Teachers and Generative AI: A Systematic Review of Adoption, Competence, Professional Development, and Socio-Technical Experiences

Abstract

Generative artificial intelligence (AI) has rapidly become integrated into teaching, yet evidence on how teachers actually adopt, experience, and implement it remains fragmented. This systematic review synthesizes 105 peer-reviewed empirical studies published since 2023, identified through a PRISMA-guided search of three leading educational technology journals indexed in Scopus and Web of Science. The findings show pronounced geographical disparities: research is concentrated in China, Turkey, the United States, and Spain, while Africa and Oceania are markedly underrepresented, skewing the evidence base toward high-resource systems. Across studies, AI literacy, pedagogical knowledge, and self-efficacy, often theorized through the Technological Pedagogical Content Knowledge (TPACK) framework, are the strongest predictors of effective integration. Adoption depends less on technological access than on socio-technical conditions: institutional support, leadership, professional development, trust, and teacher autonomy. Most studies use cross-sectional self-reports that capture intention rather than practice, leaving classroom implementation, longitudinal change, and equity barriers underexamined. This study argues that effective integration requires aligning technological, human, and organizational systems and outlines an agenda for context-sensitive, practice-oriented research.

1. Introduction

Artificial intelligence (AI) has emerged as a fundamental catalyst for systemic pedagogical and structural change in contemporary education. While AI has long been embedded in education through systems such as intelligent tutoring, learning analytics, and automated assessment tools, the emergence of generative AI has fundamentally reshaped its scope, accessibility, and pedagogical implications. Large language models now enable the generation of instructional content, real-time feedback, adaptive explanations, and decision-support functions that significantly expand the role of AI in teaching and learning (Kasneci et al., 2023). This development has shifted AI from a supportive educational tool to an important component of instructional and institutional transformation.

The rapid expansion of AI in education has generated a substantial and growing body of research. Existing studies have examined AI-supported instruction, teacher–AI collaboration, learning analytics, professional development, and the ethical implications of AI integration in educational contexts (e.g., Ayubian et al., in press; Bergdahl & Sjöberg, 2025; Hoang, 2025; Ouyang et al., 2022). While much of the literature emphasizes the potential of AI to enhance efficiency, personalization, and instructional design, it simultaneously highlights concerns related to bias, transparency, privacy, academic integrity, and the evolving professional role of teachers (Kasneci et al., 2023; Selwyn, 2024; Williamson & Eynon, 2024). Collectively, these studies suggest that AI integration is not a purely technical innovation but a socio-technical transformation that reshapes pedagogical practice, institutional structures, and professional identities.

Within this transformation, teachers remain central actors. Despite the increasing sophistication of AI systems, their educational impact depends largely on how teachers interpret, evaluate, and integrate these tools into their instructional practices. Teachers are expected to critically assess AI-generated outputs, guide students in responsible AI use, and make informed pedagogical decisions while maintaining ethical standards and instructional quality. Consequently, effective AI integration requires more than access to technology; it depends on teachers’ professional knowledge, pedagogical beliefs, emotional readiness, and institutional conditions (Chiu et al., 2024; Ding et al., 2025; Nazaretsky et al., 2022).

Although the literature on AI in education is rapidly expanding, important gaps remain. Previous reviews have largely focused on technological developments, student learning outcomes, or broad applications of artificial intelligence in education (Celik et al., 2022; Zawacki-Richter et al., 2019). In comparison, less attention has been devoted to synthesizing empirical evidence specifically focused on teachers’ experiences of AI integration. This limitation has become increasingly significant in the post-2022 era, during which generative AI has accelerated both the volume and complexity of educational research. As a result, the literature is highly fragmented across contexts, methodologies, and theoretical approaches, making it difficult to develop a coherent understanding of how teachers engage with AI across different educational settings.

Moreover, much of the existing research emphasizes teachers’ intentions to adopt AI rather than examining how AI is implemented in instructional practice (e.g., X. An et al., 2023; Elyakim, 2025; Lucas et al., 2024; Wut et al., 2025). While technology acceptance models such as TAM and UTAUT have provided useful insights into perceived usefulness and ease of use, they offer limited explanation of how teachers navigate the pedagogical, emotional, ethical, and institutional dimensions of AI integration in real-world classrooms. There is therefore a need for a more comprehensive synthesis that integrates these dimensions and provides a holistic understanding of AI implementation in teaching.

To address these gaps, this systematic literature review synthesizes empirical research on AI in teaching and for teachers published between 2023 and 2025. Rather than focusing primarily on technological capabilities or student outcomes, this review places teachers at the center of analysis and examines how AI is being adopted, experienced, and implemented across diverse educational contexts.

2. Literature and Theoretical Foundations

2.1. AI in Teaching

Artificial intelligence (AI) has rapidly transitioned from a specialized field into a foundational pillar of modern education. The release of large language models (LLMs) and generative tools has accelerated interest in how machine learning can support, enhance, or alter the role of educators. However, meaningful integration of these AI-driven tools involves more than simply deploying software. It requires a much larger shift that impacts teacher identity (Kashif et al., 2025), pedagogical beliefs (Prestridge et al., 2025), and institutional systems (Ghiasvand & Seyri, 2025).

Additionally, the current wave of educational AI tools differs from previous educational technologies. While early technological tools focused on embedded automated grading applications and auto-generated summaries of student performance on assessment platforms (Zawacki-Richter et al., 2019), modern AI platforms can independently adapt, generate new content, and evaluate complex data patterns (Kasneci et al., 2023). Today, research shows that, while the deployment of these advanced AI technologies is primarily concentrated in higher-education settings, applications in secondary and primary school environments are rising rapidly (Lee & Kwon, 2024).

In educational spaces across the globe, this influx of technology is shifting instruction away from traditional, teacher-centered classrooms and leaning toward a hybrid approach where educators and smart tools work together. Implemented well, this division of labor offloads administrative tasks such as scheduling and attendance, reducing teacher workload and supporting teachers’ behavioral and emotional investment (Ding et al., 2025). In contrast, poorly integrated technology can cause role confusion and increase prep time, creating or perpetuating a stressful cycle of burnout (Ding et al., 2025).

