Teachers’ perceptions of artificial intelligence integration in education: a multi-campus cross-sectional study from Riphah International University, Pakistan

The integration of artificial intelligence (AI) into education is accelerating worldwide, yet evidence on teachers’ perceptions remains scarce in the Global South, particularly in resource-constrained Pakistan. This cross-sectional study examined teachers’ perceptions of AI integration at Riphah International University, Pakistan. A 14-item Likert-scale questionnaire was administered to 270 in-service teachers (86.7% male, 13.3% female) across BS Undergraduate through PhD levels. Guided by a proposed Dual-Lens conceptual framework treating perceived benefits and risks as coexisting dimensions, the instrument assessed pedagogical advantages and structural concerns of AI. Descriptive analysis revealed a moderately favorable overall perception (composite M = 3.40, SD = 0.55). Teachers rated interactivity enhancement (M = 3.81), student enjoyment (M = 3.73), and assessment support (M = 3.72) most favorably, while AI replacing human teachers received the lowest endorsement (M = 2.96). An independent-samples t-test showed no significant gender difference (t(268) = −0.245, p =.807, d = −0.044). One-way ANOVA revealed significant effects for age (F(4, 265) = 3.200, p =.014, η 2 =.046) and level of instruction (F(3, 266) = 3.471, p =.017, η 2 =.038), with teachers under 29 and BS Undergraduate instructors holding less favorable perceptions. Qualification and experience showed no significant association. A moderate positive correlation emerged between perceived benefits and risks (r =.537, p <.001), indicating co-existence rather than opposition. These findings provide baseline evidence from a Global South private university with multiple campuses, highlighting the need for targeted teacher training and policies to mitigate the risks of AI integration in similar institutional environments in developing countries amid their development gaps.

The use of artificial intelligence (AI) in professional and public domains has accelerated rapidly, exerting new pressure on educational systems worldwide to ready students for a society influenced by AI (Holmes et al., 2019; Zawacki-Richter et al., 2019). Beyond delivering content knowledge, schools are now expected to be more responsible for developing computational thinking skills, digital literacy and adaptive problem-solving abilities that are needed in the AI economy (Luckin et al., 2016; Ng et al., 2021). However, this aspiration exceeds the current readiness of most educational systems, particularly in the Global South, where shortcomings in infrastructure, teacher training, and policy inertia often lead to a gap between technology use and classroom practices (Pedro et al., 2019; Selwyn, 2019; UNESCO, 2024).

Teachers are the key mediators between educational policy and classroom practice, making the understanding of teachers’ perceptions of AI integration critical. The beliefs, attitudes, and self-efficacy they hold regarding technology greatly influence their willingness to use, and therefore their successful adoption, of innovations (Davis, 1989; Ertmer and Ottenbreit-Leftwich, 2010). According to the Technology Acceptance Model (TAM), the intention to adopt technology is determined by perceived usefulness and perceived ease of use (Davis, 1989; Scherer et al., 2019), and these factors are mediated by other variables in AI-specific contexts, such as AI anxiety, ethical awareness, and perceived threats to professional identity (Celik, 2023). It is consistently shown that even the most well-resourced technology initiatives fall short when teachers feel that the innovation is threatening, irrelevant, and inconsistent with their values around pedagogy (Knezek and Christensen, 2016; Tondeur et al., 2017). With respect to AI, teacher perceptions are given extra weight as AI raises qualitatively different concerns that prior waves of educational technology did not: algorithmic opacity, teacher displacement fears, ethical issues regarding data privacy, and whether AI tools can support or hinder equitable access to learning (Borenstein and Howard, 2021; Holmes et al., 2022). The challenges are further exacerbated in developing areas where digital infrastructure is still weak, and most teachers lack formal experiences with AI in pre-service or in-service education (Sperling et al., 2024; UNESCO, 2024).

Empirical studies on teachers’ attitudes toward AI in education have been gradually increasing. Research on AI in high-tech environments finds generally positive attitudes towards using AI for personalized learning, formative assessment, and administrative tasks, but also ongoing concerns regarding job loss and a lack of preparation (An et al., 2023; Celik et al., 2022). Chiu et al. (2024) revealed that higher self-efficacy and technological pedagogical content knowledge about AI predicted higher teacher willingness to use AI in the classroom. Celik (2023) found that AI literacy and outcome expectations were important in determining teachers’ intention to use AI tools. Sperling et al. (2024) argue that AI literacy is not only about technical proficiency but also about pedagogical and ethical abilities. In a similar vein, Muthukrishnan et al. (2024) found that prospective teachers’ AI literacy was closely linked to responsible AI use, underscoring the importance of embedding ethical awareness alongside technical competence in teacher education. Pekkarinen et al. (2025) similarly found in their study that the usefulness of AI tools was a significant predictor of Estonian teachers’ willingness to use AI. Moreover, studies based on surveys in East Asian and European contexts have shown that there are significant gaps in the perception of AI based on factors like gender, teaching experience, educational qualification, and level of instruction taught (Ayanwale et al., 2022; Galindo-Domínguez et al., 2024). There is something that emerges in all these studies: teachers recognise the potential of AI in pedagogical practice but also have some concerns regarding socio-professional implications of AI. However, recent challenges due to resource scarcity and limited professional development have led to greater teacher resistance to implementing new practices (Crompton and Burke, 2023; Jamil and Muschert, 2023).

