Development and psychometric analysis of 5-item Artificial Intelligence Attitude (AIAT-5) for students Authors: Guogang Xin, Shengqin Yang, Xiaoyu Hou, Ruobing Wang, Shuaishuai Mi, Jipeng Wang, Bing Shen Published: 2026 ABSTRACT AbstractIn the context of accelerating integration of artificial intelligence (AI) technologies into education, rapidly assessing students’ AI attitude is crucial for advancing digital transformation in education.The aim of this study is to develop and validate a new five-item version of the Artificial Intelligence Attitude Scale (AIAT-5). For examining the quality of the scale, a total sample of 1576 middle and high school students (study 1: 858; study 2: 614) from diverse regions in China completed the survey. Results: (1) Study1: AIAT-5 exhibited strong internal consistency (Cronbach’s α = 0.923, McDonald’s ω = 0.923, Guttman’s λ6 = 0.911) and good discrimination (α = 3.201 ~ 4.513). Both EFA and EGA supported a robust unidimensional structure. (2) Study 2: CFA results indicated excellent model fit for unidimensional structure (CFI = 0.993, TLI = 0.976, and RMSEA = 0.028), and all factor loadings ranged from 0.879 to 0.953. Furthermore, AI attitude was positively correlated with knowledge of generative AI, life satisfaction, internet addiction, and epistemic curiosity. Overall, within the specific population of Chinese middle and high school students, the AIAT‑5 demonstrates initial validation as a reliable and practically efficient tool for large‑scale assessment of student AI attitude.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40359-026-05020-4.Keywords: Artificial intelligence, Attitude, Student, Exploratory graph analysis, IRT, Scale development Introduction Artificial intelligence (AI) technology is rapidly permeating the field of education, becoming an important tool for promoting personalized learning and optimizing teaching. Data show that 45.1% of Chinese adolescents have used AI products [4], and 86% of university students worldwide apply AI in their studies [13]. As the main participants and beneficiaries of education, students’ AI attitude directly influences the effectiveness of its application in education [8]. Therefore, studying students’ AI attitude in educational contexts is of key importance for promoting the in-depth integration of AI in education and advancing educational development. AI attitude is conceptualized as an evaluative orientation toward AI-based technologies [33]. Previous theoretical frameworks generally distinguish attitudes from two related constructs: (a) beliefs about expected outcomes and (b) downstream intention and behavior. For example, in the Technology Acceptance Model (TAM), beliefs such as perceived usefulness and ease of use shape attitudes toward using a system, which in turn influence behavioral intention and ultimately actual use [11]. Similarly, in the Theory of Planned Behavior (TPB), beliefs toward a behavior predict intention, and intention is one of the most proximal determinants of behavior [1, 5]. Consistent with this theoretical distinction, the present study defines AI attitude more specifically as attitudes toward using AI-based tools. It refers to overall evaluation of whether using AI is desirable and beneficial, as reflected in their beliefs about AI and their tendency to engage with AI. Based on this conceptualization, researchers have developed a variety of AI attitude scales. However, most of these instruments primarily focus on adult populations, particularly workers’ AI attitude in the workplace [16, 34, 36], and thus fail to adequately capture students’ unique experiences and psychological responses in educational contexts. For students, whose core goals include learning and personal growth, the use of AI is largely concentrated on educational assistance. Some scholars have therefore designed AI attitude scales tailored more closely to educational settings. A representative example is the Student Attitudes Toward Artificial Intelligence (SATAI) scale, which includes cognitive, behavioral, and affective dimensions, with a total of 26 items [37]. The scale has since been revised and validated in different cultural contexts, such as among university students in Turkey [12] and Greece [20]. However, because of its relatively large number of items, the SATAI may increase respondents’ cognitive burden and completion time, limiting its utility in large-scale surveys or in educational settings where students have restricted time. In contrast, the brief four-item Artificial Intelligence Attitude (AIAT-4) measure developed by Xia, Chiu, and colleagues uses concise wording while covering both learning and daily-life contexts [8, 42]. Its items combine students’ beliefs about the desirability of AI (e.g., “I think it would be very wise to use artificial intelligence in my learning/daily life”) with students’ behavioral engagement tendencies (e.g., “I use artificial intelligence in my learning/daily life”). Consistent with the logic of “beliefs shape attitudes” in TAM and TPB, these belief items operationalize students’ positive evaluative judgments about using AI as direct reflections of their attitudes. Meanwhile, the behavioral engagement tendency items capture the actual behavioral manifestations associated with AI