As teachers navigate the challenges of this technological shift, these tools are already transforming the traditional roles of teachers. Chiu and Rospigliosi (2025) point out that teachers are transitioning into facilitators who review AI recommendations to guide students down differentiated learning paths, lead discussions about the nuances of digital citizenship, and hone their evaluation skills as they look out for biases and “hallucinations” from AI content. However, many teachers avoid AI technologies due to a lack of confidence and fear of replacement. Nazaretsky et al. (2022) found that explaining how AI tools make decisions clarifies misconceptions and builds trust. Supporting this, Ding et al. (2025) reported a positive relationship between institutional support and teachers’ self-efficacy. When educators feel leadership provides clear policies, training, infrastructure, and encouragement, their confidence increases. Ultimately, to successfully utilize AI to enhance student learning, teachers require ongoing support from the entire educational ecosystem, including school systems, policymakers, and preparation programs.

2.2. Theoretical Frameworks in AI Implementations

When looking at how technology is integrated into educational settings, guiding theories have shifted from looking at individual choices to focusing on large, interconnected systems. In the context of AI integration, these frameworks can be understood as complementary rather than competing perspectives, with each addressing a different dimension of teachers’ implementations of AI. Early foundational models highlighted psychological acceptance, such as the Technology Acceptance Model (TAM) by Davis (1989), which argues that users adopt a tool based on its perceived usefulness and ease of use. While the TAM provides a useful explanation of individual perceptions of technology, it does not account for the human dynamics of working in a school. Parallel to this, Hord et al. (1987) introduced the Concerns-Based Adoption Model (CBAM). Instead of viewing the use of technology as a choice that teachers make, the CBAM maps out the developmental stages teachers go through during a school rollout. Building on both individual and institutional paradigms, Venkatesh et al. (2003) introduced the Unified Theory of Acceptance and Use of Technology (UTAUT). The UTAUT grouped existing models and expanded them by adding social influence and organizational support as factors to predict technology use. Taken together, the TAM, CBAM, and UTAUT provide a foundation for understanding perceptions, intentions, concerns, and institutional conditions in shaping teachers’ technology adoption.

With the rapid integration of AI technology into pedagogical tools, researchers have shifted acceptance models toward instructional and knowledge-based frameworks. P. Mishra and Koehler (2006) introduced the Technological Pedagogical Content Knowledge (TPACK) framework, pointing out that teachers must understand how technology blends with their specific content, pedagogy, and context. In this review, TPACK provides a lens for understanding teachers’ AI competence and their ability to integrate AI with content and pedagogy. To measure how deeply teachers are integrating new technology in the classroom, Puentedura (2006) developed the Substitution, Augmentation, Modification, and Redefinition (SAMR) model. Rather than simply quantifying usage, the SAMR model evaluates whether an educator is utilizing a tool to execute a functional substitution or to facilitate fundamental pedagogical shifts that redefine the learning environment. Therefore, TPACK highlights teachers’ knowledge, while SAMR focuses on the depth of technology integration in teaching practice.

However, instructional technology frameworks like TPACK and SAMR struggle to fully address the nature of AI tools which can learn, adapt, and interact. Desveaud and Bawack (2026) most recently proposed a framework built on Socio-Technical Systems (STS) Theory. The model distinguishes three components: AI-related attributes such as reliability, human-related attributes such as teachers’ aspirations and goals, and the interactional space where they meet, including user experience and emotional connection. In this review, the STS perspective provides a broader perspective by considering the interaction among human, technological, and institutional factors in AI integration.

2.3. Aims of Review

The rapid growth of educational technology research after the development, release, and utilization of AI powered education tools has created a vast yet fragmented body of research. While previous reviews have mapped the use of machine learning in education, they have focused on the design of AI tools and student outcomes, leaving teachers’ perspectives underexamined (Celik et al., 2022). This systematic literature review synthesizes recent peer-reviewed empirical studies to clarify how teachers navigate the technical, pedagogical, and emotional demands of modern educational AI. By examining the exponential growth of AI research published between 2023 and 2025, this study charts the global development of this field during an era of widespread generative AI adoption.

This review first gauged an overview of AI research based on a central question: what are the developmental and emerging patterns of research on AI in teaching across the world?

As part of unpacking the patterns of research as the focus of this review, we also explored the following questions:

(1)

What are teachers’ perceptions of AI integration?

(2)

What is the role of AI competence and professional development in fostering/hindering AI adoption?

(3)

What are the psychological and ethical dimensions of AI implementation?

3. Methods

A systematic review is defined as a set of methods to reduce systematic biases in selecting, reviewing and synthesizing relevant studies on a topic (Petticrew & Roberts, 2008). To avoid individual bias, as was common in traditional reviews, systematic reviews should always have a transparent overview of their methods and can be conducted in teams (Macaro, 2022). Our researcher/educator team, working in the same university, holds familiarity and expertise with AI in teaching.

3.1. Search Protocol

The current review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) standards for conducting a systematic review (Page et al., 2021). According to the PRISMA guidelines, an SLR generally requires two or more comprehensive databases covering scientific journals (Moher et al., 2009). Elsevier’s Scopus database, as a comprehensive source over 20,000 globally ranked journals (D. Mishra et al., 2017; Vijaya & Mathur, 2023), and the Web of Science database, a well-established platform for educational research, were selected to obtain the required literature data for this study. With the assistance of a university librarian with expertise in conducting systematic reviews, the search terms were carefully combined to ensure the inclusion of broad yet relevant studies. To maximize the rigor, consistency and quality of the dataset, the search scope was deliberately limited to three peer-reviewed leading journals in educational technology research: Computers & Education, Education and Information Technologies, and Education Sciences. These three journals were selected as the highest-volume peer-reviewed outlets publishing post-2022 empirical research specifically on AI integration in teaching, enabling a focused thematic synthesis of the field’s most concentrated evidence base. The final thematic search query comprised the following: (“Artificial Intelligence and Education” OR “AI and Education” OR “Artificial Intelligence and Teachers” OR “AI and Teachers”) AND (“Education” OR “Teacher” OR “AI*”) AND PUBYEAR > 2022.