However, there are three key gaps in the expanding evidence base. Firstly, most of the existing literature is regionally concentrated in high and upper-middle income countries, with limited empirical research originating from South Asia and Pakistan specifically (Pedro et al., 2019). The socio-cultural, economic, and institutional conditions in which teachers’ perceptions develop in Pakistan are different from those in Seoul, Riga, or Abu Dhabi, and outcomes from those conditions should not be presumed to be applicable. Secondly, few studies have investigated both benefits and risks, and almost none have investigated both benefits and risks in the same sample and using the same instrument. Even as educators recognize the potential of AI in the classroom, there is a potential for conflicting emotions regarding its impact, given the diverse landscape. Third, the voices of teachers at different levels of education (undergraduate to doctoral level) are rarely heard in one study, which restricts comparison of perceptions along the whole range of university education.

These gaps highlight the need for contextually grounded empirical research on teachers’ perceptions in under-researched educational settings, using instruments that capture both the perceived opportunities and threats of AI across multiple dimensions of teaching and learning. The present study is situated within Riphah International University (RIU), a federally chartered private university established in 2002.

The university has campuses in different areas (Islamabad, Rawalpindi, Lahore) and regional locations (Malakand, Faisalabad, etc.). It provides a combination of digital infrastructure, socio-economic contexts, and geographic environments in one institutional setting.

This study presents a Dual-Lens conceptual framework of Teacher AI Perceptions (Figure 1) in which perceived benefits and perceived risks are treated as two distinct constructs, endorsed by the same teacher along a spectrum rather than at opposite extremes. The structure is not intended as an empirically tested structural model, but as a way of organizing the inquiry. This framework is based on the TAM (Davis, 1989), the multidimensional belief systems perspective (Ertmer and Ottenbreit-Leftwich, 2010), and Holmes et al.’s (2022) ethical framework for AI in education. This framework positions five demographic variables; gender, age, educational qualification, teaching experience, and level of instruction — as factors potentially associated with teachers’ overall AI perceptions.

These perceptions consist of two parallel sub-dimensions: the perceived benefits (including pedagogical utility, interactivity, assessment support, inclusivity, and student engagement) and the perceived risks (such as digital divide, compulsory instruction, teacher displacement, and instructional level restriction).

This study aims to achieve three objectives: (1) to assess teachers’ overall perceptions of AI integration in teaching and assessment at Riphah International University; (2) to examine whether demographic variables are associated with variations in teachers’ perceptions of AI's role in these processes; and (3) to explore the connection between teachers’ perceived pedagogical benefits and perceived structural risks associated with AI integration at the university. This research contributes to the literature by offering among the first quantitative datasets on teachers’ perceptions of AI integration within a multi-campus private university network in Pakistan.

It employed a Dual-Lens conceptual framework and provided actionable baseline data to guide institutional policies, teacher training, and curriculum reforms, supporting responsible AI integration in Pakistani educational institutions.

Research design

This study adopted a quantitative, cross-sectional survey design. A cross-sectional approach was chosen, as the main goal was to capture a point-in-time assessment of teachers’ perceptions, not to see changes over time (Creswell and Creswell, 2018). The design aligns with most of the teacher perception studies in the AI-in-education literature that used self-report questionnaires with a specific population over a limited time of data gathering (Celik et al., 2022; Galindo-Dominguez et al., 2024).

Participants and setting

The target population comprised in-service teachers affiliated with the Riphah International University, Pakistan. Riphah International University is a federally chartered private university established in 2002, accredited by the Higher Education Commission (HEC) of Pakistan and certified with ISO 9001:2015 QMS. The university has several campuses in 6 cities in Pakistan and a diverse student and faculty population. Convenience sampling was used for recruitment of the participants. To be eligible, respondents had to be current teachers. 270 valid responses were gathered from February to April 2026.

Respondents were spread across seven campuses: I-14 Campus (Main), Islamabad (n = 58, 21.5%); Raiwind Campus, Lahore (n = 45, 16.7%); Al-Mizan Campus, Rawalpindi (n = 42, 15.6%); Faisalabad Campus (n = 35, 13.0%); Malakand Campus, KPK (n = 32, 11.9%); Gulberg Green Campus, Islamabad (n = 30, 11.1%); and Peshawar Campus (n = 28, 10.4%). The sampling strategy is of several campuses and represents the Riphah system in different education environments.