attitudes. This interpretation is broadly consistent with theoretical frameworks such as TAM [11] and TPB [1, 5], which emphasize attitudes toward performing a behavior. Given this item content, we conceptualize the measure as capturing students’ global, acceptance-oriented attitude toward engaging with AI across learning and everyday contexts. Besides, due to its short length, it is therefore well suited for large-scale surveys, teaching evaluations, policy-oriented assessments, and classroom settings where time is limited. The AIAT-4 is grounded in Self-Determination Theory (SDT). SDT posits that individuals develop positive attitudes toward a technology and subsequently engage in actual usage behaviors when the technology satisfies their basic psychological needs for autonomy, competence, and relatedness [42]. When students perceive that using AI can enhance their learning efficiency or everyday convenience, they are more likely to view it as a wise choice, which in turn increases their behavioral intention to seek assistance from AI in learning and daily life. Nevertheless, although the AIAT-4 has been applied in prior research [42], its psychometric properties have not yet undergone rigorous validation. Given its favorable features, further examination of its reliability, validity, and other psychometric characteristics is warranted. However, most existing scales, including the SATAI and AIAT-4, were developed before large language models (LLMs) such as ChatGPT and DeepSeek became widely used [37, 42]. Prior to 2022, AI technologies were primarily based on deep learning or more traditional algorithms, which were typically designed to perform narrow tasks such as facial recognition, playing Go, or protein prediction [22]. By contrast, the emergence of LLMs has enabled AI to engage in open-ended dialogue, assist users in gathering and organizing information, support reasoning, and even aid in decision-making processes [44]. This conversational and knowledge-supporting role represents a shift in the nature of AI [28]. Consequently, existing instruments do not explicitly include items assessing students’ behavioral tendencies toward interacting with conversational AI systems (e.g., asking AI for information or seeking help). To address this gap, we extended the original four-item AIAT by adding one new item specifically focusing on students’ tendencies to ask questions to AI systems. This addition aims to capture behavioral inclinations related to student–AI interactions in conversational contexts. With the inclusion of this new item, the scale was expanded into a five-item version of the Artificial Intelligence Attitude (AIAT-5). In addition, this study applied Item Response Theory (IRT) for item analysis and employed Exploratory Graph Analysis (EGA) to further explore the dimensional structure of the new scale. While Classical Testing Theory (CTT) assesses a scale from a more integrative perspective by examining overall psychometric properties, such as reliability and factor structure, IRT focuses on rich statistics for the item discrimination, difficulty, information, and item-person fit [17, 23, 31]. These indicators are valuable for scale refinement, item selection, and the development of computerized adaptive testing [39]. Therefore, the present study integrates both CTT and IRT approaches to provide a comprehensive evaluation of the scale’s psychometric properties. EGA is an emerging approach for determining the number of dimensions in psychological data and has several advantages [15]: (1) Visualization: through network graphs with color-coded communities, the dimensional allocation of items can be identified intuitively; (2) Automatic estimation: it does not require consideration of rotation methods; and (3) Stability test: stability can be evaluated through resampling procedures. Unlike traditional factor analysis, EGA treats observed variables (items) as nodes in a network connected by edges, and the dimensional structure of a scale is identified through community detection algorithms that cluster nodes [10]. Simulation studies have shown that EGA outperforms factor analysis in accurately detecting the dimensionality of psychological scales [15]. Therefore, this method was applied to examine the dimensional structure of the scale, with the aim of providing deeper validation of its structural characteristics. To enable the rapid and accurate assessment of students’ AI attitude, this study extended the existing AIAT by adding an AI interaction item, thereby developing AIAT-5. The first study was designed to assess the psychometric properties of the newly developed scale and to uncover its underlying measurement structure. To further enhance the rigor of the psychometric evaluation, item parameters were estimated using IRT, and EGA was additionally employed to validate the structural characteristics of the scale. The second study aimed to examine the measurement structure and criterion-related validity of the scale. Confirmatory Factor Analysis (CFA) was conducted, and relevant criteria were selected for correlation analyses. In summary, this study developed a unidimensional scale that is concise and practical for large-scale administration, offering a reliable tool for assessing students’ attitudes toward