You’re listening to “Governance of Generative Artificial Intelligence: A Contemporary and Institutional Perspective,” by A.K.M.K. Hasan. Published in 2025.

Abstract.

The field of knowledge management and knowledge management systems is evolving and dynamic. In the era of developed information technology systems, the dynamics of knowledge creation and dissemination have also changed. Generative artificial intelligence (GenAI)—an embedded entity in the knowledge management system—has become a prominent area of research nowadays, while accountability, transparency, and ethics are common research agendas in institutional economics related to GenAI. The research in this paper has investigated the convictions behind GenAI adoption and how to develop a GenAI governance framework. The research adopts a qualitative approach to investigate the problem and surveys undergraduate students to explore their motive for using GenAI. The study sheds analytical light on institutional economists’ view on the governance of GenAI.

The study has found a positive relationship between perceived benefits and the adoption of GenAI in education by students. The theoretical model will have a considerable impact on the ongoing debate on the governance of GenAI and knowledge management systems.

INTRODUCTION.

Knowledge management (KM) and Knowledge management system (KMS) refer to the application of knowledge and a system created to facilitate the capture, storage, retrieval, and reuse of knowledge, respectively. Information and communication technology plays a great role in designing a KMS. Once the advanced machine learning system was adopted in the KM process, there were massive revolutionary changes in KM and KMS. After McCarthy coined the term artificial intelligence (AI) in 1956, AI became the most used term in contemporary KMS and machine learning research areas. The emerging usage of AI raises a valid question about how to ensure better usage of AI or create responsible AI. AI is broadly used in almost every sphere of life, including managing big data, microdata, games, online shopping and advertising, web search, digital personal assistants, cybersecurity, healthcare, and education.

Although there is no specific academic definition of AI that is used for academic purposes, the literature on AI has existed in academia for around three decades. In this paper, we attempt to focus on the uses of generative artificial intelligence (GenAI) in the education industry and how to govern GenAI once it is used in the education sector by students and researchers. In addition, our main concern is how to ensure quality information and integrity in GenAI for the education industry. Contemporary research shows evidence that GenAI can enhance knowledge sharing and organizational learning in academic contexts. Wu and Wan (2025) show that personalized test papers prepared by AI help students grasp programming knowledge. Qiu and Fang (2024) find evidence that a computer-aided instruction system is effective in improving students' performance on exams.

In addition, in business, AI could help to reduce human cognitive biases during recruitment in corporations. In short, the role of GenAI in KM is increasing significantly.

The term GenAI was first adopted by a number of scholars in the early 2000s. GenAI focuses on creating new textual and multimodal content using large language models, art-based models, and video-based models. Examples of well-known GenAI tools include OpenAI's ChatGPT, GPT-4, Playground, DALL ·E 3, and Sora; Anthropic's Claude; Google's Gemini (previously Bard); Stability AI's Stable Diffusion; and Runaway's Gen-2. GenAI allows users to acquire quick output on text, images, code, and other forms in an ordered manner so that users quickly get ready-made information. The development of GenAI is so quick that the answer to a specific question may rapidly change over time. It quickly adopted the changes that occurred in the real world.

However, how to develop a governance system for GenAI is still a topic of debate among the various stakeholders, including industry experts, academicians, and policymakers. Within this paper, we articulate the argument about the extent to which GenAI could be allowed for academic purposes and how to ensure the information provided by GenAI can be validated or authenticated. There is a chance for GenAI to create disinformation. For instance, as long as academic users and consumers (mainly faculty and students) are gradually dependent on GenAI for academic purposes, it is difficult to judge the trustworthiness and integrity of the data and information provided by GenAI until it is checked by a third party or reviewer. We have observed that, although every student uses the same GenAI tools, they get different outputs and information and receive different grades on exams as well.