Out of 270 participants, 234 were males (86.7%), and 36 were females (13.3%). About age, 124 (45.9%) were under 29 years old, 57 (21.1%) were aged 29 to 36, 36 (13.3%) were aged 36 to 43, 44 (16.3%) were aged 43 to 50, and 9 (3.3%) were aged 50 or above. In terms of education, 255 (94.4%) subjects were MPhil and 15 (5.6%) subjects were PhD holders. As to teaching experience, 142 (52.6%) participants had 2 years or less teaching experience, 44 (16.3%) had 3–5 years of teaching experience, 53 (19.6%) had 6–15 years of teaching experience and 31 (11.5%) had 16 years or more of teaching experience. The teachers were spread between the different stages of university teaching; 94 (34.8%) were teaching BS (Undergraduate) students, 104 (38.5%) were teaching MS (Master) students, 35 (13.0%) were teaching MPhil students and 37 (13.7%) were teaching PhD students. A summary of the sample is given by their demographic characteristics in Table 1.

Instrument

The data collection tool was a structured, self-administered questionnaire consisting of 2 sections. The questionnaire included 14 substantive items that were distilled from themes found in the literature on AI in education, such as the pedagogical benefits of interaction (Holmes et al., 2019), support for assessment (Luckin et al., 2016; and Crompton and Burke, 2023), inclusive learning, and student engagement, alongside the structural risks of digital divide, teacher displacement, and compulsory implementation (Borenstein and Howard, 2021; Selwyn, 2019; Holmes et al., 2022). Items were crafted to cover these dimensions, with two content experts in each area reviewing them for content appropriateness, clarity, and the balance between benefit and risk statements before administering. A limitation of this study is that no pilot study was conducted. Section A recorded five demographic variables, gender (male and female), age (five categories: less than 29, 29–36, 36–43, 43–50, and 50 and above), highest educational qualification (two categories coded as MPhil or PhD), years of teaching experience (7 categories combined into four groups) and level of instruction currently taught (four categories: BS Undergraduate, MS Master's, MPhil, and PhD). The second section, B, comprised 14 substantive items that assessed teachers’ perceptions of the integration of AI in classrooms with a 5-point Likert scale (Strongly Disagree = 1, Disagree = 2, Neutral = 3, Agree = 4, and Strongly Agree = 5). The 14 items explored the following aspects: usefulness of AI in teaching and assessment (Items 1 and 2); usefulness of the integration of AI in the curriculum to enhance the interactivity of learning experiences (Items 3 and 4); usefulness of the integration of AI to facilitate interdisciplinary and multidisciplinary teaching (Item 5); need for time and flexibility to develop deeper understanding (Item 6); students’ enjoyment of learning with AI (Item 7); opposition to compulsory learning with AI (Item 8); opposition to using AI in lower grades (Item 9); usefulness of AI for inclusive learning (Item 10); usefulness of AI to reduce drop out levels and improve students’ performance (Item 11); increase of the digital divide due to AI integration (Item 12); concern that AI-powered mechanical teachers will replace human teachers (Item 13); and usefulness of AI in real-life contexts (Item 14). Items 8, 9, 12 and 13 were negatively worded and reflected perceived risk and structural issues. These items were coded using two coding schemes, depending on the purpose of analysis. These four items were reverse-coded (6 – original score) for the overall composite score, with higher composite scores indicating more positive perceptions of AI integration. The original coding (not reversed) was used for the perceived risks subscale in the benefit–risk correlation analysis, with higher scores indicating greater perceived risk. In this way, the correlation supported the Dual-Lens framework's hypothesis that teachers who endorsed more pedagogical benefits also endorsed more structural risks. Descriptive statistics for individual items are presented in the original coding to maintain the item's intuitive meaning.

Data collection procedure

The questionnaire was carried out online using a structured digital questionnaire from 1 February to 15 April 2026. The link of the survey was shared with the teachers through institutional communication at 7 campuses of Riphah International University. There was no mandatory participation and at the outset, participants were informed that the responses they provided would be confidential and anonymous. No specific identification data were gathered other than the demographic data mentioned above. Before the respondents could go to the survey items, informed consent was obtained digitally. Average completion time was about 8–10 min.