AI-education integration. Methods Participants Data were collected from two junior high schools and three senior high schools in Zhejiang, Fujian, and Henan provinces. To ensure representativeness, several classes from each grade were randomly selected, and online questionnaires were administered collectively. Two samples were collected for Study 1 and Study 2. Sample 1 consisted of 858 participants (49.9% girls; Mage = 15.40, SDage = 1.25). Sample 2 consisted of 614 participants (46.9% girls; Mage = 16.64, SDage = 1.70). The study was approved by the Shandong Normal University’s Ethics Committee (Approval No. 20250305). The participants of questionnaire survey also reviewed the informed consent form in the Wenjuanxing system and indicated their agreement to participate by checking a “consent to participate” box. All data were anonymized and stored securely in password-protected databases, with no personally identifiable information linked to the responses. Since the data was collected in the school, informed consent was obtained from parents by the teachers. Measures Artificial intelligence attitude This scale was adapted from the AIAT, which was designed to measure students’ beliefs about AI and their tendency to engage with AI in both learning and daily life contexts [42]. To account for the widespread adoption of conversational AI, one additional item was added on students’ tendency to ask questions to AI systems, resulting in AIAT-5. The scale adopts a five-point Likert-type format (1 = strongly disagree, 5 = strongly agree). Total scores are calculated by summing all items, providing a quick assessment of students’ AI attitude. During the adaptation process, the original items were slightly reworded to better reflect the context of current AI tools and student usage scenarios. The revised items were reviewed by several experts in educational psychology and educational technology to evaluate their clarity, relevance, and content validity. Minor wording adjustments were made based on their feedback. In addition, the wording of the items was examined to ensure linguistic clarity and cultural appropriateness for the target student population. Criterion-related scale Research has shown that AI attitude are closely related to knowledge of generative AI, life satisfaction, internet use, and curiosity. Students’ understanding and familiarity with AI are positively associated with their positive attitudes toward using AI [6]. Individuals with a more positive AI attitude generally report higher levels of life satisfaction [32]. Moreover, those who frequently use the internet tend to have more opportunities to engage with AI. As a result, they are more likely to develop positive attitudes toward it. Additionally, individuals with higher curiosity are more willing to explore and adopt new technologies such as AI [2]. Based on this evidence, the present study selected these three constructs as criteria variables for validating the scale. Knowledge of generative AI Knowledge of generative AI was measured as one dimension of the generative AI perception scale [6]. The scale consists of 5 items rated on a 5-point Likert scale ranging from 1 (“strongly disagree”) to 5 (“strongly agree”), asking participants to indicate the extent to which they understand key limitations and characteristics of generative AI technologies (e.g., factual inaccuracies, contextual errors, biases, and limited emotional intelligence). Total scores were obtained by summing all items, with higher scores indicating a higher level of knowledge about generative AI. In this study, the Cronbach’s α of the scale was 0.936. Life satisfaction Students’ life satisfaction was measured using the life satisfaction scale [27]. A translated version of this scale demonstrated good reliability and validity [7]. The scale consists of 5 items rated on a 7-point Likert scale ranging from 1 (“strongly disagree”) to 7 (“strongly agree”), asking participants to indicate the extent to which each statement reflected their actual situation. Total scores were calculated by summing all items, with higher scores indicating greater overall life satisfaction. In this study, the Cronbach’s α of the scale was 0.944. Internet addiction Students’ internet addiction was assessed with the 8-item internet addiction scale [35]. A translated version of this scale has been widely used in China [43]. Each item was scored dichotomously (1 = “yes,” 0 = “no”), and the sum of the 8 items was calculated, with higher scores indicating greater levels of internet dependence. In this study, the Cronbach’s α of the scale was 0.919. Epistemic curiosity Students’ curiosity was measured using the epistemic curiosity scale [21]. The scale contains 10 items rated on a 5-point Likert scale (1 = “strongly disagree,” 5 = “strongly agree”). Total scores were computed by summing all items, with higher scores reflecting stronger curiosity traits. In this study, the Cronbach’s α of the scale was 0.973. Statistical analyses Item analysis, reliability testing, EFA, CFA, and correlation analysis were conducted using JASP 0.19 [19]. EGA was performed with the EGAnet package in R 4.4.3. IRT was performed with the mirt package in R 4.4.3. First, item quality and internal consistency were evaluated based on CTT. Item analysis