Similarly, academics use GenAI like ChatGPT to reshape their research work; some of their scientific work is published, and some is not. Why does this happen? To understand this, we conducted a survey among the students to understand their views and opinions about GenAI. In other words, what leads them to adopt GenAI for academic purposes? In the second part of our research, we conducted an institutional analysis of the GenAI governance framework. Finally, we proposed a new model for the governance of GenAI to ensure data quality while adopting agency theory. Our findings help to understand the GenAI governance structure from an institutional economics lens and provide new evidence in the debate about the GenAI governance model.

The existing literature has investigated the determinants of GenAI adoption in the education sector in different countries, such as China, Pakistan, Sri Lanka, and Bangladesh. However, the identification of factors that affect the usage of GenAI among Central Asian countries, such as Uzbek students, has not been addressed yet. Hence, our sample size for the survey is based on Uzbekistan. In addition, the governance system of GenAI is a fiery issue in the literature of KMS. This study attempts to address both issues through this research.

To explore the above-mentioned research objectives of the study and the research gap, we raise the following two research questions:

• RQ1: What convictions constitute the frame of the GenAI governance system?

• RQ2: What are the key themes and models in GenAI governance, especially in the education sector?

The paper is organized as follows: the next section outlines a literature review regarding the theories of technology adoption and contemporary developments in AI regulations. Then we describe the adopted methodologies and data. The next two sections report the results of the empirical investigation and discuss the study’s empirical findings. In the discussion section, we describe the key contemporary themes of the governance framework of GenAI and offer a model to contribute to the existing debate. The final section provides brief implications for the study, a few limitations of the study, and areas for further research.

LITERATURE REVIEW AND CONTEMPORARY DEBATE

KM principles are well articulated by contemporary scholars. For instance, DeLone and McLean (2003) have demonstrated that the success of an information system depends on three things: system quality, knowledge quality, and service quality. These three factors are crucial for the intent to use a specific system. Mohammed (2022) has found similar evidence while experimenting with DeLone and McLean’s (2003) information system model in 13 Iraqi private universities in 2020. In an empirical study, Chiu et al. (2023) identified four challenges in knowledge retention: coordination complexity, insufficient resources for knowledge retention, insufficient attention to knowledge retention, and slow staff replacement and handover processes. Therefore, in the case of KM and KMS, the intention to adopt the KM matters.

Many social science theories have been developed to explain the factors that affect the adoption of technology. For example, the technology acceptance model, the theory of reasoned action, the theory of planned behavior, the innovation diffusion theory, the model of personal computer utilization, the unified theory of acceptance and use of technology (UTAUT) theory, the UTAUT 2 theory, and so on. While framing the UTAUT model, Venkatesh et al. (2003) integrated all the above-mentioned theories with their core constructs, conducted comparative descriptions, and then integrated the main influencers of the above-mentioned theories into the UTAUT model. Hence, once we adopt the UTAUT model, it will cover the theoretical concepts of the previous theories on technology adoption.

According to the UTAUT model, there are four key constructs that affect the decision to adopt technology: performance expectancy, effect expectancy, social influence, and facilitating conditions. The four moderators of gender, age, experience, and voluntariness affect the decision as well. Venkatesh et al. (2003) have extended the UTAUT model to the UTAUT 2 model, where they incorporate more variables like hedonic motivation, price value, and habit in order to enhance the explanatory power of the UTAUT model. Several investigations have been conducted to understand the individual benefits that support the adoption of information technology in the education sector using the UTAUT and UTAUT 2 models.

For example, Nawaz et al. (2024) conducted a detailed study on the acceptance of ChatGPT by undergraduate students in 17 state-owned universities in Sri Lanka, and their sample size was around 500 students. They found that using personal incentives and behavioral intentions are the prime factors for students using ChatGPT. However, a debate exists in academia about whether the UTAUT 2 model really explains the behavior of technology acceptance, as the theory has a scale of measurement bias and accuracy and validity of issues. In short, this study employs the “perceived compatibility” construct from the UTAUT model, whereas it employs the “perceived benefits” construct from the UTAUT 2 model. We attempt to generalize the theory while taking only two constructs and testing our hypotheses as valid in the case of Uzbekistan, while adopting GenAI for undergraduate students.