Data analysis

Data were analyzed statistically using R version 4.4.2 (R Core Team, 2024). Analysis was done in five steps. First, descriptive statistics (frequencies, percentages, means, and standard deviations) for all 14 Likert scale items and demographic variables were computed. Second, an independent-samples t-test was conducted to investigate gender differences in perceptions of AI integration, as the gender variable was dichotomous. Third, one-way analysis of variance (ANOVA) was used to compare mean perception score differences across the multi-category demographic variables of age group, educational qualification, teaching experience, and level of instruction. Tukey's Honestly Significant Difference (HSD) post hoc test was employed to identify group differences following a significant ANOVA result. All inferential testing was conducted at the 0.05 significance level. Parametric tests were preceded by the assumptions of normality and homogeneity of variance. Normality was assessed using the Shapiro–Wilk test and by examining skewness and kurtosis values, and homogeneity of variance was assessed using Levene's test. When assumptions were not met, results for robust alternatives (Welch t-test) are presented in addition to the standard tests. Fourth, Pearson's product-moment correlation analysis was performed to analyze the relationship between each of the 14 items to explore the correlation between teachers’ perceived pedagogical advantages and perceived structural risks of AI integration. In addition to significance testing, effect sizes were calculated for both t-tests (Cohen's d) and ANOVAs (partial eta-squared) (Cohen, 1988). Fifth, an exploratory factor analysis (EFA) was performed to analyze the dimensionality of the 14-item instrument. In the pre-extraction stage, sampling adequacy was assessed using the Kaiser-Meyer-Olkin (KMO) measure and Bartlett's test of sphericity. The principal axis method was used to extract factors, and the number of factors to be retained was determined by the Kaiser criterion, which selects factors with eigenvalues greater than 1. To simplify understanding of the factor structure, Varimax rotation was applied.

Ethical considerations

This study was conducted in accordance with the principles of the Declaration of Helsinki. The necessary institutional ethics approval from Riphah International University was obtained before data collection. Participants consented to take the survey before taking part. Participation was voluntary, and respondents were informed that they could drop out at any time without any repercussions. The data collected did not include personally identifiable information, and all responses were kept in a password-protected online space accessible only to the research team.

Instrument reliability

To determine internal consistency of the survey instrument, Cronbach's alpha was calculated prior to conducting the main analyses. The total 14-item scale (Items 8, 9, 12, and 13 were reverse-coded) achieved an alpha of.733, which is above the generally accepted level of.70 for exploratory studies (Nunnally, 1978). The perceived benefits subscale (alpha =.934) consisted of 10 items (Items 1, 2, 3, 4, 5, 6, 7, 10, 11, and 14), and the perceived risks subscale (alpha =.763) consisted of 4 items (Items 8, 9, 12, and 13, reverse-coded). These values suggest that both subscales are reliable, with values in the acceptable to excellent range for purposes of this analysis.

Exploratory factor analysis

An exploratory factor analysis (EFA) was conducted on the 14 items, including Items 8, 9, 12, and 13 (which were reverse coded), to evaluate the instrument's dimensionality. The Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy was.922, which is considered excellent for factor analysis (Kaiser, 1974). The result of Bartlett's test of sphericity was significant, chi-squared(91) = 2386.1, p <.001, indicating that the correlation matrix was not an identity matrix.

In the principal axis extraction, four factors were identified, with two factors having eigenvalues over 1.0 (7.22 and 1.69). These two factors explained 63.6% of the total variance. A two-factor solution was also suggested by the scree plot. Following a varimax rotation, the factor structure was found to be closely aligned with the proposed Dual-Lens conceptual framework. The ten perceived benefit items loaded heavily on factor 1, which explained 42.1% of the variance: AI enhanced teaching interactivity (item 4;.868), AI aids assessment (item 2;.841), AI supports teaching (item 1;.839), AI increased teaching interactivity (item 3;.832), time efficiency and adaptability (item 6;.830), interdisciplinary teaching (item 5;.773), fostering an inclusive learning environment (item 10;.758), students enjoy learning with AI (item 7;.716), reduces dropout rates (item 11;.477), and real-world applicability (item 14;.469). The four perceived risk items (reverse-coded) loaded onto Factor 2, which accounted for 21.5% of the variance: Item 13 (teacher replacement; -.821), Item 12 (digital divide; -.746), Item 8 (AI not compulsory; -.691), and Item 9 (higher classes only; -.614).

Two items exhibited cross-loadings: Item 11 loaded on both Factor 1 (.477) and Factor 2 (.588), and Item 14 loaded on Factor 1 (.469) and Factor 2 (.457). Overall, the pattern does offer preliminary structural support for the conceptualization of perceived benefits and perceived risks as separable aspects of teachers’ AI perceptions, despite these cross-loadings. The cross-loading items suggest that the two dimensions are not completely pure, however, so further studies are needed to purify this instrument using an independent sample and confirmatory factor analysis.

Descriptive statistics for perception items

The means, standard deviations, and response distributions of the 14 survey items in their original coding are shown in Table 2.