was conducted using the extreme-groups t-test and the item–total correlation method [25]. In the extreme-groups t-test, the top 27% and bottom 27% of participants based on total scores were defined as the high- and low-scoring groups, respectively, and an independent-samples t-test was performed to compare the two groups. If the mean difference between the groups on a given item was not significant, the item was considered for deletion [25]. The item–total correlation method calculated the correlation coefficient between each item and the total score. Items with coefficients below 0.40 were removed [25]. Items were evaluated to ensure that they met the required discriminative standard. Following previous research [2, 32], three reliability indices were applied to assess the internal consistency of the new short scale: Cronbach’s α, McDonald’s ω, and Guttman’s λ6. In addition, changes in reliability indices were compared after item deletion. If removing an item substantially increased the reliability coefficient, it indicated that the item lowered the internal consistency of the scale or dimension and should be removed [18]. Second, item parameters and item information were estimated based on IRT [24]. The item parameters include item discrimination (α) and threshold parameters (β). The α parameter reflects the sensitivity of an item to respondents’ latent trait. In IRT, an item discrimination value of α < 0.35 indicates poor discrimination, 0.35–0.65 indicates fair, 0.65–1.34 good, 1.35–1.69 very good, and α ≥ 1.70 is considered excellent [3]. The β parameter reflects the point on the latent trait at which the probability of choosing a particular response category or higher reaches 50%, thereby describing the “threshold difficulty” between different response options. For Likert scales, the β parameters of each item should increase in succession, which indicates that the difficulties of the options align as they are designed. Item information refers to the amount of information each item provides when estimating respondents’ latent ability levels. The sum of all item information yields the test information (TIF), which reflects the overall precision of the test in estimating ability levels. A larger information value indicates greater measurement precision [38]. According to the formula that reliability = 1–1/TIF, for getting an acceptable reliability of 0.9, the TIF in the given range should be greater than 10 [14]. In addition, item characteristic curves (ICCs) were inspected as a visual supplement to these analyses, illustrating the relationship between latent trait levels and the probability of endorsing each response category. Furthermore, Yen’s Q3 statistic was used to examine local independence. A Q3 value below 0.20 indicates that the assumption of local independence is satisfied, suggesting that the items do not share substantial residual variance beyond the latent trait. Finally, the M2 statistic and associated global model fit indices were used to assess the unidimensionality of the scale, whereas the S-χ2 statistic was used to evaluate item fit to the model. Subsequently, both EFA and EGA were used to extract the dimensional structure of the scale. Prior to EFA, the Kaiser–Meyer–Olkin (KMO) test and Bartlett’s test of sphericity were performed, ensuring that KMO values exceeded 0.70 and Bartlett’s test was significant. Factor numbers were determined using parallel analysis, and principal axis factoring was adopted. EGA was conducted in R 4.4.3 using the EGAnet package. The glasso algorithm and the walktrap community detection algorithm were first used to identify the number of item clusters, i.e., the number of dimensions in the scale [9]. To evaluate the stability of the EGA results, bootstrap analysis with 1,000 resamples was applied. Key indicators included the frequency of detecting the same number of dimensions, the frequency of structural consistency within dimensions, and the frequency of categorization of items into their respective exclusive dimensions; all indices were required to exceed 0.75 [9]. Following prior research recommendations, expected influence was selected as the centrality measure [30]. It represents the sum of the strengths of signed associations between a node and all other nodes in the network, with higher values indicating greater centrality. Then, CFA was performed to test the structural validity of the scale. Model fit indices included the chi-square statistic with degrees of freedom, CFI, TLI, and RMSEA. For CFI and TLI, values above 0.90 indicate acceptable fit, and values above 0.95 indicate good fit. RMSEA values below 0.08 indicate acceptable fit, and values below 0.05 indicate good fit. Finally, based on empirical evidence from previous studies [2, 32], life satisfaction, internet addiction, and epistemic curiosity were selected as criterion-related variables. These variables were hypothesized to be positively correlated with AI attitude, and if the results were consistent with this assumption, the evidence would further support the criterion-related validity of the AIAT scale. Results Study 1 Study 1 employed item analysis and reliability test under both CTT and IRT to evaluate the quality of the scale items. EFA and EGA were further conducted to determine the measurement structure of the scale. CTT analysis Item analysis results are shown in Table 1, based on extreme group t-tests and item–total correlation analysis. All items showed significant differences between the high- and low-scoring groups (p < 0.001). The correlations between item scores and the total score ranged from 0.855 to 0.891, all reaching statistical significance (p < 0.001). These findings indicate that all items demonstrated good discriminative power, and no items needed to be removed. Table 1Item analysis results of AIAT-5Itemt ar b1.