The searches were conducted on 6 November 2025, and included records published from 2023 through to the date of the search, including any early online or ahead-of-print records available at that time. To encompass the most recent and relevant literature, the search timeframe was set from 2023 due to the release of ChatGPT (GPT-3.5) by OpenAI in November 2022. Moreover, this study only included peer-reviewed journal articles written in English. After removing duplicated records, 654 records were identified and were entered into a systematic review software called Covidence (Veritas Health Innovation, 2025).

3.2. Inclusion and Exclusion Criteria

For the purposes of this review, artificial intelligence (AI) was defined as computational systems capable of performing tasks that typically require human intelligence. This definition included learning from data, generating novel content, adapting behavior based on input, or supporting decision making. This encompasses generative AI and large language models (e.g., ChatGPT), intelligent tutoring and adaptive learning systems, AI-driven learning analytics, and AI-enabled robotics. Technologies without a demonstrable AI or machine learning component (e.g., non-adaptive augmented or virtual reality and standard learning-management system features) were excluded unless the study explicitly examined an AI-enabled application of that technology.

As the first step in this review, all records’ titles and abstracts were screened to exclude those unrelated to AI in teaching and for teachers. A total of 429 studies were irrelevant. In the next step, a full-text review was conducted to apply inclusion and exclusion criteria. Out of 225 records, 120 studies were excluded based on the following criteria:

• Non-human (review records): Studies that did not report original empirical data, such as systematic literature reviews, conceptual papers, and review articles.
• Various participants: Studies involving mixed participant groups were excluded when teachers/faculty were not the sole participant group.
• Scale development: Papers developed an instrument or a scale.
• Research with students as the participants.
• Research not related to AI.
• Research conducted before 2023.

Studies were included if they were:

• Reporting on empirical data, using qualitative and/or quantitative methods.
• Written in English.
• Involved teacher/faculty participants at K-12 or university settings.
• Involved preservice teachers or teacher candidates enrolled in tertiary-level teacher-education programs (distinct from the general student populations excluded under criterion 4 above).
• Empirical studies published in peer-reviewed journals.
• No geographic restriction: studies were eligible regardless of the country or region in which they were conducted.

Ultimately, 105 studies moved forward to the extraction stage for detailed review in terms of different factors. For all stages, each study was randomly assigned for two researchers familiar with the field. If a reviewer was unsure about targeted data at any stage from title and abstract screening to extraction, the second reviewer would decide on the target information. When both reviewers had a conflict, then it was screened by a third reviewer, minimizing bias in evaluation. The screening process is presented in the PRISMA flow diagram (Figure 1).

4. Findings

4.1. Descriptives

Countries were classified into six geographic regions (Asia, Europe, Middle East, Americas, Africa, Oceania) based on the country in which each study’s data were collected, with Western Asian countries (e.g., Turkey, Saudi Arabia, Iran) reported separately as “Middle East” to align with common regional conventions in the AI-in-education literature. The 105 research studies included in this systematic review were conducted across a range of geographic areas. The majority of studies were conducted in Asia (n = 35) and Europe (n = 22), followed by the Middle East (n = 20), and the Americas (n = 13). Africa (n = 2) and Oceania (n = 2) were comparatively underrepresented among the research studies included. Several studies were conducted with mixed- or multi-country geographical settings (n = 11). Looking more closely at the geographical distribution, we can see that China (n = 25), Turkey (n = 12), the United States (n = 11), and Spain (n = 6) are the most representative areas in the sample. These are followed by Germany (n = 5), Sadi Arabia (n = 5), and South Korea (n = 3). Figure 2 illustrates the distribution of studies by geographical areas.

Most of the studies included in this systematic review collected data from a very small sample of teachers, defined as less than 30 teachers (n = 30), or a medium sample size of between 101 and 300 participants (n = 29). A considerable number of studies used a large sample size of 301–1000 teachers (n = 24). A few studies used a sample size of between 31 and 100 teachers (n = 16), and only a few studies used a very large sample size of more than 1000 teachers (n = 6). Figure 3 shows the sample size distribution by range.

In terms of design and data-collection tools, about half of the studies in this systematic review used a quantitative design (n = 49). The data collection for these studies included surveys, Likert-scale questionnaires, pre/post-tests, and system analytics. Some studies adopted a mixed-methods design (n = 30) with a combination of surveys and interviews, classroom observations, and system logs as data-collection tools. The remaining studies used a qualitative design (n = 26) with various data-collection methods including semi-structured interviews, focus groups, and field notes. Figure 4 shows the distribution of studies by design.

Most studies focused on AI at the higher-education level (n = 54). Within K-12, 11 studies addressed general K-12 education, 11 addressed mixed K-12 levels, and 10 addressed secondary education (middle and high school) specifically. Nine studies focused on elementary/primary education, and four focused on early childhood/preschool. An additional six studies examined AI across mixed K-12 and higher-education levels. Figure 5 indicates the distribution of research by level of study.

To summarize, almost half of research studies included in our systematic review were conducted in higher-education settings with teacher participants. About half of the studies adopted quantitative design and recruited a sample of more than 100 teachers. While studies were geographically distributed across the world, research on AI in teaching and for teachers was more dominant in Asia and Europe. We will revisit these themes in our discussion.

4.2. AI Acceptance, Attitudes, and Behavioral Intention

The largest cluster of studies (n = 35) examined acceptance, attitudes, and behavioral intention toward AI (e.g., Bozkuş & Canoğulları, 2025; Cabero-Almenara et al., 2024; L. Hu et al., 2025; Jatileni et al., 2024; Kim, 2024a; Polat, 2025; Sanusi et al., 2024; Stupurienė et al., 2024; Theodorio et al., 2024; Wut et al., 2025; C. M. Zhang et al., 2025). Most were set in higher education, fewer in school contexts, and only a handful spanned mixed settings (Alagöz Hamzaj, 2025; Alwaqdani, 2025); in-service and preservice samples were roughly balanced.