The items with the top 3 mean scores were Item 3 (AI integration made teaching more interactive; M = 3.81, SD = 1.12), Item 7 (students enjoy learning about AI; M = 3.73, SD = 1.18), and Item 2 (AI helps in assessment and evaluation; M = 3.72, SD = 1.19). These scores show that, in general, teachers were satisfied that AI has pedagogical benefits. The lowest mean score (M = 2.96, SD = 1.33) was achieved on Item 13 (mechanical teachers will replace human teachers), which is the only one of the 14 items to score below the scale midpoint of 3.00, indicating that teachers were divided on this item, with slightly more disagreeing (35.6%) than agreeing (35.2%) that AI will replace human teachers. There was some disagreement about the scope and mandate for AI instruction, as seen in items 8 and 9 being moderately agreed upon (AI should not be compulsory; M = 3.31, SD = 1.23; AI should be restricted to higher classes; M = 3.33, SD = 1.26). AI integration enhancing the digital divide (Item 12, Mean = 3.54, SD = 1.03) was answered higher than the midpoint, meaning that there are concerns that AI integration can widen the digital divide. A moderately favorable overall perception of AI integration was indicated by an overall mean of 3.40 (SD = 0.55) for the 14 items on the composite mean on the reverse-coded scale. The mean score for each of the 14 items is shown in Figure 2.

Gender differences in AI perceptions

Table 3 summarizes the inferential analyses across all five demographic variables. Levene's test prior to analysis showed that the homogeneity of variance assumption was not violated, with F(1, 268) = 3.14 and p =.078. Subsequently, an independent-samples t-test compared the composite AI perception scores between male and female teachers. The results revealed no statistically significant difference between male teachers (M = 3.40, SD = 0.53) and female teachers (M = 3.42, SD = 0.67), t(268) = −0.245, p =.807, Cohen's d = −0.044. A Welch's t -test was also calculated as it is a robustness check due to unequal group sizes (234 males vs. 36 females) and it also yielded substantively identical results, t = −0.206, p =.838. The effect size was close to zero, indicating that there was no meaningful association between gender and teachers’ overall AI perceptions in this sample, although the unequal group sizes (234 males and 36 females) resulted in limited statistical power to detect small differences between groups.

Differences by age group

One-way ANOVA was used to compare the composite AI perception scores for five age groups. There was a significant violation of the homogeneity of variance assumption, as indicated by Levene's test F(4, 265) = 2.79, p =.027, suggesting unequal variances across age groups. However, regular ANOVA is presented, as it is known to be not overly sensitive to moderate deviations at these sample sizes. The omnibus test was statistically significant, F(4, 265) = 3.200, p =.014, partial eta-squared =.046, indicating a small but significant relationship between age and perceptions of AI. Because of the Levene's test result, caution is warranted when interpreting pairwise differences, especially comparisons involving the smallest group (age 50+, n = 9). Teachers aged 29 to 36 reported the highest composite scores (M = 3.56, SD = 0.54), followed by those aged 43 to 50 (M = 3.52, SD = 0.37), while those under 29 reported the lowest (M = 3.29, SD = 0.57). Post hoc pairwise comparisons revealed that teachers under 29 held significantly less favorable perceptions than teachers aged 29 to 36 (mean difference = −0.26, p =.004) and teachers aged 43 to 50 (mean difference = −0.23, p =.013). A significant difference was also observed between teachers aged 43 to 50 and those aged 50 and above (mean difference = 0.28, p =.036). No other pairwise comparisons reached significance.

Differences by educational qualification

Levene's test indicated equal variances, F(1, 268) = 0.15, p =.700. An independent-samples t-test comparing composite scores for MPhil holders (M = 3.39, SD = 0.55) and PhD holders (M = 3.44, SD = 0.54) found no significant difference, t(268) = −0.632, p =.528, Cohen's d = −0.092. Similarly, Welch's t -test produced a nearly identical result, t = −0.641, p =.523. The near-zero effect size suggests no meaningful association between educational qualification and AI perceptions in this sample; however, this finding should be interpreted cautiously given the unequal group sizes.

Differences by teaching experience

A one-way ANOVA examining differences across four experience categories produced a non-significant result, F (3, 266) = 1.709, p =.165, partial eta-squared =.019. Teachers with 16 or more years of experience reported the highest composite score (M = 3.61, SD = 0.46), while those with two years or fewer reported the lowest (M = 3.36, SD = 0.58). Although the direction of this trend suggests that more experienced teachers may hold somewhat more favorable perceptions, the difference did not reach statistical significance.