我认为在我的学习中使用人工智能是非常明智的 (I think it would be very wise to use artificial intelligence in my learning)−29.130***0.877***2.我认为在我的日常生活中使用人工智能是非常明智的 (I think it would be very wise to use artificial intelligence in my daily life)−29.790***0.891***3.我在做出决策时询问人工智能 (I consult artificial intelligence when making decisions)−26.776***0.855***4.我在学习中使用人工智能 (I use artificial intelligence in my learning)−28.805***0.880***5.我在日常生活中使用人工智能 (I use artificial intelligence in my daily life)−25.794***0.868***aThe score difference on this item between participants in the low-scoring group (bottom 27%) and the high-scoring group (top 27%)bThe correlation coefficient between the item score and the total questionnaire score* p <.05** p <.01*** p <.001 Item analysis results of AIAT-5 aThe score difference on this item between participants in the low-scoring group (bottom 27%) and the high-scoring group (top 27%) bThe correlation coefficient between the item score and the total questionnaire score * p <.05 ** p <.01 *** p <.001 The internal consistency of the scale was examined using three indices: Cronbach’s α, McDonald’s ω, and Guttman’s λ6, which were 0.923, 0.923, and 0.911, respectively. In addition, corrected reliability was calculated after deleting each item. As shown in Table 2, Cronbach’s α ranged from 0.900 to 0.911, McDonald’s ω ranged from 0.901 to 0.911, and Guttman’s λ6 ranged from 0.872 to 0.893 after the deletion of any single item. These results indicate that removing items would reduce the internal consistency of the scale; therefore, no items needed to be deleted. Table 2Internal consistency reliability of the scaleReliability coefficientNo item droppedlf item dropped12345Cronbach’s α0.9230.9040.9000.9110.9040.907McDonald’s ω0.9230.9050.9010.9110.9040.907Guttman’s λ60.9110.8780.8720.8930.8820.885 Internal consistency reliability of the scale IRT analyses Item discrimination parameter estimates range from 2.285 to 3.565 (> 1.7), as shown in Table 3, indicating that the items effectively distinguish respondents at different trait levels. The threshold difficulty parameters are widely distributed across the scale, approximately from –2 to + 2 in succession. Item information values range from 2.177 to 3.426, with a total test information value of 13.953 (> 10). The empirical reliability is 0.914, and the marginal reliability is 0.918. Table 3Item parameter estimates and item information of the scaleItemαβ1β2β3β4Average item information13.131−1.315−0.8030.3581.5093.06223.565−1.484−0.8700.1801.3403.42632.285−1.318−0.7560.4191.6432.17742.816−1.220−0.7780.2101.4972.93752.501−1.331−1.0070.0781.4342.351α represents the item discrimination parameter, and β represents the threshold parameters. Since this scale adopts a five-point response format, there are four threshold parameters in total. Specifically, β1 corresponds to the threshold between option 1 and option 2; β2 corresponds to the threshold between option 2 and option 3; β3 corresponds to the threshold between option 3 and option 4; and β4 corresponds to the threshold between option 4 and option 5 Item parameter estimates and item information of the scale α represents the item discrimination parameter, and β represents the threshold parameters. Since this scale adopts a five-point response format, there are four threshold parameters in total. Specifically, β1 corresponds to the threshold between option 1 and option 2; β2 corresponds to the threshold between option 2 and option 3; β3 corresponds to the threshold between option 3 and option 4; and β4 corresponds to the threshold between option 4 and option 5 An inspection of the ICCs revealed steeply rising trace lines, consistent with the high discrimination parameters. The categories demonstrated well-separated and ordered peaks, which suggested that the scoring categories are exclusive and respondents could meaningfully distinguish between adjacent options (see Fig. 1). In summary, these results indicate that the items demonstrate good discrimination and reliability, and the options demonstrate difficulties that align with they were designed. Category response curves for 5 response options (Ps) per item. P1 = probability of responding 1; P2 = probability of responding 2; P3 = probability of responding 3; P4 = probability of responding 4; P5 = probability of responding 5 To evaluate the assumption of local independence, Yen’s Q3 statistics were calculated for all item pairs (Supplement Material, Section 1, Table S1). All Q3 values were below the commonly accepted threshold of 0.2. The absence of significant positive residuals indicates that the high discrimination parameters observed are not the artifact of semantic overlapping. The unidimensionality assumption of the Generalized Partial