These studies were predominantly quantitative, using cross-sectional surveys grounded in UTAUT and TAM and analyzed with regression, PLS-SEM, and SEM to model relationships among ease of use, trust, perceived usefulness, anxiety, behavioral intention, and digital literacy (e.g., Al-Abdullatif, 2024; X. An et al., 2023). Across these models, effort expectancy and perceived usefulness consistently predicted adoption (X. An et al., 2023; L. Hu et al., 2025). A smaller group extended beyond standard acceptance scales to capture effective and belief-based predictors (Chai et al., 2024; Huertas-Abril & Palacios-Hidalgo, 2023); notably, Sosa-Alonso et al. (2025) found that teachers’ pedagogical convictions often outweighed institutional pressures. Another subset combined surveys with qualitative data to capture teachers’ lived experiences and institutional barriers (e.g., Alwaqdani, 2025; Blundell et al., 2025; Cordero et al., 2025), while a few used purely qualitative designs to probe localized expectations and the systemic realities of AI integration (e.g., Dehghani & Mashhadi, 2024; Elyakim, 2025).

A second strand foregrounded structural and contextual conditions, arguing that adoption is structurally embedded rather than an individual choice (Adawurah & Buabeng-Andoh, 2025, studying preservice teacher-training students; Bozkuş & Canoğulları, 2025; Choi et al., 2025; Matschke et al., 2026). Fewer studies moved beyond intention altogether; Elyakim (2025), for instance, used reflective analysis to surface persistent gaps between teachers’ expectations and classroom reality. The remaining studies linked acceptance to scaffolding, institutional policy, and personal traits (e.g., Habib, 2025; Harakchiyska, 2025; Kölemen & Yıldırım, 2025; E. M. Lim, 2023; Mohamed, 2024; Runge et al., 2025; Strzelecki et al., 2024; Traga Philippakos & Rocconi, 2025).

Overall, the literature conceptualizes AI integration as intentional behavior rather than enacted outcome, relying on single-timepoint self-reports. With few studies triangulating sources or following teachers longitudinally, the field captures perceptions well but leaves the realities of classroom integration largely unmeasured.

4.3. AI Competence and TPACK

Thirteen studies addressed AI competence and teachers’ knowledge frameworks, with most reporting positive associations between competence and AI integration (Cabero-Almenara et al., 2025; Chiu et al., 2024; Fan et al., 2025; Karataş & Ataç, 2025; Lan et al., 2025; Lucas et al., 2024; Milutinović, 2025; Velander et al., 2024; Xu et al., 2025a, 2025b; Yau et al., 2023; Yue et al., 2024; Z. Zhang et al., 2023). Ten found that teachers with stronger digital competence, pedagogical knowledge, and self-efficacy were more ready to integrate AI (e.g., Karataş & Ataç, 2025; Yue et al., 2024), and several showed that AI-focused professional development raised readiness and confidence (e.g., Chiu et al., 2024; Lan et al., 2025; Yau et al., 2023). Others reported only moderate or uneven readiness, pointing to persistent gaps in AI literacy and pedagogical preparation (e.g., Cabero-Almenara et al., 2025; Yue et al., 2024).

Competence was operationalized unevenly across studies. Most studies (n = 11) used perception-based self-reports, including AI-TPACK scales, AI-literacy surveys, and self-efficacy instruments (e.g., Xu et al., 2025a; Yue et al., 2024), while two studies used practice-oriented or task-based measures (e.g., Z. Zhang et al., 2023). Multi-instrument studies linked AI-TPACK to its predictors: Xu et al. (2025a) tied AI attitudes and self-efficacy to AI-TPACK development, and Karataş and Ataç (2025) identified pedagogical knowledge as a key predictor of integration via SEM. A growing set extended TPACK to reflective and disciplinary dimensions, including ethical awareness (Milutinović, 2025) and AI pedagogical content knowledge for non-STEM elementary teachers (Fan et al., 2025).

Samples skewed toward preservice teachers in teacher-education settings (n = 9, including one mixed-sample study) over in-service teachers (n = 5, including the same mixed-sample study); Fan et al. (2025) was the only study to sample both groups. Several studies called for AI-competency measures to be updated continually to keep pace with rapidly shifting AI technologies. Overall, the evidence is dominated by perception-based readiness and conceptual modeling; few studies examine authentic classroom implementation or the longitudinal impact of competence on pedagogical change.

4.4. Professional Development and Training

Nine studies examined professional development (PD) and training aimed at teachers’ AI integration (Bergdahl & Sjöberg, 2025; Hong et al., 2025; L. Huang et al., 2025; Kohnke et al., 2025; J. Lim et al., 2025; Song et al., 2025; J. Sun et al., 2023; Xiao et al., 2025; Yang, 2025). PD emerged as a central driver of adoption, particularly when structured, sustained, practice-oriented, and tailored to context (e.g., Xiao et al., 2025; Yang, 2025); however, its effectiveness varied with regional and institutional conditions, indicating that PD outcomes are mediated by systemic factors (e.g., Bower et al., 2024; Ghamrawi et al., 2024; Hoang, 2025). Designs were mostly quantitative or quasi-experimental (n = 4), with fewer mixed-methods (n = 3) and qualitative (n = 2) studies; surveys predominated, supplemented by reflective journals, lesson artifacts, and performance tasks.

Most evidence came from intervention studies. Hybrid online–in-person training (Xiao et al., 2025), TPACK-based PD (J. Sun et al., 2023), and simulation-based PD (J. Lim et al., 2025) each improved teachers’ AI competence, integration skills, or instructional decision making. Emerging formats showed similar gains: robotics-focused PD for early-childhood teachers (Yang, 2025) and digital micro-learning (Kohnke et al., 2025) all increased confidence and innovation. Beyond individual interventions, several studies emphasized systemic support, institutional leadership (Bergdahl & Sjöberg, 2025) and peer-led professional learning communities for co-designing AI lessons (Hong et al., 2025; Song et al., 2025). Collectively, the evidence indicates that structured, sustained, practice-based PD, when aligned with institutional context, strengthens teachers’ readiness, confidence, and AI competence.