Based on the above discussion, in this study, we hypothesize that:

• H1: There is a significant positive relationship between perceived benefits and the adoption of GenAI. If so, it signals that there is a need for governance in GenAI because the information is sorted automatically by machine learning and big data analytics.

• H2: There is a significant positive relationship between perceived compatibility and the adoption of GenAI. If we accept this hypothesis, then there is a need to have a check and balance system in GenAI, as it could enhance knowledge transformation.

What is more, there is still debate on how to govern AI. Scholars argue that the development of sustainable AI depends on identifying the key factors for knowledge transfer, and ruling those through strict guidelines could be a good option. In the real world, several developed and developing countries have adopted AI strategies. It is worth noting that we have to understand the real flavor of “what is happening in the world now.” Hence, it is worth quoting an excerpt from a recent report published by the World Bank (2024) regarding the state of AI governance within the European Union (EU), the United States, Japan, and China.

The European Union has opted for a structured and risk-based legislative framework, with the AI Act proposing exhaustive regulations governing AI applications across diverse sectors. This act classifies AI systems based on their associated risk levels and imposes corresponding obligations, focusing on mitigating potential harm. While lower-risk AI systems, like spam filters, are subject to minimal transparency requirements, high-risk systems, prevalent in sectors such as healthcare, must comply with stringent obligations before market placement. Additionally, the act prohibits AI systems that endanger safety and fundamental rights, like real-time biometrics and predictive policing. However, extensive and costly compliance measures have raised concerns among businesses about the impact on firms’ global competitiveness and productivity levels.

The European Union’s push for similar AI regulations in Asian countries has been met predominantly with a more cautious “wait and see” response. The United States has adopted a more diverse and flexible approach to AI regulation, characterized by a combination of soft law, self-regulation, responsible use, and legislation at various levels within different domains. Federal initiatives include the AI Bill of Rights, an AI risk management framework by the National Institute of Standards and Technology, and plans for a national AI research resource aimed at enhancing public access to AI infrastructure. Various federal agencies are formulating road maps and best practices for AI within their domains, addressing potential discrimination and other issues stemming from AI systems. At the state level, numerous laws related to AI use and protection are being either proposed or enacted.

For instance, California’s proposed act will allow citizens to opt out of AI systems, while New York City’s local law mandates transparency in AI use during hiring processes and annual bias assessments. Several other states are either enacting or preparing AI-related legislation, and the U.S. Congress is anticipated to pass the Algorithmic Accountability Act of 2022.

Japan is employing a soft law approach to AI, focusing not only on minimizing AI-related harm and stimulating economic growth but also on harnessing AI to achieve societal objectives such as human dignity, diversity, inclusion, and sustainability. Rather than imposing rigid obligations or prohibitions, Japan’s AI strategies emphasize maximizing AI’s societal benefits through a flexible, risk-based, and multistakeholder approach. The country has promulgated the Social Principles of Human-Centric AI, which seeks to realize these values through AI use without imposing undue restrictions.

There are extensive legal constraints on AI, with the Ministry of Economy, Trade, and Industry advocating for a non-restrictive, agile governance structure that respects voluntary governance efforts by companies, offers nonbinding guidelines, and fosters multistakeholder dialogue. China has developed a diverse regulatory framework, focusing on both hard and soft law, technology-specific AI regulations, draft rules for GenAI, guidelines for AI in various applications, and specific provisions regarding automated decision-making. The country has introduced rules to enhance consumer protection and maintain competition, with special provisions on biometric data privacy, albeit with exceptions for national security and law enforcement needs.

The Cyberspace Administration of China and the Ministry of Science and Technology are crafting regulations and guidelines, emphasizing the responsibility of providers of GenAI products and services for the content generated (section: AI governance principles and divergent regulatory trends, page 97).