Credit Model (GPCM) was evaluated using the M2 statistic and associated global model fit indices. The results yielded M2(5) = 76.725, p < 0.001, RMSEA = 0.129, SRMSR = 0.034, TLI = 0.966, CFI = 0.983. Although the M2 statistic was significant and the RMSEA value appeared relatively high, this may reflect inflation due to the limited degrees of freedom in short scales. The high comparative fit indices (CFI > 0.9 and TLI > 0.9) and low residuals (SRMSR < 0.05) nevertheless indicate an adequate overall model fit, supporting the unidimensionality assumption. Item-level fit was assessed using the S − χ2 statistic and RMSEA (Supplement Material, Section 1, Table S2). Although the p-values for all five items were significant (p < 0.001), this can be attributed to the high sensitivity of the χ2 test in a large sample. The RMSEA values suggest that the degree of misfit is modest. When considered alongside the acceptable global fit indices, these results suggest that the items remain within an interpretable range. Exploratory factor analysis First, the KMO test and Bartlett’s test of sphericity were conducted. The results showed that KMO = 0.874 (> 0.7), χ2 = 3158.343 (df = 10), p < 0.001, indicating that the data were suitable for factor analysis. Subsequently, principal axis factoring was performed to extract common factors. Parallel analysis suggested the extraction of a single factor, which explained 70.6% of the total variance. As shown in Table 4, the factor loadings of all items ranged from 0.806 to 0.873 (> 0.30), and the uniqueness values ranged from 0.238 to 0.350 (< 0.60). Table 4Factor loadings from the EFA of the scaleItemFactor loadingsUniqueness10.8480.28120.8730.23830.8060.35040.8450.28650.8290.313Eigenvalues3.824Explained variance ratio0.706 Factor loadings from the EFA of the scale Exploratory graph analysis The results of the EGA are presented in Fig. 2. On the left side of the figure, the item network graph shows nodes AI1 to AI5, representing items 1 through 5 of the scale, with node colors indicating their dimensional membership. The thickness of the edges between nodes reflects the strength of associations, with thicker edges indicating stronger associations; green edges represent positive associations. Except for the relatively weaker connections between Item 1 and Item 5, and between Item 2 and Item 4, the remaining items were closely connected, and Item 3 showed relatively balanced associations with all other items. All items were displayed in red, indicating that the scale exhibited a unidimensional structure. Next, a bootstrap procedure was conducted to further examine the stability of this unidimensional structure. The results showed that both the detection frequency of the same number of dimensions and the frequency of structural consistency within dimensions were 1 (> 0.75). In addition, as shown on the right side of Fig. 1, the item stability analysis revealed that the consistency frequency of each item being classified into a single dimension was also 1 (> 0.75), suggesting good item stability. Taken together, the bootstrap sampling results of the EGA provide strong evidence that the scale possesses a stable unidimensional structure. Graphical network and item stability from the EGA of the scale The expected influence indices for the AI attitude network are depicted in Fig. 3. AI2 had the highest expected influence, indicating that this node exerted the most influence on the overall network. Estimates of expected influence indices for the items. AI1: I think it would be very wise to use artificial intelligence in my learning. AI2: I think it would be very wise to use artificial intelligence in my daily life. AI3: I consult artificial intelligence when making decisions. AI4: I use artificial intelligence in my learning. AI5: I use artificial intelligence in my daily life Study 2 Study 2 employed CFA to further examine the measurement structure of the scale. Criterion-related validity was assessed by analyzing the correlations between AI attitude and life satisfaction, internet use behavior, and epistemic curiosity. Confirmatory factor analysis Following previous methodological recommendations [26, 29], the Likert-type items were treated as ordinal variables and the CFA was conducted using the Weighted Least Square Mean and Variance adjusted (WLSMV) estimator. The model fit indices were χ2 = 4.48 (df = 3, p = 0.214), CFI = 0.993, TLI = 0.976, and RMSEA = 0.028 (90% CI [0.000, 0.079]). The RMSEA value and the comparative fit indices suggest a good model fit. As shown in Fig. 4, the factor loadings of all items ranged from 0.879 to 0.953, indicating that the items were strongly associated with the latent factor. Overall, these results support the single-factor model. The CFA model for the scale Criterion-related validity Based on previous research [2, 32], life satisfaction, internet addiction, and epistemic curiosity were selected as criterion-related variables and examined in correlation analyses with students’ AI attitude. The results showed that AI attitude was significantly correlated with knowledge of generative AI (r = 0.276, p < 0.001), life satisfaction (r = 0.311, p < 0.001), internet addiction (r = 0.189, p < 0.001), and epistemic curiosity (r = 0.430, p < 0.001). These findings