4.5. Instructional Practice and Classroom Pedagogy

The largest body of work (n = 36) examined how AI reshapes instructional practice, lesson planning, feedback, decision making, classroom interaction, and teaching strategies (e.g., Celik et al., 2026; Egara & Mosimege, 2024; Hong et al., 2025; Hwang et al., 2025; Starks & Reich, 2023; Z. Sun et al., 2023; N. Zhang et al., 2025).

One strand examined AI-supported feedback and classroom discourse (Alanazi et al., 2025; S. An et al., 2025; Chang & Sun, 2026; Filiz et al., 2025). Automated feedback improved teachers’ questioning quality (Demszky et al., 2025), and real-time interpretation of student responses supported responsive teaching (N. Zhang et al., 2025), while NLP methods identified effective discourse patterns (Shin et al., 2023) and student misconceptions (Kökver et al., 2025), together positioning AI as a support for data-driven instructional decisions. A second strand examined teacher–AI collaboration (Dahri et al., 2025; Kang et al., 2025; Yorulmaz et al., 2025), which was found to be most effective when AI complements rather than replaces the teacher (Jeon & Lee, 2023) and when teachers are ready and favorably disposed toward it (Bao et al., 2025).

A third strand addressed AI in content creation and instructional design (H. Chen & Wang, 2023; Giaouri & Charisi, 2025; Y. Huang et al., 2023; Kim, 2024b; Lyu et al., 2025; Mayer & Schwemmle, 2023). AI-assisted planning enhanced creativity and efficiency, though teachers struggled to judge the quality of generated materials (Şimşek, 2025), improved mathematics design (Martínez-Zarzuelo et al., 2025), supported language teaching workflows (Liu & Yao, 2025), and helped customize materials to diverse learner needs (Almuhanna, 2025). A fourth examined immersive technologies, where augmented reality and simulations improved conceptual understanding, retention, engagement, and cognitive load (Dann et al., 2024; Segaran & Moltudal, 2025; Z. Sun et al., 2023; Uygun et al., 2025; Wei, 2025).

Finally, studies in specialized contexts qualified these gains: AI analytics aided real-time monitoring in science (Shi et al., 2024), but special-education teachers raised accessibility and integration concerns (Alhaif et al., 2025), and rural settings showed that weak infrastructure limits implementation despite positive teacher attitudes (López Costa, 2025; Novoa-Echaurren et al., 2025; Starks & Reich, 2023). Across contexts, instructional outcomes were generally positive for efficiency, personalization, innovation, and responsiveness, but ultimately depended on teachers’ capacity to critically evaluate AI outputs and adapt them to classroom realities.

4.6. Emotion, Anxiety, Trust, and Wellbeing

Eight studies addressed the emotional and psychological dimensions of AI integration, anxiety, trust, wellbeing, and motivational readiness (Cambra-Fierro et al., 2025; Darancik et al., 2025; Delello et al., 2025; Hopcan et al., 2024; Y. Hu et al., 2025; Jabali et al., 2025; Li et al., 2025; Verano-Tacoronte et al., 2025). Anxiety featured as a central barrier: fear of professional displacement, technological uncertainty, and academic integrity concerns drove negative emotions among higher-education faculty (Verano-Tacoronte et al., 2025), and low digital competence was associated with greater apprehension among preservice teachers (Darancik et al., 2025).

Trust and perceived reliability also shaped emotional responses, with familiarity, institutional scaffolding, and emotional proximity influencing willingness or resistance (Hopcan et al., 2024; Y. Hu et al., 2025; Jabali et al., 2025). Studies of wellbeing identified a recurring tension: although AI reduced workload, over-reliance raised concerns about pedagogical ownership and professional identity (Cambra-Fierro et al., 2025), and mounting administrative expectations introduced psychological pressure, affecting mental health (Delello et al., 2025). Li et al. (2025) captured this ambivalence directly, finding that curiosity and enthusiasm coexisted with performance anxiety and frustration among preschool teachers. Collectively, these studies indicate that teachers’ willingness to implement AI depends heavily on resolving emotional and psychological concerns.

4.7. Ethics, Leadership, and Policy

Only four studies focused specifically on the ethical, leadership, and policy dimensions of AI integration (Al-Zahrani & Alasmari, 2025; Bower et al., 2024; Ghamrawi et al., 2024; Hoang, 2025). Surveying 318 teachers across disciplines and regions, Bower et al. (2024) found that institutions lacked clear guidelines for responsible AI use and called for institution-wide governance addressing academic integrity, authorship, and sustainable implementation. Leadership mattered: school leaders with stronger AI-leadership competencies were more supportive of innovation and teaching readiness (Hoang, 2025). At the system level, Al-Zahrani and Alasmari (2025), studying 19 MENA countries, documented substantial regional inequalities in AI policy and institutional preparedness, and Ghamrawi et al. (2024) cautioned that AI can build leadership capacity through peer coaching but may erode teacher autonomy when it dominates curricular decisions. Together, these studies indicate that effective integration requires institutions to move beyond tool adoption toward stronger ethical and macro-level policy.

Across all themes (see Appendix A), the evidence base rests largely on single-timepoint self-reports from in-service and preservice teachers. Far less is known about how integration unfolds over time, across contexts, or from the perspective of other stakeholders, and limited evidence addresses systemic equity and access. What remains consistent is that AI literacy, pedagogical knowledge, self-efficacy, and institutional support are fundamental to effective integration.

