The research design or planning stage is an important time investment. During this stage research team has an interaction to clearly define the scope, the research question and objective(s). In this scope literature review, we agree that scope lies at the intersection of three issues (Figure 2). The topic area is “generative artificial intelligence”, the context is education and educational research, and our approach is governance. Governance is the structure of relationships that bring about organizational coherence, authorize policies, plans with decisions, and account for their probity, responsiveness and cost-effectiveness (Gallagher, 2001).

The formulation of questions is one of the main premises of critical and reflective thinking (Stern et al., 2014). Building questions can be supported by standardized structures/acronyms, such as PICO (Population, Intervention, Comparison, and Outcome), that can impact research strategy. The PICO model is the most widely used in the construction of research questions, not only in evidence-based medicine but also in other health science disciplines; variations of the PICO model and specific adaptations have emerged. For the Social Science, the Spider (Sample, Phenomenon of Interest, Design, Evaluation, Research type) model can be used to improve research quality (Methley et al., 2014; Booth, 2001; Cooke et al., 2012; Mohamed et al., 2021).

Given GenAI’s multifaceted role in education and research, we outline two main items for the literature review protocol:

The bibliographic search was done in the WoS database in 2025 (Feb 28). The keywords used to search include “generative artificial intelligence” and its synonyms or acronyms, and “governance” or related issues. We limited to WoS categories: Education Educational Research or Education Scientific Disciplines. The result Search string is “generative artificial intelligence” or “GenAI” or “Generative AI” or “ChatGPT” (Topic) and governance or gover* or “AI Governance” or “AI Policy” or “Responsible AI” or “AI Regulation” or ethics (Topic) and Education Educational Research or Education Scientific Disciplines (Web of Science Categories). This search resulted in a total of 195 documents, whose primary information about data is shown in Table 1.

Six review articles were identified that can be considered as starting points for future studies in some subtopics: Use of AI Chatbots among Students (Schei et al., 2024); AI Ethics in Medical Education (Weidener and Fischer, 2023); basic concepts of AI and Gen-AI (Kalota, 2024); Legal and ethical considerations (Cornwall et al., 2025), Teacher Professional Development (Brandao et al., 2024), and AI Chatbot authorship from the perspective of copyright law (Lee, 2023).

The screening process applied inclusion and exclusion criteria to select articles for the sample. Inclusion criteria select relevant articles for analysis; exclusion criteria remove unrelated ones. We decided to include all document types published in peer-reviewed journals; we also considered proceedings papers, because these publications reveal the research front in this emerging theme. After removing one duplicate, 194 articles were selected for bibliometric analysis.

The analysis is done with two techniques: bibliometrics and content analysis. To produce bibliometric maps, it was chosen two open-source software. VOSviewer gives an overview of our research field (Generative AI Governance in Educational Research), and Bibliometrix reveals the conceptual structure of that field.

VOSviewer is a software tool for creating and visualizing maps based on network data, particularly bibliometric networks (van Eck and Waltman, 2020). Constructing a map involves three steps: a similarity matrix is calculated based on the co-occurrence matrix, a map is constructed by applying the VOS mapping technique to the similarity matrix, and finally the map is applied three transformations (translated, rotated, and reflected) to ensure consistent results. This process ensures that the distance between two items reflects the strength of the relation between the items (van Eck and Waltman, 2010). By default, VOSViewer assigns each term (e.g., author keywords) exactly to one cluster. The cluster technique used is discussed in Waltman et al. (2010).

To analyze the conceptual structure we use Bibliometrix which offers functionality to generate a thematic map (Aria and Cuccurullo, 2017; Cobo et al., 2011, Callon et al., 1991) with the following options: Keyword Plus, author keywords, unigrams, bigrams, and trigrams from titles or abstracts. The thematic map is predicated on the assumption that keywords represent concepts, the density and centrality of which can be utilized for categorization and conceptual mapping in a two-dimensional diagram. The thematic map functionality was based on the concepts of strategic diagrams. The steps required to reveal the conceptual structure of a scientific field are succinctly illustrated in Figure 3.

Detected communities can be represented by degrees of relevance (Callon’s centrality) and of development (Callon’s density) (Callon et al., 1991). Callon’s centrality (CC) measures the intensity of links between a given community and others, representable as a measure of a theme’s significance across the entire corpus. Callon’s density (CD) gauges the internal strength of the community, representable as a measure of the theme’s development. Utilizing these two measures, research themes can be mapped onto a two-dimensional strategic diagram with four quadrants: (1) upper right quadrant: motor themes; (2) lower right quadrant: basic themes; (3) lower left quadrant: emerging or disappearing themes; (4) upper left quadrant: niche or highly specialized.