provide preliminary evidence of criterion-related validity. Discussion This study conducted a psychometric validation of AIAT-5. The scale not only offers the advantage of brevity (consisting of five items) but also strikes a balance between students’ cognitive evaluations and behavioral tendencies toward AI, as well as between its use in different contexts (i.e., learning and daily life). To account for the widespread adoption of conversational AI, an additional item on students’ tendency to ask questions to AI systems was included, with the resulting AIAT-5 demonstrating good reliability and validity. Both the CTT- and IRT-based item and reliability analyses indicate that the items are of good quality and the test demonstrates high reliability. Structural validity analyses, including EFA, EGA, and CFA, yielded consistent results, all supporting a unidimensional model. The introduction of EGA not only supplemented traditional validation methods but also provided visual support for the structure. Regarding criterion validity, AI attitude was found to be significantly positively correlated with life satisfaction, internet use behavior, and epistemic curiosity. In summary, the scale demonstrated good reliability and validity, providing a reliable tool for understanding students’ AI attitude. From the perspective of item analysis and reliability, AIAT-5 demonstrated strong psychometric properties across multiple indicators. First, extreme group t-tests and item–total correlation analysis both showed that all items effectively distinguished individuals with different levels of attitudes. In addition, Cronbach’s α, McDonald’s ω, and Guttman’s λ6 all reached high standards, indicating excellent internal consistency [42]. Furthermore, deleting any item would reduce reliability, confirming that each item is an essential component and that the scale maintained robust measurement performance even after the addition of the new item. The high quality of the scale may be attributed to the focused design and clear wording of its items, which helped avoid issues of content overlap [16]. Based on IRT, it was likewise found that all items showed high discrimination values, suggesting that they were sensitive to differences in respondents’ levels of AI attitude. Meanwhile, the overall test information reached 13.953, with empirical reliability and marginal reliability both at high levels, indicating that the test can provide precise measurement of students’ AI attitude. In terms of structural validity, three methods—EFA, EGA, and CFA—were jointly applied, and all results supported a unidimensional structure of the scale. This multi-method cross-validation strategy enhanced the robustness of the structural validity assessment. In the CFA, the factor loadings were relatively high. Since the items of AIAT-5 cover learning use, daily-life use, decision-making, and beliefs about whether AI use is wise, the construct represented by this scale is relatively broad. Therefore, the high loadings are more likely to reflect a coherent latent construct rather than simple item redundancy. Among them, EGA not only confirmed the stability of the unidimensional structure through bootstrap resampling but also provided a visual representation of the inter-item relationships [15]. Most items were closely and positively associated, except for the relatively weaker connections between Item 1 and Item 5 and between Item 2 and Item 4. These differences may reflect context-specific distinctions in AI usage as well as a potential separation between attitudes and behaviors. The findings are consistent with previous studies that have employed EGA as a complementary tool for factor validation [40, 41], further supporting the value of EGA as an important method for examining structural validity, particularly in its ability to intuitively reveal patterns of relationships among items [9]. Interestingly, AI2 had the highest expected influence, meaning it is the most central node in the network. This suggests that attitudes toward using AI in daily life play a key role in shaping the overall attitude network. Although the AIAT items cover slightly different aspects of attitudes toward AI, including evaluative judgments and behavioral inclinations, they collectively reflect a broader underlying orientation toward AI technologies. In attitude research, such global evaluations are often treated as manifestations of a single latent construct representing individuals’ overall stance toward an object [1, 5, 11]. Therefore, the AIAT conceptualizes AI attitude as a unidimensional evaluative disposition toward AI technologies. Nevertheless, given that attitudes may involve multiple related components, the possibility that different facets (e.g., beliefs or behavioral tendencies) may partially diverge cannot be entirely ruled out. This conceptual consideration should be acknowledged when interpreting the scale’s unidimensional structure. The study found that students’ AI attitude was significantly and positively associated with knowledge of generative AI, life satisfaction, internet addiction, and epistemic curiosity, providing preliminary evidence for the criterion-related validity of the scale. First, a significant correlation between AI attitude and knowledge of generative AI