Here we can summarize that in the real world, AI regulations can be categorized by the concepts of “hard law” and “soft law”, and “self-regulation,” alongside the “risk-based approach,” “technology-specific regulatory approach,” and “responsible-use approach”.

In a word, the governance of AI is still undergoing research. For instance, Memarian and Doleck (2023) have conducted a systematic literature review on fairness, accountability, transparency, and ethics in AI and higher education, where they articulated 33 studies from Scopus and Web of Science-referred journal articles between 2015 and 2023 focusing on this theme. According to Memarian and Doleck (2023), the reviewed journals used the term accountability to refer to quality assurance in education while using big data by AI, regulating the AI programs in higher education, and pedagogical changes in the classroom while using AI. While describing the AI’s accountability, some scholars divided it into two parts: criminal accountability and civil accountability.

Bearman et al. (2022) advocated building up a relationship between AI and the teaching and learning environment rather than considering AI as simply an innovation. Some scholars argue for the emphasis on big data governance from an external accountability and internal accountability perspective. Recent versions of the EU AI Act mentioned four types of risk that AI poses and argued that AI systems should be regulated to address such kinds of risks. These are unacceptable risks, high risk, limited risk, and low and minimal risk. In addition, regarding GenAI, the EU Act has further identified three basic concerns about the uses of GenAI: disclosure of the content produced by GenAI, preventing illegal content through GenAI, and misuse of copyrighted data. Hence, it is urged to regulate and control GenAI because of its opacity and multifunctionality.

However, there are a few studies to clearly identify the governance model of GenAI and its regulating procedures. Scholars argue about the responsible governance of AI and GenAI. Triguero et al. (2023) referred to the fact that responsible governance includes the auditability and accountability of AI systems. Whereas Wirtz et al. (2022) mentioned that governance should be focused on regulation, and they stress a focus on the governance system of AI from a four-layer perspective. These are the AI risk layers, the AI risk management and guidance process layer, the AI guidance layer, and the AI governance layer. According to Wirtz et al. (2022), the governance layer of AI governance mainly depends on regulatory guidelines, which are based on six pillars. To sum up, there is a debate in existing literature about the extent to which GenAI should be regulated.

And if it is regulated, which model of governance system would optimize the benefits of GenAI? Our research contributes to the existing debate by adding new evidence from institutional analysis of GenAI (the code of conduct) and providing a new model for the governance of GenAI.

DATA AND METHODS

The framework of the study is constructed based on several kinds of literature and theories related to AI and technology adoption. To explore our first research question (the convictions constitute the frame of the GenAI governance system), we follow the interpretations based on the UTAUT and UTAUT 2 models. We consider two independent variables and one dependent variable to predict the adaptation of GenAI. The model is shown in Figure 1.

To verify our hypotheses that were developed in the previous section, we conducted a survey among university undergraduate students regarding the uses of GenAI for their academic purposes. To explore our second research question, we discuss conventional governance theory (agency theory) and its presumed relation to GenAI. In addition, we shed analytical light on the contemporary GenAI governance model, the EU Act 2023, and the hourglass model coined by Mäntymäki et al. (2022). Based on this theoretical and institutional analysis and the primary data retrieved from surveys, we propose a model for GenAI governance, especially to ensure quality and trustworthy data.

To identify the determinants of the adoption of GenAI, we have opted for a quantitative approach using structural equation modeling under the partial least squares (PLS) approach. The questionnaire items used in this study were drawn from previous empirical studies. Data from this research includes three latent variables, namely perceived compatibility (three items), perceived benefits (four items), and adoption of GenAI (two items). To measure each variable, we adopted a five-point Likert-type scale. The agreement options are from one (strongly disagree) to five (strongly agree).

The questionnaire (see appendix) was conducted in a single stage within two days. It was structured into two parts. The first part focused on collecting data regarding the sociodemographic profile (education level, gender), while the second part measured the three latent variables of the current study (perceived compatibility, perceived benefits, and adoption of GenAI). Printed forms were distributed among the students, and a total of 113 usable responses were collected from 65 male students (57.52%) and 48 female students (42.47%), followed by educational level 4 (49.44%) and level 5 (50.44%), respectively. All forms were input into an Excel sheet, and there were no incomplete forms; hence, all forms were retained for the analytical phase.

