be interpreted as simple or unconditional acceptance of surveillance-oriented AI. Surveillance-AI remains normatively distinct from routine urban-AI because it raises heightened concerns about privacy, discrimination, institutional power, unequal visibility, and accountability. Critical sociotechnical perspectives on surveillance show that algorithmic monitoring can normalise intrusive forms of visibility and reshape relationships between public institutions and communities (Lyon, 2007; Browne, 2015; Eubanks, 2018). Therefore, public openness to surveillance-AI should be understood as conditional and governance-dependent, rather than as acceptance driven only by familiarity, usefulness, or perceived benefits.

This gateway effect raises important governance considerations. The gradual normalisation of AI through visible urban services may reduce public resistance to more intrusive applications, shifting the boundaries of acceptable governance practices. While such pathways can support innovation diffusion, they also risk depoliticising decisions about surveillance, data governance, and civil liberties if not accompanied by transparent deliberation. Local government AI acceptance, therefore, is not only an experiential process but also an institutional and political one, shaped by how cities sequence, frame, and legitimise technological interventions.

From a policy standpoint, these findings support experience-first strategies, but only when paired with strong safeguards. Governments seeking to introduce ethically sensitive AI should first build positive experiences with urban-AI services that demonstrate clear public value, while simultaneously ensuring transparency, accountability and public justification. Ethical oversight and continuous public dialogue remain essential to ensure that familiarity does not replace critical awareness, particularly in applications that involve monitoring, identification, or law-enforcement-related decisions.

Across all three samples, perceived risk emerged as a strong driver of responsible-AI prioritisation. This indicates that public calls for responsible-AI are shaped by caution, particularly the demand for transparency, accountability, and safeguards against potential harms in cities where algorithmic systems increasingly influence services, mobility, safety, and administrative interactions. Concerns about misuse and unintended consequences therefore remain a central motivator for stronger ethical governance.

At the same time, positive attitudes toward urban-AI also significantly influenced responsible-AI prioritisation. This suggests that when AI is perceived as useful and beneficial in tangible local applications, the public also expects it to be governed responsibly. Demand for responsible-AI is therefore not driven solely by fear or concerns about potential risks. It is also linked to sustaining beneficial innovation: residents may support technologies that improve public services while expecting them to be fair, safe, secure, transparent, and aligned with societal values. Responsible-AI expectations are therefore both precautionary and aspirational.

The relative weight of these drivers varied across contexts. In the Australian sample, risk was the stronger driver, reflecting expectations of accountability, institutional checks, and safeguards. In the Saudi Arabian sample, positive urban-AI attitudes had a stronger association with responsible-AI prioritisation, suggesting that public expectations for responsibility may be linked to the perceived promise of AI-enabled service improvement and national digital transformation. In Hong Kong, both trust-mediated pathways and risk sensitivity were important, indicating that residents may evaluate AI through both practical service experience and governance concerns.

Trust also produced divergent indirect effects on responsible-AI prioritisation. In Australia, trust indirectly reduced responsible-AI prioritisation through its negative association with risk perception, whereas in Hong Kong and Saudi Arabia, trust indirectly reinforced responsible-AI prioritisation. This pattern supports the broader interpretation that trust does not always reduce caution. In some contexts, particularly where respondents are younger, more educated, or more familiar with digitally mediated public services, trust may coexist with stronger expectations for safeguards and responsible implementation. These differences should be interpreted cautiously, as they may reflect sample composition, governance context, and the pathways through which trust shapes perceived risk and urban-AI attitudes.

Overall, these patterns show that responsible-AI prioritisation is shaped by the interaction of risk sensitivity, trust formation, and application-specific attitudes. This moves the discussion beyond a simple adoption model. Public support for responsible-AI is not merely a question of whether residents accept or reject innovation; it reflects how they negotiate the relationship between public value, perceived risk, and institutional responsibility. Effective AI governance must therefore respond to both sides of public concern: addressing fears of misuse while reinforcing confidence in AI's fair, safe, and beneficial deployment (Kim and Kwon, 2024; Bullock et al., 2025).

Before interpreting the cross-context patterns, it is important to emphasise that these comparisons reflect structural relationships observed within the collected samples, rather than representative portraits of national public opinion. This caution is necessary because full scalar invariance was not established and because the sampling strategies and demographic compositions differed across contexts. Therefore, Australia, Hong Kong, and Saudi Arabia are compared in terms of observed relationships among experience, understanding, trust, risk perception, urban-AI attitudes, and responsible-AI priorities, not as direct claims about absolute differences in wider public attitudes.

Although baseline perception patterns were broadly consistent across the three samples, the findings reveal nuanced differences in how local government AI is interpreted and prioritised across urban governance contexts. These differences are shaped by governance traditions, digital maturity, institutional trust, and demographic composition, which together influence how risks, benefits, and responsible-AI expectations are formed.