Differences by level of instruction

The Levene's test did not show any significant violation of the homogeneity of variance assumption, F (3, 266) = 0.33, p =.802, for the level of instruction. Results of a one-way ANOVA revealed a statistically significant difference between composite scores for the four levels of instruction, F(3, 266) = 3.471, p =.017, partial eta-squared =.038. Teachers instructing MPhil students reported the highest composite score (M = 3.57, SD = 0.52), followed by those instructing PhD students (M = 3.53, SD = 0.57), MS Master's students (M = 3.41, SD = 0.54), and BS Undergraduate students (M = 3.27, SD = 0.55). Post hoc analyses indicated that teachers instructing BS Undergraduate students held significantly less favorable perceptions than those instructing MPhil students (mean difference = −0.29, p =.007) and PhD students (mean difference = −0.25, p =.021). No other pairwise comparisons reached significance. This finding suggests that teachers working with advanced-level students tend to perceive AI integration more favorably than those working with undergraduate students. Figure 3 visualizes the composite scores across the four levels of instruction.

Relationship between perceived benefits and perceived risks

A Pearson product-moment correlation analyzed the relationship between the perceived benefits composite (Items 1–7, 10, 11, and 14; M = 3.67, SD = 0.90) and the perceived risks composite (Items 8, 9, 12, and 13; M = 3.29, SD = 0.93). Original scoring (not reversed) was applied to the perceived risks composite in this analysis, with higher scores indicating higher perceived risk (see Section 2.3 for the coding rationale). The Pearson correlation between the two scores was r =.537, p <.001, which is considered moderate. This was confirmed by a Spearman rank order correlation, with rho =.489, p <.001. This empirical finding supports the Dual-Lens framework presented in this study (Figure 1), demonstrating that perceived benefits and risks can coexist within the same individuals rather than being mutually exclusive. But more validation studies via CFA and replication studies are required before we can conclude that the framework is empirically supported.

In terms of items, among the highest inter-correlations for the benefit dimension were between Item 1 (AI helps in teaching) and Item 2 (AI helps in assessment), r =.740, p <.001, and Item 7 (students enjoy learning AI) and Item 10 (inclusive learning environment), r =.653, p <.001. The highest correlation in the risk dimension was between Item 8 (AI should not be compulsory) and Item 12 (AI increases the digital divide), r =.539, p <.001. Interestingly, the correlation across dimensions between Item 1 (benefit: AI helps in teaching) and Item 13 (risk: mechanical teacher replacement) was weak but significant, r =.167, p =.006, indicating that while some might believe that AI can be used to benefit teaching and learning, some also worry about that possibility leading to the replacement of teachers.

The investigation was guided by a Dual-Lens conceptual framework, viewing perceived benefits and perceived risks as co-existing dimensions, and had the following three objectives: to examine the overall levels of perceived benefits and perceived risks; to examine whether the five demographic variables were related to variations in the overall levels of perceived benefits and perceived risks; and to examine the relationship between perceived pedagogical benefits and perceived structural risks. Each objective's findings are presented and discussed below in relation to the existing literature.

Overall perceptions of AI integration

The teachers at Riphah International University had a score of 3.40 (SD = 0.55) on the reverse-coded scale, which represents a moderately positive attitude towards AI integration in education. This finding warrants attention for two reasons. First, it is lower than the higher composite means reported in studies in more technologically advanced places, which suggests that the limited exposure to AI and the lack of resources across the Riphah campus network dampen the optimism seen in other places (An et al., 2023; Galindo-Dominguez et al., 2024). Secondly, the score is still above the scale's midpoint, indicating that even in a context that lacks extensive digital infrastructure and any formal AI training programs, teachers perceive a certain potential of AI for teaching and learning. This is consistent with Davis’ (1989) TAM that assumed that attitudes can be positive even if the user had no direct experience with the technology, because of perceived usefulness. Kowang et al. (2025) similarly identified perceived usefulness and ease of use as primary determinants of AI acceptance among undergraduates in a Malaysian higher education context, reinforcing the cross-cultural applicability of TAM-based explanations.

The three top-scoring items are consistent with the broader literature, which ranks these pedagogical benefits as the most readily apparent (Crompton and Burke, 2023; Holmes et al., 2019; Luckin et al., 2016): three aspects of interactivity enhancement, student enjoyment, and assessment support. The single item below the midpoint was mechanical teachers replacing human teachers (Item 13, M = 2.96), consistent with international studies that indicated that teachers are not fully ready to replace themselves with AI (Pekkarinen et al., 2025; Selwyn, 2019). However, 35.2% agreed or strongly agreed with this statement, suggesting that whilst not all people in the sample were concerned about displacement, a significant minority were.

A notable point is the above midpoint score on Item 12 (AI integration will enhance the digital divide; M = 3.54). In fact, over half (53.3%) agree or strongly agree that AI might worsen the gap instead of narrowing it. This is an important point to note in the Riphah campus network as there is a disparity in the digital infrastructure of metropolitan and regional campuses (Jamil and Muschert, 2023; Pedro et al., 2019). This finding indicates a heightened awareness of structural dangers of AI among teachers in contexts with limited resources compared to teachers in contexts with more resources, where its presence is less obvious in everyday teaching and learning. This aligns with Borenstein and Howard’s (2021) argument that the ethical risks of AI are not shared equally across communities but rather fall disproportionately on communities that are already facing socio-economic vulnerabilities.