indicates that individuals who understand AI better tend to hold more positive views toward its use. Knowledge may reduce uncertainty and increase confidence, thereby fostering more favorable attitudes [6]. Second, the positive correlation between AI attitude and life satisfaction suggests that individuals with more favorable attitudes are more likely to experience higher levels of life satisfaction, consistent with previous findings [32]. This may be because students perceive AI technologies as enhancing efficiency and convenience in daily life, thereby strengthening positive evaluations of life. Third, the association between AI attitude and internet addiction suggests a potential link between AI engagement and broader online behavior patterns [32]. It may reflect a more complex relationship between technology engagement and problematic internet use. On the one hand, a positive AI attitude may indicate greater engagement with digital technologies and AI-related environments. On the other hand, it may also be associated with broader patterns of technology use that warrant attention. However, the observed correlation was relatively weak. Therefore, this finding should be interpreted cautiously. Finally, the positive correlation between AI attitude and epistemic curiosity demonstrates that highly curious individuals are more willing to explore the unknown and show greater acceptance of and support for emerging technologies. This finding reinforces the close link between technology use and individual traits [2, 32]. Overall, students’ AI attitude reflects not only their perspectives on new technologies but also their broader life satisfaction, internet use patterns, and personal characteristics. These results provide an empirical foundation for exploring the psychological mechanisms and behavioral outcomes of AI attitude, as well as for informing interventions and applications. In summary, the findings provide empirical support for AIAT-5 in terms of item discrimination, reliability, and validity. The strong psychometric properties of the scale may be attributed to the design of items that closely align with students’ contexts of AI use, thereby enabling accurate assessment [37]. Nevertheless, several limitations remain. First, the sample consisted exclusively of middle and high school students from the central and eastern regions of China. As a result, the generalizability of the findings may be constrained by cultural context and developmental stage. Future studies should conduct cross-cultural validation to further evaluate the scale’s broader applicability. Additionally, measurement invariance testing across gender, school level, and region is recommended. Second, the present study did not include measures of a negative attitude toward AI. Future research is encouraged to incorporate such assessments to build a more comprehensive framework. Third, the focus on students excluded other key groups such as teachers and parents. Subsequent studies should broaden the scope of participants to gain a more complete understanding of how different stakeholders perceive the integration of AI in education. Finally, test–retest reliability evidence is absent. This is a notable limitation, as students’ attitudes toward AI may change rapidly while they gain more experience with generative AI tools. Conclusion AIAT-5 demonstrated good item discrimination and test reliability. Evidence from EFA, EGA, and CFA confirmed stable structural validity, supporting its unidimensional construct. Students’ AI attitude scores were significantly and positively correlated with knowledge of generative AI, life satisfaction, internet addiction, and epistemic curiosity. Overall, within Chinese middle and high school students, the scale shows initial promise as an efficient tool for large-scale AI attitude assessment. Supplementary Information Supplementary Material 1. Authors’ contributions Conceptualization, Guogang Xin, and Jipeng Wang; methodology, Guogang Xin, Ruobing Wang, and Xiaoyu Hou; formal analysis, Guogang Xin and Jipeng Wang; resources, Ruobing Wang; writing—original draft preparation, Guogang Xin, Bing Shen, and Ruobing Wang; writing—review and editing, Guogang Xin, Xiaoyu Hou, Shengqin Yang, Shuaishuai Mi, and Jipeng Wang; project administration, Guogang Xin and Jipeng Wang. All authors have read and agreed to the published version of the manuscript. Funding No Funding. Data availability The data that were used in the present study can be provided by the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate The study was approved by the Shandong Normal University’s Ethics Committee (Approval No. 20250305). The study was performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and itsd following updates. After being informed about the survey’s aims, all participants gave informed consent and joined the study voluntarily. For adolescents, informed consent was obtained from their parents/legally authorized representatives. Relevant guidelines and regulations carried out all methods. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Guogang Xin and Shengqin Yang contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The data that were used in the present study can be provided by the corresponding author upon reasonable request.