The analysis with VOSViewer gives an overview of the research field and reveals the most used terms by authors (van Eck and Waltman, 2010). The authors’ keywords were restricted to at least two occurrences, which yielded a total of 96 keywords (out of 625 total author keywords). From those 96 keywords, VOSViewer calculated the total link strength of the co-occurrence links with other keywords. We present the network with 50 keywords with the highest total links (Figure 4). In the Figure, the size of a bubble is directly proportional to the number of publications that contain the keyword analyzed. Usually, terms co-occurring often tend to be located close to each other.

In this network, the 50 author’s keywords are divided into eight clusters. The most frequent keywords are AI (106), ChatGPT (100) and GenAI (58) and they are in the center of the map. It seems that clusters 1, 4, and 7 are more linked to Ethics, clusters 2, 3, 6 to Education, and cluster 5 to both.

In cluster 1 (red) there is the connection between Gen AI (58) and Ethics (43). In cluster 4 (yellow) there are two areas, Medical and Ethics in AI, with terms: academic integrity (18), medical education (13), machine learning (6), and OpenAI (6). In cluster 7 (orange), the focus is on plagiarism (5) and student perceptions (5).

The cluster 2 (green) shows terms like higher education, K-2 (elementary and secondary education), teaching, and learning, thus is related to Levels in Educations. In the cluster 3 (blue) the keywords are related to Education Technology: chatbot (18), education technology (9), and GPT (4). The cluster 6 (light blue) display terms related to practical education: education (26), assessment (4), pedagogy (3). The cluster 5 (purple) the focus is on AI (106)/ChatGPT (100), and its relationship with Ethics (AI ethics—10), and Education (AI education—7).

This overview shows how this topic is multidisciplinary and complex; next, we use Bibliometrix to deep understanding how themes are organized and have a thematic big picture.

We use Bibliometrix to build a thematic map (Strategic diagram) from 194 documents retrieved in the WoS database (min cluster frequency (per thousand docs): 20; clustering algorithm: Leading Eigenvalues). Thematic map is display in Figure 5.

The R-Bibliometrix package allows the co-occurrence network of keywords and their relationship with the thematic map to be obtained. The words can be extracted from the following metadata: indexing keywords, author’s keywords, unigrams, bigrams or trigrams of titles or abstracts. In this work, we used the author’s keywords. The words represent concepts whose density and centrality can be used in categorization and conceptual mapping in a two-dimensional diagram. The detected communities can be represented by two measures: the degree of relevance (Callon centrality) and the degree of development (Callon density) (Callon et al., 1991). This thematic map define four quadrant; based on Callon et al. (1991) and Aria et al. (2022), a brief description is carried out of each quadrant where we can observe the localization of the themes (see Figure 5).

The clusters/themes in Basic and Motor quadrants have high centrality. These clusters are central to the general network, so they are relevant. The difference between the quadrants is the density. In the Basic quadrant, the internal links density is relatively low, the two clusters/themes (cluster 1—AI; cluster 6—chatbot) are basic and transversal to the research area. In the Motor quadrant the density is high, so the two clusters that belong to this quadrant, Higher education (cluster 7) and Large Language Model (cluster 4), are core themes and well-developed. Notice that GenAI cluster/theme is located on the border between the motor and basic theme quadrants; we can classify them as Basic/Motor frontier.

The clusters in the other quadrants (Emergent or Declining and Niche) are not central to the research area. In the Emergent or Declining quadrant, the clusters have low density, so they have low relevance and development. They are on the margins of the network. We consider that AI ethics (cluster 9) and Education (cluster 8) are in the initial stages of development (emerging issues) because is very recent the field under study. In the Niche quadrant, the clusters Research ethics (5), Medical education (3), and Skills (10) are also peripheral but have high internal density. Those clusters can be classified as specializations because of their weak interaction with other clusters while being well-developed. Cluster 11 (AI literacy) falls within the Basic/Emerging frontier.

The results of the bibliometric analysis allow us to proceed to a more in-depth analysis. At this stage, we start by presenting one table for each quadrant (Basic, Motor, Emerging, Niche) with the respective clusters, the frequency of author keywords, and examples of publications. Two clusters are on the frontier between one quadrant and another, so we had to present two more tables (Basic/Motor frontier and Basic/Emerging frontier). The choice of these publications results from reading the title and abstract of each of the 194 publications. Following these procedures allowed an organized reading of each of the selected publications, but with interconnection and coherence. Thus, the complexity of the theme acquires an internal organization that facilitates understanding and reflection.