Demographic correlates of AI perceptions

Among the five demographic variables examined, only age and level of instruction showed statistically significant associations with composite AI perception scores, while gender, educational qualification, and teaching experience did not. No significant difference between the genders was found (t(268) = −0.245, p =.807, d = −0.044), which aligns with the results of several recent studies that did not find significant gender differences in perceptions of AI (Ayanwale et al., 2022; Galindo-Domínguez et al., 2024; Pekkarinen et al., 2025). This is significant given the gendered nature of the teaching profession in the current context where the sample consisted of 86.7% males. The near-zero effect size suggests that male and female teachers held similar views on AI integration, although the unequal group sizes limit statistical power to detect small differences.

The significant age-group difference (F(4, 265) = 3.200, p =.014, partial eta-squared =.046) indicated that teachers in the youngest age group held significantly less favorable perceptions than those aged 29 to 36 and those aged 43 to 50.

This finding runs counter to the assumption that younger teachers, as “digital natives,” would be more receptive to new technologies (Ng et al., 2021; Prensky, 2001). A possible explanation is that the under-29 cohort in this sample is primarily made up of early-career teachers (52.6% of the full sample had two years or less of teaching experience), who may feel overwhelmed with the demands of classroom practice and have less opportunity to see how an unfamiliar technology could be integrated. Conversely, teachers aged 29 to 36 and 43 to 50 may have already gained a certain level of teaching confidence, allowing them to engage with the idea of AI in the classroom without fear. This interpretation is supported by the direction (if not the statistical significance) of the experience effect: teachers with 16 years or more experience reported the highest composite score (M = 3.61), suggesting that maturity in the profession may contribute to open-mindedness to innovations.

Composite scores varied significantly by level of instruction, F(3, 266) = 3.471, p =.017, partial eta-squared =.038. Teachers at the BS Undergraduate level (M = 3.27) reported less favorable perceptions than those at the MPhil level (M = 3.57, p =.007) and PhD level (M = 3.53, p =.021).

This is in line with the nature of the university curriculum in which the concepts of AI, computational thinking and research-based digital literacy are more likely to be taught at higher levels (MPhil and PhD), while the BS Undergraduate level is more likely to prioritize knowledge of the discipline. For teachers at the more advanced levels, they might experience more direct contact with the potential use of AI in research and advanced teaching, which could mean that they feel AI is more relevant and useful (Su et al., 2022). Teachers, on the other hand, working with undergraduate students might not see how AI could apply to teaching at the BS level, which is more basic and general.

The non-significant effects of educational qualification (t (268) = −0.632, p =.528) and teaching experience (F (3, 266) = 1.709, p =.165) deserve brief comment. The small effect sizes (d = −0.092 and partial eta-squared =.019, respectively) suggest that neither MPhil nor PhD qualifications nor longer careers were meaningfully associated with perceptions of AI in this context. One reason is that AI literacy has not yet been integrated into the systematic teacher education curriculum in Pakistan, so formal qualifications might not accurately reflect AI readiness in this context (Muthukrishnan et al., 2024; Sperling et al., 2024).

Co-existence of perceived benefits and perceived risks

The key conceptual finding of this study is the moderate positive correlation between the perceived benefits and perceived risks composites (r =.537, p <.001). This finding is consistent with the Dual-Lens conceptual framework outlined in this study (Figure 1) and challenges the notion that benefit and risk are opposing extremes of a single continuum, as posited in simple unidimensional models. Rather, those teachers who agreed more with the pedagogical benefits of AI (increased interaction, assessment, and an inclusive classroom) also agreed more with the potential risks of AI (digital divide, compulsory instruction, and teacher replacements). This suggests that partial familiarity with AI — knowing enough to recognize its promise while remaining aware of its uncertainties — may simultaneously produce both optimism and caution.

The dual-awareness pattern is vital to the grasping of theory. This is consistent with Ertmer and Ottenbreit-Leftwich’s (2010) belief model, which is that teachers’ attitudes towards technology are multifaceted, at times contradictory. It also aligns with the conclusion that an ethical and community-focused approach to AI in education is necessary, as AI has the potential to either exacerbate or alleviate educational inequalities (Holmes et al., 2022). Additional structural evidence for this interpretation comes from the exploratory factor analysis reported in Section 3.2, which revealed a two-factor solution with varimax rotation, with the two factors accounting for 63.6% of the total variance, with a largely clear separation between benefit items (loading on Factor 1) and risk items (loading on Factor 2), although Items 11 and 14 cross-loaded on both factors. Additionally, the present study contributes to the TAM (Davis, 1989; Scherer et al., 2019) by demonstrating that perceived usefulness and perceived threat are not mutually exclusive; rather, the higher the cognitive engagement with the technology, the greater the perceived usefulness and perceived threat.