5. Discussion

5.1. Geopolitical Gaps in Research Coverage

The overarching pattern indicates that China, Turkey, the U.S., and Spain emerge as the leading contributors to the studies included in this review on AI in teaching across different educational contexts. This aligns closely with broader bibliometric trends documenting the global expansion of technology-enhanced learning (X. Chen et al., 2020; Hinojo-Lucena et al., 2019; Tang et al., 2023; Zawacki-Richter et al., 2019). The prominence of these nations may be associated with different factors such as substantial state-level funding, national AI strategies, dense concentrations of high-resource tech institutions, extensive infrastructure investments, and targeted education policies (State Council of the People’s Republic of China, 2017; Turkey’s National Artificial Intelligence Strategy 2024–2025; National Artificial Intelligence Initiative Act, 2021; Government of Spain, 2020; Ministry for Digital Transformation and Civil Service, 2024).

Historically, field-level mappings of artificial intelligence in education (AIED) have identified China and the United States as the dominant contributors to the field (Radu et al., 2024). However, their contributions reflect distinct patterns. While the United States remains a leading source of pedagogical frameworks, ethical guidance, and advanced generative AI models that underpin many educational applications (Cruvinel Júnior et al., 2025; OpenAI, 2023), China has surpassed the United States in overall publication output and research examining AI implementation in classroom settings (N. Zhang et al., 2025). Together, these patterns in the present review suggest a globally interconnected AI research landscape in which large-scale educational implementation and the development of core AI technologies are often concentrated across different regions.

Furthermore, recent studies challenge the notion that global research leadership can be measured by publication counts alone (Agarwal, 2026; Omaar, 2024). This nuance is reflected in our systematic review’s data: while Asia led in absolute volume with 35 studies, predominantly driven by China (n = 25), and Europe followed with 22 studies, the United States accounted for 11 studies. Our findings show that, despite having a lower volume of research than countries in Asia, contexts such as the U.S. and Spain remain central to shaping the field’s structural paradigms.

A significant countertrend strongly validated by our review data, however, is the limited representation of Africa and Oceania, with each region accounting for only two studies out of the 105 reviewed. This highlights stark disparities consistently reflected in broader AIED bibliometric analyses (Guo et al., 2024; Mittal et al., 2026). These gaps may reflect unequal access to research funding, technological infrastructure, internet connectivity, institutional support, and policy implementation.

This geographical imbalance suggests that the literature reviewed may not fully capture the realities, opportunities, and challenges faced by low-resource educational systems (Okolo, 2023). This is further compounded by the specific demographics and methodologies dominant in the sampled literature. For instance, the vast majority of the reviewed studies focused strictly on AI in higher education, leaving critical foundational levels such as preschool and elementary/primary less represented. Furthermore, because nearly half of the included studies relied on rigid quantitative designs, the evidence base may provide less insight into the contextual and qualitative realities of marginalized and underrepresented contexts.

5.2. Understanding AI Acceptance Versus Classroom Realities

This thematic focus addresses teachers’ perceptions of AI integration and highlights the growing global interest in teachers’ willingness to adopt AI in their teaching. Recent scholarship emphasizes that successful AI adoption in school systems depends less on technological infrastructure and far more on the cognitive and affective dimensions of teacher preparation (Cabero-Almenara et al., 2024; Fteiha et al., 2025). This shift is reflected in global readiness frameworks such as the World Economic Forum’s (2026) AI Readiness Framework, which treats teachers’ psychological readiness for new technologies as a core instructional competency. At the same time, this global trend is supported by results from this study on established technology acceptance models like the TAM and the UTAUT.

To address these theoretical limitations, researchers are actively expanding traditional frameworks to incorporate nuanced variables, including educators’ specific pedagogical beliefs, ethical anxieties, and perceptions of machine intelligence (Cabero-Almenara et al., 2024). Sosa-Alonso et al. (2025) illustrate this shift, finding that teachers’ pedagogical beliefs often outweigh external institutional pressures. Explaining adoption therefore requires attending to these beliefs, not perceived utility alone.

The strong AIED research and use in higher education reported in this study mirrors global trends of several broader bibliometric analyses and systematic reviews in the field (X. Chen et al., 2020; Hinojo-Lucena et al., 2019; Tang et al., 2023; Zawacki-Richter et al., 2019). This may be explained by universities’ greater access to digital technologies, research funding, institutional autonomy, and experimentation with emerging AI tools, including generative AI platforms and intelligent learning systems (Ouyang et al., 2022). Higher education researchers have rapidly documented generative AI adoption due to immediate pressures related to academic integrity and student-led use. A 2025 meta-summary by the Digital Education Council, for example, points out that 86% of students use AI in their studies, with 54% using it weekly and nearly one in four using it daily (Kelly, 2024). This high rate of adoption is mirrored in other surveys, such as the Chegg Global Students Survey of 11,706 undergraduate students across 15 countries, which found that 80% of students worldwide have used generative AI to support their university studies (Chegg, Inc., 2025). The divergence, however, is that, while over 80% of higher education students actively use AI, faculty adoption lags significantly behind, with 88% of professors using it minimally (Kelly, 2024). Meanwhile, the limited number of studies at the preschool and primary levels suggests that AI integration in early education and inclusive education settings remains underexplored. This gap may reflect concerns related to developmental appropriateness, ethics, digital safety, teacher preparedness, and infrastructural limitations in younger learning environments (Khoo & Jamaludin, 2025). Nevertheless, the increasing number of K-12 studies reported in the findings indicates growing scholarly attention to understanding how AI can support teaching and learning across school systems (Lee & Kwon, 2024), particularly following the rapid expansion of generative AI technologies.

The researchers, however, opine that, while some progress has been made in K-12 teaching, balancing this rapid expansion requires anchoring technological tools in foundational learning theories such as social constructivism (Vygotsky, 1978), which will ensure that AI teaching tools are designed to facilitate teacher–learner interactions rather than isolating learners in ways that place them in a siloed and hyper-personalized bubble. Distributed cognition frameworks (Hollan et al., 2000) can help prevent AI from being seen as a cheating mechanism or an external brain doing the work for the student and instead as an integrated component of the student’s cognitive toolkit. In line with human-centered machine learning, this would mandate that developers prioritize algorithmic fairness, transparency, and culturally and locally specific transparency, as well as cultural and localized educational needs, over technocentric metrics that marginalize context.