We used computer-aided text analysis (CATA) tools to organize and manage data, code bibliographic categories, and analyze the content of key publications (Costa and Amado, 2018). Defining categories is an essential stage in content analysis (Saldaña, 2021; Krippendorf, 1980; Neuendorf, 2017). Categories can be created by reading the data and/or a priori reading (deductive model and the open/inductive model). The method for building the category system, including main categories and subcategories, combines both inductive and deductive approaches. We began by considering Category1: ontological levels, with 3 subcategories: Macro, Meso, and Micro.

Ontological levels refer to different layers of reality or existence, often organized hierarchically based on their complexity, dependence, or fundamentality. The concept is used in philosophy (especially metaphysics), systems theory, and some areas of science to explain how different kinds of things exist or relate to one another (Tahko, 2021).

We can organize publications on an ontological level and classify them in those levels; this close approach is deductive. Classifications allow to assign a “descriptive label” to an entire document, such as an article. But, when we read each article, new categories may emerge that were not foreseen in the theory, such as GenAi Literacy (Annapureddy et al., 2024) or explainable AI (Sharma et al., 2024); in this case we are taking an open or inductive approach.

The quality assessment of these publications is structured according to how they can contribute to answering our review question and how they can structure the construction of the GenAI governance model. We read those publications and present the most relevant content of each. This presentation has two components: six tables and the main ideas captured from selected publications.

This GenAI Governance model must align with those levels and have a dynamic and living process (UNESCO, 2023; OECD, 2024). GenAI Governance model must be living because all the components are interconnected and GenAI has a lifecycle nature that evolves in response to users’ needs and contextual changes (Mahajan, 2025; Zhang D. et al., 2024). We build a Living GenAI Governance Model as a synthesis and as a product of this scoping review (Figure 7). This is a living model because it requires an adaptive and iterative approach that surpasses traditional static regulatory support in addressing emerging risks and challenges associated with technological development (Taeihagh, 2025).

This model considers multiple stakeholders and actors involved in tasks and processes that need to address ethical, regulatory, and operational challenges. The key factor is the ‘scale’ or ‘granularity’ of the analysis, that gives the contextualization where GenAI are used. This is divided into three levels: micro, meso, macro. Micro refers to individual actors or team projects. Macro refers to the system, like Global or Big Regional Geographical scope. Meso occupies the space between, referring to institutional, organizational and communities, like Universities or National Science Funding agencies. This model is a possible representation of a multilevel phenomenon, such as the use of GenAI.

This multi-scalar approach defines social spaces where individual and institutional agents have open possibilities, and causation flows from any of the interacting local, national and global scales. It should be noted that this multi-scalar phenomenon can have a top-down approach, if we consider that there are supranational recommendation documents (macro) that aim to harmonize the responsible use of GenAI, which can be adopted and adapted by national institutions and universities (meso), which in turn will have an impact on individual behavior (micro) where individual and institutional agents have open possibilities and causation flows from any of the interacting local, national and global scales. Another approach can be bottom-up if we consider that the actions of individuals and their interactions are the starting point. Individual attitudes, intentions, motivations, and behaviors are determinants of the results of the use of GenAI, both at the individual level, at the team level, at the network level, and at the institutional level. The various stakeholders are part of this system, but the different spaces or arenas are interconnected. For example, individuals should identify gaps and seek to acquire skills to use GenAI responsibly and effectively, but institutions should publicize their codes of ethics and provide training in the use of GenAI for the various user segments (Gupta et al., 2007; Marginson, 2022; Moresi et al., 2020; Pinho, 2021; Bozkurt, 2024b; Jobin et al., 2019).

The model was created based on the review of existing literature and an organized examination of concepts related to the subject. It is a framework that can be refined in future systematic reviews. Additionally, it will serve as a structured approach for analyzing the articles and documentation. To better clarify the description of this model, in the following section we will apply it to the context of Education Research.

GenAI Governance refers to choosing structures and mechanisms that can influence the processes of build, implement and monitoring GenAI responsible use, looking the interrelation between micro, meso and macro levels, with a Human Centered strategic focus. By taking this holistic approach, we can look at each component but also the integration and alignment of dimensions.

We choose to apply this governance model to a specific context: Education research. This is a particularly space where knowledge mobilization can ensure that education systems are better equipped to integrate evidence into policy and practice (OECD, 2025).Some authors refer to the need to build a holistic pedagogical framework, from the Human-Centered Artificial Intelligence perspective (Anastasiades et al., 2024).In times of misinformation schizophrenia, where technocracy overthrows politics, the value of research and evidence structured in ethical values and respect for Human Rights becomes even more pressing.