Limitations

There are some limitations that must be acknowledged. First, the cross-sectional design shows perceptions at one specific time and cannot show causal relationships between demographic variables and perceptions of AI. Second, the findings of this study cannot be generalized to the wider population of university teachers in Pakistan due to the use of convenience sampling. Third, the sample was mostly male (86.7%), which, while representing the local teaching population, limits the statistical power of gender-based comparisons. Overall, there was also a difference in the sizes of cells between some of the demographic groups (gender: 234 vs. 36; educational qualification: 255 vs. 15; oldest age group: n = 9), so the non-significant results for these comparisons must be interpreted with caution and the power to detect small effects might be limited. Hence, caution should be exercised on the non-significant results of gender, educational qualification and teaching experience in interpreting that there was no association between them. Fourth, the exploratory factor analysis presented in Section 3.2 suggests that the instrument may have two factors, but it has not been confirmed by confirmatory factor analysis (CFA) with an independent sample. Items 11 and 14 were cross loaded on both factors, indicating that there might not be a clear separation between perceived benefits and perceived risks. Future studies should use a larger and independent sample for CFA to verify the dimensional structure of the instrument. Fifth, self-report survey data may be prone to social desirability bias, with the possibility that positive perceptions of AI may have been over-reported. Finally, Items 3 and 4 are identical in wording (“AI integration in curriculum/another subject made teaching more interactive”), and this may have led to increased inter-item correlation and response-set effects. The relationship between these two items was high (r =.72), which may reflect a methodological artifact due to similar wording rather than a substantively different construct. In future versions of this instrument, one of these items should be changed to a uniquely worded version so that it does not create this confound.

Implications for policy and practice

The results have several implications for action. Moderately favorable overall perceptions indicate that teachers in the Riphah campus network do not seem resistant to the integration of AI but need targeted support to move from openness to practice. The leadership of the universities and other education departments involved should focus on creating AI literacy initiatives that build on teachers’ current knowledge, because not all teachers are entirely unfamiliar or hostile to AI. Second, it was found that teachers instructing BS Undergraduate students and the youngest teachers had less positive perceptions, which suggests that professional development should be differentiated based on teacher career stage and level of instruction. Training could help teachers of undergraduate programs understand how AI can enhance foundational teaching and learning at the BS level. Third, the high level of concern around the digital divide (Item 12) suggests that AI integration policies must go hand in hand with investments in digital infrastructure; otherwise, AI integration can exacerbate the inequalities it is meant to overcome, as a lack of internet access and device ownership are barriers to AI adoption. Finally, the dual nature of the perceived benefits and risks (as illustrated in the Dual-Lens conceptual framework) provides a rationale for a critical approach to AI training, in which teachers are prepared to appraise AI tools in the light of pedagogical, ethical, and contextual considerations (UNESCO, 2024).

This study provides one of the first multi-campus quantitative investigations of teacher AI perceptions from a private-sector university in Pakistan, which is underrepresented in the global AI-in-education literature. A key conceptual contribution is the introduction of the Dual-Lens framework of Teacher AI Perceptions, demonstrating that positive and negative perceptions of AI can exist simultaneously within the same individuals, rather than cancel each other out. The results presented here in a preliminary correlational context are in line with this framework, but formal structural validation using confirmatory factor analysis and replication across different samples are needed to consider this framework empirically established. Despite this, the pattern challenges one-dimensional assumptions and suggests a more nuanced view of teachers’ interactions with AI integration.

The evidence shows that teachers at Riphah International University generally perceive AI positively, acknowledging its potential for education while also being aware of its possible structural challenges. Age and level of instruction were associated with variation in perceptions, suggesting that professional development initiatives should be differentiated by career stage and instructional level rather than designed as one-size-fits-all interventions.

Overall, the study indicates that policies for the adoption of AI in educational institutions in Pakistan should shift from a mere enthusiastic embrace of AI to an informed approach that accounts for the technology's impact on the learning landscape and infrastructure gaps, especially in multi-campus educational institutions in metropolitan and regional areas. An empirical question of whether these findings can be generalized to public-sector universities and other institutional settings in Pakistan needs to be explored. Although the structural validity and applicability to other institutional and cultural settings would be worthy of further empirical study, the Dual-Lens conceptual approach may prove to be a suitable organizing lens for future research and policy design. Longitudinal studies employing validated multi-dimensional instruments and probability-based sampling would strengthen the evidence base and help track how teacher perceptions evolve as AI tools become increasingly embedded in Pakistani university settings.