A critical insight within this theme lies in the methodological design of the reviewed studies. The reliance on cross-sectional, self-reported Likert scales collected at a single point in time frames AI integration primarily as an intentional behavior rather than an empirical outcome. This trend directly contributes to what Elyakim (2025) describes as the “price-value paradox” (p. 16929) or the distinct pedagogical “expectation-reality gap” (p. 16933), where high expectations mismatch everyday classroom constraints. Measuring an educator’s intent via a one-time survey fails to capture actual utilization or illustrate how teachers navigate complex, dynamic classroom ecosystems. Furthermore, longitudinal adoption studies in the broader educational technology literature demonstrate that initial positive attitudes toward new tools often decay sharply over time. This decline is typically driven by immediate technical friction, a lack of ongoing instructional scaffolding, and systemic institutional bottlenecks such as deficient infrastructure or unsupportive school policies.

5.3. Pedagogical Competence and TPACK as Prerequisite Frameworks

Across reviewed studies, professional development (PD) emerged as a foundational mechanism through which teachers develop the knowledge, confidence, and skills needed to implement AI effectively in classroom settings. These findings address the following question: what is the role of AI competence and professional development in fostering or hindering AI adoption?

This finding aligned with contemporary technology integration theories, particularly the TPACK framework, which emphasizes that meaningful technology integration occurs when teachers develop an understanding of the complex interplay among technology, pedagogy, and content knowledge (P. Mishra & Koehler, 2006). Studies in this systematic review have demonstrated that AI-focused PD, grounded in TPACK principles, significantly enhances teachers’ instructional design capabilities, pedagogical decision making, and classroom implementation of AI-supported learning activities.

5.4. Evolving Professional Development for Socio-Technical Support

The findings further suggest that effective professional development has evolved beyond traditional one-time workshops toward sustained, practice-oriented, and contextually relevant learning experiences. Hybrid learning models, coaching and mentoring approaches, simulation-based training, and collaborative professional learning communities have been shown to strengthen teachers’ AI competencies and instructional self-efficacy (Ngongpah & Oni, 2025). These approaches enable educators to engage directly with AI technologies in authentic teaching contexts, thereby fostering confidence in their ability to use AI effectively. From the perspective of the TAM (Davis, 1989) and the UTAUT (Venkatesh et al., 2003), such PD initiatives increase both perceived ease of use and performance expectancy, which are two factors consistently associated with technology adoption. As teachers gain a clearer understanding of how AI systems function and how they can support teaching and learning, concerns about job displacement and technological uncertainty tend to diminish, leading to a greater willingness to integrate AI into instructional practice (U.S. Department of Education, Office of Educational Technology, 2023).

Conversely, the review also highlights how inadequate PD can impede AI adoption. A recurring finding across the studies is that PD initiatives often fail when they are disconnected from teachers’ instructional realities, lack contextual relevance, or are not supported by broader institutional structures. Studies have revealed substantial regional, institutional, and cultural variation in the effectiveness of AI-related PD, suggesting that training outcomes are heavily influenced by contextual factors such as technological infrastructure, leadership support, policy environments, and resource availability (Roshan et al., 2024; Zawacki-Richter et al., 2019). When PD is insufficiently tailored to educators’ needs or lacks opportunities for ongoing support, teachers frequently experience uncertainty, role ambiguity, and implementation challenges. These findings can also be interpreted through the Concerns-Based Adoption Model (CBAM), which proposes that educators progress through identifiable stages of concern during innovation adoption (Hall & Hord, 2020). Without sustained support and opportunities for continued learning, teachers often remain in the early stages of concern, characterized by uncertainty, anxiety, and self-focused apprehensions, which prevent successful implementation.

Furthermore, the findings support the assumptions of the Socio-Technical Systems Theory (STS), which argues that technological innovations are most effective when technological, human, and organizational systems are aligned (Baxter & Sommerville, 2011). AI adoption therefore cannot be understood solely as a matter of technological competence, as it also depends on teachers’ psychological readiness, institutional culture, leadership support, and the availability of resources necessary to sustain innovation. When these factors are absent, implementation may increase workload, exacerbate stress, and contribute to teacher burnout (Ding et al., 2025), ultimately undermining adoption efforts.

5.5. Balancing Adaptive Personalization with Pedagogical Agency

A major contribution of these findings is the evidence that AI can support more responsive and personalized instruction across diverse educational settings. The positive outcomes reported in science education and special education align with the recent literature, emphasizing AI’s capacity to facilitate adaptive teaching and individualized learning pathways (Tan et al., 2025). Research by Holmes and Tuomi (2022) argues that AI systems are particularly valuable in learning environments that require continuous monitoring of student progress, as they enable teachers to identify learning difficulties and adjust instruction more rapidly than traditional approaches. Similarly, Garzón et al. (2025) contend that AI-driven learning analytics can enhance instructional responsiveness by providing educators with actionable insights about student engagement, performance, and misconceptions. The study’s findings from science and special education contexts therefore reinforce the notion that AI can strengthen teachers’ ability to differentiate instruction and address diverse learning needs.

However, the study also highlights that technological effectiveness alone does not guarantee successful instructional transformation. The concerns regarding accessibility, infrastructure, and integration barriers reflect a broader theme emerging in contemporary AI-in-education research about the persistence of contextual inequalities (Fitas, 2025). Recent studies have warned that unequal access to digital resources may exacerbate existing educational disparities, particularly in rural and under-resourced settings (UNESCO, 2024; OECD, 2025). The challenges identified in rural districts therefore suggest that, even when teachers hold favorable attitudes toward AI, inadequate technological infrastructure can limit its meaningful implementation. Supporting broader arguments that successful AI adoption requires not only teacher willingness but also institutional investments in connectivity, technical support, and digital ecosystems capable of sustaining AI-enhanced teaching.