In the Australian sample, public perceptions reflect cautious optimism shaped by democratic norms and strong expectations of accountability. The influence of perceived risks in driving responsible-AI prioritisation underscores a governance setting where accountability and institutional checks are highly valued (Noetel et al., 2024). This risk-sensitivity does not preclude support for AI; rather, it suggests a pragmatic stance in which benefits are recognised but expected to be justified against potential downsides. Trust also exerted an important influence on both benefit and risk perceptions, indicating that confidence in local government AI remains closely tied to expectations of responsible implementation.

In the Hong Kong sample, perceptions were closely tied to experience, understanding, and trust. This may reflect a dense and digitally mature urban environment where residents are more likely to encounter smart urban systems in everyday life. Positive experiences and familiarity with AI enhanced trust, which in turn encouraged recognition of benefits. However, this pattern should not be interpreted simply as institutional deference. Instead, the findings suggest that trust may be shaped through lived interactions with technology, where residents evaluate AI according to its visible usefulness, reliability, and governance implications.

In the Saudi Arabian sample, institutional trust and positive urban-AI attitudes played a particularly important role in shaping AI acceptance and responsible-AI prioritisation. However, these findings should be interpreted with caution given the younger, more female-skewed, and highly educated profile of the sample. Within this sample, AI may be associated with progress, efficiency, and national digital transformation, particularly in visible applications such as traffic management and urban service delivery. Exposure to government-led AI initiatives, together with trust in state capacity within a centralised governance context aligned with national digital transformation agendas, may strengthen positive attitudes and support for responsible-AI (Harshan, 2025; Khayyat, 2025). The comparatively strong positive urban-AI pathway suggests that responsible-AI expectations can be driven not only by risk sensitivity but also by a desire to sustain beneficial innovation under conditions of rapid digital transformation.

Across the three samples, these findings show that public acceptance of responsible-AI is neither uniform nor automatic. Instead, responsible-AI priorities are shaped by the interaction between institutional trust, spatial exposure, perceived risk, and application-specific attitudes. The divergent pathways also show that governance contexts can shape whether trust functions mainly as risk reduction, informed scrutiny, or support for innovation. For local governments, this means that responsible-AI strategies should be sensitive to the specific public expectations, demographic profiles, and governance conditions of each context, rather than assuming that a single model of AI acceptance will apply across all urban settings. Recognising how AI is seen, encountered, and debated in the city fundamentally shapes its legitimacy.

The findings offer practical guidance for local governments seeking to implement AI in socially acceptable and responsible ways. Rather than presenting AI adoption as a broad technical modernisation agenda, local governments should consider how residents build understanding, trust, and risk awareness through visible urban service applications. Applications such as traffic management, service scheduling, infrastructure maintenance, or administrative support can function as experience-building entry points because they provide tangible public value. However, this experience-first pathway should not be used to normalise more intrusive technologies without additional safeguards. Extensions to surveillance-AI or high-stakes automated decision support should be accompanied by clear public justification, transparency notices, privacy safeguards, human oversight, and opportunities for community feedback. Table 5 translates the study's main perception pathways into practical governance implications for local governments.

The cross-context findings also suggest that responsible-AI strategies should be adapted to the perception patterns observed in each sample. In the Australian sample, where risk sensitivity strongly shaped responsible-AI prioritisation, local governments should emphasise independent oversight, contestability, privacy protection, and transparent explanation of how AI systems affect residents. In the Saudi Arabian sample, where positive urban-AI attitudes strongly contributed to responsible-AI prioritisation, responsible-AI communication should connect AI-enabled service improvements with expectations for fairness, reliability, security, and accountability, rather than framing governance only around risk avoidance. In the Hong Kong sample, where experience and understanding were closely tied to trust, public demonstrations, service-specific AI explanations, and interactive communication tools may help residents evaluate both the benefits and governance conditions of municipal AI.

For research, the findings highlight the need to move beyond treating experience as a single uniform construct. Future studies should distinguish passive exposure to AI-enabled services, such as encountering automated traffic systems, from active interaction, such as using a chatbot, digital permit system, or service platform. These forms of experience may operate through different psychological mechanisms: passive exposure may build familiarity and normalisation, while active interaction may shape perceived understanding, trust, satisfaction, and concern more directly. Longitudinal and mixed-method research could further examine how these different forms of experience influence trust, perceived risk, and responsible-AI expectations over time. In this way, Table 5 closes the loop between the urban-AI gateway pathway and practical governance by showing how visible, service-oriented AI applications can build familiarity and trust, while requiring stronger safeguards before extending to surveillance-AI or other high-stakes uses.

This study examined public perceptions of AI in local government, focusing on how experience, understanding, trust, perceived risks, perceived benefits, and application-specific attitudes shape the prioritisation of responsible-AI. Drawing on survey data from Australia, Hong Kong, and Saudi Arabia and using structural equation modelling, the study provides comparative insights into the relational pathways through which residents form supportive and critical evaluations of local government AI.