We used the two-stage partial least squares structural equation method (PLS-SEM) advocated by Anderson and Gerbing (1988) on SmartPLS4 to analyze and interpret the collected data. This model provides a high level of statistical power when the data size is small. The first stage is to develop a measurement model using confirmatory factor analysis. It shows the structural relationship between latent constructs and their indicators. It is also used to assess convergent validity and discriminant validity. Next, the inner model is carried out while checking a number of criteria like the coefficient of determination, the effect size, the predictive relevance, the goodness of fit of the model, and to test the hypothesis of our research. There were no ethical issues related to the questionnaire survey, as all respondents were kept anonymous.

The detailed steps followed in the SmartPLS to conduct the study are shown in Table 1.

Note. CR = composite reliability; AVE = average variance extracted.

The outer model is a measurement model used to assess the reliability and validity of the constructs. Convergent validity shows that the variables studied really reflect the latent constructs that they are designed to measure. Factor loading represents the strength of the relationship between indicators, that is, observed variables and their corresponding latent constructs. Normally, factor loading over 0.70 is recommended. To check convergent validity, each latent variable’s average variance extracted (AVE) is evaluated. It should be 0.5 or higher. To check the reliability and internal consistency, composite reliability is used as a yardstick, and it should be 0.7 or higher.

According to Hair et al. (2022), discriminant validity verifies that a reflective construct exhibits stronger relationships with its own indicators than with those of any other construct in the PLS path model. Subhaktiyasa (2024) mentioned that “to check discriminant validity, the classical approach is proposed by Fornell and Larcker (1981) who suggest that the square root of AVE in each latent variable can be used to establish discriminant validity, if this value is larger than other correlation values among the latent variables” (p. 357).

The inner model is also known as the structural model. The inner model refers to the component of the overall model that focuses on the relationships between latent variables. The inner model in PLS-SEM tests hypotheses about the relationships between constructs or latent variables. The structural model in PLS-SEM consists of the paths or arrows representing the hypothesized relationships between the latent variables. These paths indicate one construct's directional influence or impact on another. The instruments used in the inner model include R2, F2, Q2, beta value, t-value, and P-value.

RESULTS.

Assessment of the Outer Model

The outer model (measurement evaluation model) shows solid convergent validity according to various criteria such as loading, Cronbach’s alpha, composite reliability, and AVE. The value of the perceived benefit indicator’s outer loading is above 0.60 (Figure 2), which indicates its significant contribution to the model construct. The Fornell-Larcker criterion and outer-loading measures are considered conventional methodologies when evaluating discriminant validity. To establish discriminant validity using the Fornell-Larcker criterion, the square root of AVE of each construct should be greater than the correlation with any other construct in the framework, whereas to establish discriminant validity using the cross-loadings method, the outer loading of each item on its associated construct should be greater than the loading of items on other constructs.

Note. PC = perceived compatibility; PB = perceived benefits; AG = adoption of generative artificial intelligence; GenAI = generative artificial intelligence.

Table 2 shows the details of the validity criteria. For instance, the Cronbach’s alpha and composite reliability are above 0.729 and 0.733, respectively, for the perceived benefits construct, whereas the AVE is above 0.50 in all three constructs.

Note. AVE = average variance extracted; GenAI = generative artificial intelligence.

The Fornell and Larcker criteria are shown in Table 3. The square roots of AVE, 0.63, 0.652, 0.635 (from Table 2), are shown in Table 3, which are 0.794, 0.514, 0.399. The square root of construct variance (Adoption of GenAI) is 0.794, which is higher than its correlation with the other two constructs in the study (0.794 > 0.514, 0794 > 0.399). Therefore, Table 3 shows evidence that the discriminant validity is established in the study, as according to the Fornell-Larcker criterion, the square roots of the AVE values are greater than their correlations with the latent constructs.