Findings from special education settings are also important because they highlight both the potential and the limits of AI in inclusive education. The existing research indicates that AI can aid students with diverse learning needs by offering adaptive content, personalized pacing, and assistive tools (Zawacki-Richter et al., 2019). However, concerns about accessibility and implementation indicate that inclusion is not guaranteed by technological advances alone (Manzoor & Vimarlund, 2018). Instead, the successful application of AI in special education requires careful design, proper teacher training, and alignment with individual student needs. This underscores recent calls for human-centered educational AI approaches that prioritize equity, accessibility, and pedagogical relevance over technological efficiency alone (Williamson & Eynon, 2024).

A key insight from these findings on the question “What is the impact of AI on instructional practice?” is that AI’s effect on teaching primarily depends on teacher expertise. In various educational environments, teachers must interpret AI-generated data, evaluate suggestions, and modify their teaching methods based on classroom needs. This supports emerging studies that suggest that teachers are crucial mediators of AI-enhanced learning rather than just technology users (Luckin, 2024; Selwyn, 2024). While AI can increase efficiency, personalization, and responsiveness, its success in education ultimately relies on teachers’ professional judgment and their ability to integrate technological insights with pedagogical strategies in context.

5.6. Institutional Policy, Trust, and the Ethical Imperative

Regarding the institutional aspects of AI adoption, recent research indicates that supportive leadership is essential for mitigating uncertainty and encouraging positive perceptions of educational innovation (UNESCO, 2023; U.S. Department of Education, Office of Educational Technology, 2023; Kaufman et al., 2025). According to Dexter and Richardson (2020), leaders who possess digital and AI-related competencies are better positioned to create environments that encourage experimentation, collaboration, and the responsible use of technology. Such leadership is particularly important because AI implementation often generates anxiety regarding job security, ethical responsibility, and changing professional expectations, but effective leadership can help mitigate these concerns by providing clear guidance, professional learning opportunities, and institutional support structures. Findings related to policy differences among educational systems highlight significant issues of fairness and justice. Recent international reports warn that unequal access to AI infrastructure, regulatory guidelines, and professional development opportunities may exacerbate existing educational disparities (OECD, 2025; UNESCO, 2023). Therefore, implementing ethical AI goes beyond just preparing individual teachers; it also involves systemic policies that promote equitable access, safeguard user rights, and set uniform standards for responsible AI use. These findings support the growing call for comprehensive governance frameworks that combine technological advancement with ethical oversight and social responsibility, aiming to address ethical concerns in AI implementation.

A key psychological aspect emerging from these findings is the relationship between trust, perceived control, and teachers’ professional autonomy in implementing AI. Research indicates that educators tend to adopt AI more readily when they see it as a tool that complements rather than replaces their expertise. For example, Chiu et al. (2024) found that teachers’ willingness to use AI depends heavily on their confidence in interpreting AI outputs and their belief that they maintain authority over instructional decisions. Conversely, concerns about the reliability, transparency, and accuracy of AI systems can lead to resistance, particularly in critical education contexts (Kasneci et al., 2023). These worries go beyond usability and acceptance, addressing broader issues of professional agency and identity. The concern that AI systems may influence or limit curricular and pedagogical choices reflects ongoing debates about balancing technological automation with human judgment. As Selwyn (2024) notes, while AI can improve efficiency and support decision making, over-reliance on algorithms might reduce teachers’ professional discretion and creativity. Similarly, Williamson and Eynon (2024) warn against uncritically adopting AI technologies that render educators passive recipients of machine recommendations, implying that teachers’ trust in AI depends on their sense of control and autonomy; thus, they are more inclined to adopt AI when it strengthens, rather than diminishes, their professional judgment and role as decision-makers in teaching and learning.

6. Limitations and Implications for Future Research

Despite the comprehensive nature of this review, several methodological limitations must be acknowledged. First, the searches were conducted in November 2025 and were restricted to three high-impact English-language journals; consequently, the relevant literature published outside these parameters or in regional proceedings was excluded. Second, the heavy reliance on cross-sectional, self-reported data across the included studies limits causal inferences regarding long-term AI adoption. Finally, while strict PRISMA guidelines were applied, the screening process inherently involves human judgment.

These limitations highlight several critical avenues for future research. The field urgently requires longitudinal and mixed-method designs to move beyond single-point survey data and capture the complex realities of AI integration over time. Furthermore, as the current evidence base is heavily skewed toward higher education, subsequent studies must urgently examine AI integration within K-12 and early childhood settings. This research must move beyond technological adoption metrics to investigate the systemic dynamics of implementation from the perspectives of administrators, students, and policymakers, paying particular attention to the ongoing psychological impacts of AI on teacher anxiety and professional wellbeing. Finally, to ensure equitable access, more empirical evidence is needed to address the systemic challenges encountered by stakeholders in under-resourced and rural educational settings. As professional development evolves beyond traditional workshops toward sustained, practice-oriented learning, future research should heavily evaluate the efficacy of hybrid training models, coaching, and collaborative professional learning communities (Ngongpah & Oni, 2025).

7. Conclusions

The increasing presence and rapid integration of generative AI have reshaped the educational landscape, driving a global expansion of research focused on teacher preparation and professional development. Through this systematic review of empirical evidence published since the introduction of advanced generative models, the findings indicate that, although AI may offer substantial opportunities for adaptive personalization and instructional efficiency (X. Chen et al., 2020; Lee & Kwon, 2024; Fan et al., 2025), technological access alone does not necessarily translate into successful and meaningful integration.

The primary implication of this research is that effective AI adoption may require a profound socio-technical alignment. Educational institutions should consider moving beyond mere tool acquisition toward cultivating teachers’ pedagogical agency, AI literacy, and trust. To support equitable and sustainable instructional transformation, policymakers and educational leaders should provide robust institutional scaffolding—ranging from secure technological infrastructure to clear ethical guidelines. Ultimately, navigating the future of AI in education requires continued attention to a human-centered approach that prioritizes pedagogical rigor, contextual equity, and the professional judgment and agency of educators.