New 22-item scale measures how confident teachers really feel about using AI in the classroom

Artificial intelligence has moved from the research laboratory into the everyday toolkit of the classroom, but a fundamental question has gone largely unanswered: do teachers actually believe they can use it well? A research team from Istanbul 29 Mayıs University, Istanbul University-Cerrahpaşa and independent scholarship in Türkiye has now built and validated the first theory-grounded instrument designed specifically to measure teachers’ confidence in their own AI capabilities. The Artificial Intelligence Self-Efficacy Scale for Teachers, or AISES-T, published in Frontiers in Education, distills a 50-item pool into a 22-item questionnaire that captures three distinct but related strands of belief: learning with AI, teaching with AI, and understanding how AI works.

The distinction the researchers draw is subtle but consequential. Within Social Cognitive Theory, self-efficacy is not knowledge, not demonstrated competence, not a favorable attitude and not an intention to act. It is a person’s judgment of their own capability to organize and execute the actions required for a specific performance. A teacher might know a great deal about machine learning, or successfully produce an AI-generated lesson plan when asked, yet still privately doubt their ability to repeat that performance under real classroom conditions. Conversely, a teacher with limited factual AI knowledge may nonetheless believe firmly that they can evaluate AI outputs or integrate AI tools into instruction. The AISES-T deliberately measures the belief, not the skill, which is why its items are phrased as capability statements rather than test questions.

This conceptual separation matters because earlier measurement traditions blurred it. Educational technology research long relied on Computer Self-Efficacy scales, later adapted for the internet and even for robots in healthcare settings. The authors argue that transplanting these instruments to AI is theoretically flawed: traditional software executes deterministic, rule-based commands, whereas AI systems learn autonomously, rationalize their outputs and interact in anthropomorphic ways, acting as what researchers describe as a dynamic cognitive extender. Existing AI-related scales reflect this ambiguity. The general-user AISES developed by Wang and Chuang emphasizes assistance, comfort and anthropomorphic interaction; Technology Acceptance Model questionnaires capture readiness, attitudes and intentions; and even recent teacher-focused instruments such as the TAICS scale, with its six Delphi-derived dimensions, were tested by confirmatory analysis in a single sample without an independent exploratory stage. None, the team concluded, captured teachers’ confidence across their own learning, pedagogical integration and critical understanding in a single empirically validated instrument.

Construction of the new scale followed a rigorous multi-stage sequence. The researchers first reviewed the literature on AI self-efficacy, AI literacy and teacher competencies, anchoring the items in the UNESCO AI Competency Framework for Teachers and the European DigCompEdu framework. From this groundwork they wrote 50 candidate items, assigned in advance to three hypothesized dimensions. A panel of ten independent experts, including specialists in educational technology, measurement, curriculum and instruction, Turkish linguistics and two practicing teachers with classroom AI experience, then rated every item on relevance, clarity, representativeness and linguistic appropriateness. Items were retained only if at least 80 percent of experts rated them highly relevant. Twenty-three items fell below this threshold and were cut, leaving a 27-item trial form with a scale-level content validity index of 0.97.

The trial form was administered to 313 teachers working in schools across different regions of Türkiye, a sample exceeding the conventional benchmark of five to ten respondents per item. Exploratory factor analysis using principal axis factoring with oblique Promax rotation revealed a strikingly clean structure. Five further items failed to reach the retention criterion of a pattern coefficient of at least 0.60 on any factor and were removed, even though some had earned perfect expert ratings. The remaining 22 items settled into three factors explaining a cumulative 68.51 percent of the variance, with the first factor alone accounting for 54.32 percent. Velicer’s minimum average partial test, eigenvalues, the scree plot and the conceptual coherence of the loading pattern all converged on the same three-factor solution, and no item showed problematic cross-loading.

The three dimensions that emerged map directly onto distinct classes of action. AI Use and Learning Self-Efficacy, with eleven items, covers teachers’ belief that they can learn, operate and adapt AI applications for problem solving, research, self-regulated learning and time management. Pedagogical AI Integration Self-Efficacy, with seven items, concerns translating AI capabilities into instructional decisions: designing lessons, building assessment tools, monitoring student progress, supporting student-centered learning and facilitating learning for students with diverse needs. AI Awareness and Understanding Self-Efficacy, with four items, captures confidence in understanding how AI applications work, explaining their purposes, judging the reliability of AI-generated information and keeping up with developments in education. Crucially, these items assess perceived capability to perform cognitive and evaluative tasks, not factual AI knowledge itself.