With this background we decide articulate and organize the topic according to the macro, meso and micro levels. At Macro-Level, several international organizations, including the European Union (2024), with Artificial Intelligence Act, the OECD (2023b,c), and UNESCO (2023), make global efforts on responsible development and use of AI or GenAI.

The Harmonized GenAI Framework (H-GenAIGF) identifies four constituents, fifteen processes, and nine principles essential for the global governance of GenAI, emphasizing risk-based approaches for better process coverage (Luna et al., 2025).

In the researcher’s specific context, starting with broad international regulations, such as the European Commission’s Living guidelines on the responsible use of generative AI in research (Bockting et al., 2023) aligned with the European Code of Conduct (ALLEA, 2023), and then narrowing down to specific national, institutional and local standards to ensure compliant and ethical use.

At Meso-Level we can locate National (Khanal et al., 2024) and Institutional stakeholders (Dai et al., 2024). Institutions must develop flexible and iterative policy frameworks that can adapt to the evolving nature of GenAI, ensuring they address ethical concerns such as bias and privacy (Lamberti, 2024). Engagement and continuous training for policymakers and public engagement initiatives are crucial for fostering inclusive AI policymaking. This author advocates for enhanced transparency, accountability mechanisms, public engagement initiatives, and continual adaptive frameworks to address ethical considerations and algorithmic biases in AI applications within policymaking.

Many educational institutions lack specific guidelines for the ethical deployment of AI tools, highlighting the need for overarching policies that address privacy and algorithmic transparency (Ghimire and Edwards, 2024).

A concept introduced and loosely defined as “proficiency in understanding, interacting with, and critically evaluating generative AI technologies”, which “entails not only knowing how to use AI-driven tools but also understanding the ethical considerations, biases, and limitations inherent in such systems” (Bozkurt, 2023). GenAI literacy is an important topic (O’Dea et al., 2024; Annapureddy et al., 2025).

The Meso level, institutions includes teachers, students, researchers, and other stakeholders who need to identify the main skills and competencies that they must acquire by formal training or self-training, in university space in the presence of GenAI (OECD, 2023a; Chan, 2023).

Emphasizing interpretability and sustainability in AI applications, at the micro level can help mitigate risks associated with GenAI use in education (Chan and Tsi, 2024; Lamberti, 2024). At Micro-Level, at Research context, the focus is Responsible Use of GenAI (Smith et al., 2024). In the GenAi era, possessing a diverse set of skills is highly advantageous for researchers (European, Directorate-General for, and Innovation et al., 2022). ResearchComp can serve as a starting point to identify and promote research-specific skills, as well as transversal skills that enhance quality research, by offering training and learning opportunities (European Union, 2024). Skills or competencies are strong themes for exploring effectively GenAI (Borneo et al., 2025; Costa et al., 2024).

At the micro level, individual initiatives aimed at developing skills in critical evaluation and responsible use of generative AI technologies are integral to bridging gaps in understanding. Researchers, educators, and students are encouraged to cultivate GenAI literacy, which extends beyond basic operation to encompass ethical considerations, bias recognition, and system limitations. This literacy serves as a foundation for both individual agency and collective innovation, ensuring that these tools are wielded in ways that align with broader educational and ethical goals.

At the micro level, there is literature about perceptions about the use of GenAI. The Dunning-Kruger Effect is a cognitive bias where individuals with limited knowledge or skill overestimate their competence (Figure 8). This overestimation arises because their lack of knowledge prevents them from understanding it. As people improve competence and skill, they develop an understanding of how much they don’t know. As such, confidence decreases and then increases as skill and awareness are gained (Kruger and Dunning, 1999). An author focus on negative effects of GenAI on researchers, such as on skill erosion (Giray et al., 2024). A collaborative auto-ethnography explored a non-technology lecturer’s first encounter with GenAI, analyzed through the 2023 Gartner Hype Cycle for Emerging Technologies (King and Prasetyo, 2023). A revised Dunning-Kruger effect was applied to using ChatGPT or other Artificial Intelligence (AI) in scientific writing (Salvagno et al., 2023). Initially, excessive confidence and enthusiasm for this tool’s potential may lead to the belief that producing papers and publishing can be accomplished quickly and effortlessly. Over time, as the limitations and risks of ChatGPT and other AI technologies are understood, along with the complexity of their operation requiring specific prompts, enthusiasm and confidence may wane. With increased awareness, ChatGPT and other AIs can become effective and supportive tools in scientific writing, akin to computers and internet search engines, ultimately leading to conscious and correct usage (Salvagno et al., 2023).