The findings show that direct experience with local government AI enhances perceived understanding, which in turn strengthens trust. This underscores the value of experiential familiarity and comprehension as foundations for cultivating public confidence in AI-enabled services. Trust then emerged as a central anchor, shaping perceived benefits, risk perceptions, and favourable attitudes toward urban-AI applications. However, trust does not uniformly reduce perceived risk. In some contexts, trust appears to coexist with heightened risk awareness, suggesting a more conditional form of public confidence in which residents recognise the value of AI while still expecting safeguards.

The study also identifies that positive attitude toward visible, service-oriented urban-AI applications were associated with greater openness to surveillance-AI. This suggests that routine municipal AI applications can shape broader perceptions of more sensitive technologies. However, such openness should not be interpreted as unconditional acceptance. Extensions to surveillance-AI or high-stakes automated decision support require stronger transparency, privacy protection, human oversight, contestability, and public justification.

Responsible-AI prioritisation was shaped by both caution and aspiration. Perceived risk remained a strong driver, indicating that residents value responsible-AI as a safeguard against misuse, bias, opacity, and unintended harms. At the same time, positive attitudes toward urban-AI also contributed to responsible-AI prioritisation, showing that residents may demand responsible governance not only because they fear risks, but also because they want beneficial innovation to be fair, reliable, secure, and socially legitimate. This finding contributes to responsible-AI and public sector innovation literature by showing that public support for responsible-AI is not reducible to either resistance or acceptance; it is shaped by the interaction between perceived public value, trust, risk awareness, and governance expectations.

The cross-context results further show that public perceptions of local government AI are not uniform. Australia reflected stronger risk sensitivity and accountability expectations, Hong Kong highlighted the importance of experience, understanding, and trust formation, and Saudi Arabia showed stronger links between positive urban-AI attitudes and responsible-AI prioritisation. These differences should be interpreted as relational patterns within the collected samples rather than nationally representative comparisons. Nevertheless, they demonstrate that responsible-AI strategies cannot rely on a one-size-fits-all approach. Local governments need context-sensitive governance that reflects public expectations, demographic characteristics, digital maturity, and local institutional conditions.

The study makes three main contributions. First, it brings local government into sharper focus within AI governance research, addressing an area often overlooked compared to national or global perspectives. It shows that municipal contexts are important practical settings where AI is encountered, evaluated, and legitimised through everyday public services. Second, it develops and tests a relational model linking experience, understanding, trust, risk, application-specific attitudes, and responsible-AI priorities. Third, it distinguishes routine urban-AI from more contested surveillance-AI, showing how visible service-oriented applications may shape openness to more sensitive municipal AI uses while also requiring stronger governance safeguards. In doing so, the study demonstrates the value of comparative, relational analysis across different urban governance contexts.

For practice, the findings suggest that responsible-AI implementation should begin with transparent and visible service applications that allow residents to observe public value and develop understanding. However, experience-first strategies must be accompanied by clear safeguards, especially when AI moves toward surveillance, identification, or high-stakes decision support. Local governments should therefore combine service improvement with public explanation, privacy protection, human oversight, contestability, and opportunities for community feedback. Responsible-AI governance should be treated as an urban governance process, not only a technical modernisation agenda.

This study has several limitations that provide directions for future research. The reliance on self-reported survey data means that the findings capture perceived rather than observed interactions with local government AI. Sample representativeness also remains uneven across contexts, and the focus on three countries limits generalisability to other cultural, political, and regional settings. While SEM provides useful insights into structural relationships, it cannot fully capture the discursive, emotional, or narrative dimensions of how residents reason about AI in public services. Future research could extend the model to additional contexts, combine longitudinal and qualitative methods, and examine how municipal policies, participatory practices, and different forms of AI exposure shape public acceptance and responsible-AI expectations.

Overall, the study shows that responsible-AI in local government depends on more than technical capability. Public perceptions are shaped by lived experience, institutional trust, perceived risk, application context, and expectations of responsible governance. Embedding AI responsibly in local government therefore requires sustained, context-aware engagement strategies that align technological innovation with citizens' expectations and local institutional realities. By foregrounding these dynamics, the study contributes to a more grounded and socially attuned understanding of responsible-AI in urban governance.

Data collection, analysis, and writing (original draft): R.M.; Conceptualisation, supervision, and writing (review and editing): T.Y.; Writing (review and editing): R.L., R.M., A.P. All authors have read and agreed to the published version of the manuscript.

This research was conducted in accordance with the ethical principles and guidelines applicable to research involving human participants and data. Ethical approval for the study was obtained from the Queensland University of Technology Human Research Ethics Committee under Ethics Approval No. LR 2026–8381–27647.

The authors thank the editor and anonymous referees for their invaluable comments on an earlier version of the manuscript.