Confirmatory factor analysis in a second, entirely independent sample of 559 teachers put the structure to a sterner test. A one-factor model, in which all 22 items loaded on a single latent variable, fit the data poorly, with a chi-square-to-degrees-of-freedom ratio of 16.82 and a comparative fit index of just 0.702. The three-factor model fit dramatically better, and the difference between the models was overwhelmingly significant. After twelve within-factor residual covariances, identified post hoc from modification indices and justified by overlapping item content, were added as refinements, the final model achieved a chi-square ratio of 3.51, a root mean square error of approximation of 0.067, a standardized root mean residual of 0.044 and fit indices above 0.94. A supplementary higher-order model, in which a general AI self-efficacy factor predicted the three dimensions, produced mathematically identical fit, indicating that a broad general factor coexists with the domain-specific structure rather than replacing it.

The psychometric evidence for reliability and validity is unusually strong. Cronbach’s alpha reached 0.96 for the total scale and ranged from 0.92 to 0.96 across subscales, with bootstrapped confidence intervals confirming the precision of these estimates. Composite reliability ranged from 0.92 to 0.95, and average variance extracted exceeded 0.50 for every factor, supporting convergent validity. Discriminant validity was verified through three complementary criteria: average variance extracted exceeded both maximum and average shared variance for each factor, the square root of each factor’s average variance extracted surpassed its correlations with the other factors, and all heterotrait-monotrait ratios fell below the conservative 0.85 cutoff. The strongest association, between pedagogical integration and awareness, at 0.767, signals conceptual kinship while still confirming that the dimensions are empirically distinguishable.

Substantively, the results reveal an intriguing asymmetry in teachers’ confidence. Teachers reported their highest mean scores on awareness and understanding and on pedagogical integration, and their lowest on AI use and learning, suggesting they feel more capable of using AI in familiar instructional tasks and interpreting its outputs than of deploying it as a personal learning and research partner. From a social cognitive perspective this pattern is coherent: instructional planning is a domain where teachers have accumulated extensive mastery experience, while self-directed AI learning demands new forms of engagement for which such experience is still scarce. Notably, the five ethics-related items in the original pool were all eliminated during expert review, largely because they expressed attitudes and beliefs rather than capabilities. The authors are explicit that the AISES-T should not be read as a measure of ethical AI self-efficacy, and they flag capability-framed ethics items, such as recognizing bias in AI outputs or protecting student data, as a priority for future development.

The practical implications reach from teacher education to policy. Because the scale yields both a total score and a diagnostic subscale profile, pre-service programs can track growth across the three dimensions over a course of study, while schools and ministries can differentiate professional development: teachers weak in AI use and learning may need guided practice with AI as a personal tool, whereas those weak in integration or awareness may benefit from lesson-design workshops paired with training in critically evaluating AI outputs. The authors caution that the instrument measures beliefs rather than demonstrated competence and should not inform high-stakes decisions about individual teachers. They also acknowledge limitations: both samples were recruited by convenience in Türkiye, the residual covariances require replication in new samples, measurement invariance across gender, school level and experience remains untested, and the rapid evolution of generative AI may require periodic review of item wording. Even so, the AISES-T gives researchers, teacher educators and policymakers something the field has lacked: a validated, theory-grounded window into where teachers feel confident with AI, and where the confidence gap that could determine the technology’s educational impact truly lies.

Subject of Research: Psychometrics and Educational Technology

Article Title: Teachers' artificial intelligence self-efficacy: development and validation of a multidimensional scale grounded in social cognitive theory

Article References: Topcubaşı, T., Keleş, O., Demir, S., & İşçi, T. G. (2026). Teachers' artificial intelligence self-efficacy: development and validation of a multidimensional scale grounded in social cognitive theory. Frontiers in Education, 11, Article 1948951. https://doi.org/10.3389/feduc.2026.1948951

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1948951

Keywords: artificial intelligence, teacher self-efficacy, scale development, psychometrics, educational technology, Social Cognitive Theory, teacher professional development, factor analysis, Teachers', artificial, intelligence, self-efficacy