Note. GenAI = generative artificial intelligence.

Tables 4 and 5 show the cross-loadings and outer-loading matrix of the variables. Cross loading is an approach that compares the outer-loading value of an indicator against its latent variable and the outer-loading value of the indicator against other latent variables. The outer-loading value of an indicator against its latent variable must be greater than the outer-loading value of the indicator against other latent variables. It indicates that an indicator has proven better at measuring its latent variable than other variables. In short, if an indicator loads significantly higher on a construct other than its intended construct, it suggests that the cross loading supports the discriminant validity. Outer loadings are used to assess whether indicators adequately measure the construct, which should be at least 0.5 or higher for indicators to be considered reliable.

Tables 4 and 5 show that the loadings of the indicators are greater than all their cross loadings, supporting the discriminatory validity of the outer model.

Note. GenAI = generative artificial intelligence; AG = adoption of generative artificial intelligence; PB = perceived benefits; PC = perceived compatibility.

continued on following page

Volume 21 • Issue 1 • January-December 2025

Note. GenAI = generative artificial intelligence; AG = adoption of generative artificial intelligence; PB = perceived benefits; PC = perceived compatibility.

DISCUSSION.

Structural Model Analysis and Testing the Hypothesis

According to Hair et al. (2022), structural models exhibit associations between constructs. To test the hypotheses, we used a bootstrapping procedure in SmartPLS 4.0. The coefficient of determination (R2) was also calculated to assess the variance explained in the outcome variable by the predictor variables (Table 6).

Note. GenAI = generative artificial intelligence.

The effect sizes (f2) of each exogenous variable were calculated (Table 7). In addition, the predictive relevance of the model was computed through the Q2 value (Table 8). Figures 3 and 4 show the inner model’s path coefficients and corresponding p-values and t-values, respectively. Table 5 displays the results from the structural model, including the path coefficient, t-statistics, p-value, and the ultimate decision about our hypothesis. It is a two-tailed test with a probability of 5% error and a t-value of 1.96, which is considered statistically significant (according to Hair et al., 2022).

Note. GenAI = generative artificial intelligence

Note. RMSE = root mean squared error; MAE = mean absolute error; GenAI = generative artificial intelligence.

The F2 values F2 ≥ 0.02, F2 ≥ 0.15, and F2 ≥ 0.35 represent small, medium, and large effect sizes, respectively.

Note. PC = perceived compatibility; PB = perceived benefits; AG = adoption of generative artificial intelligence; GenAI = generative artificial intelligence.

From the findings (Figures 3 and 4, Table 9), it turned out that perceived benefit (H1. Β = 0.421, t = 4.172, p < 0.00) has a positive and significant effect on the adoption of GenAI, and perceived benefit (H2. Β = 0.175, t = 1.655, p > 0) has a positive effect on the adoption of GenAI. However, it has no significant effect. This statistical result supports the first hypothesis (H1), whereas the second hypothesis (H2) is not supported by our result. The coefficient determinant (R2) is 0.286 (Table 6) for the adoption of the GenAI construct, which means that perceived compatibility and perceived benefit account for 28.6% of the variance in the adoption of GenAI in our sample.

Note. O = original sample; STDEV = standard deviation; H1 = Hypothesis 1; PB = perceived benefits; AG = adoption of generative artificial intelligence; H2 = Hypothesis 2; PC = perceived compatibility.

The perceived benefit has a medium effect size as its F2 value is 0.178, whereas in the case of perceived compatibility, it has no effect size (Table 7). Table 8 shows that the predictive relevance value (Q2) is greater than zero, which supports the predictive relevance of the model.

The empirical findings of the study reveal that at least perceived benefits positively affect the adoption of GenAI among our sample students. The result is consistent with the previous studies, which provided evidence that perceived ease of use positively affects the acceptance of ChatGPT among Sri Lankan undergraduate students. Emon et al. (2023) also found a similar result: performance expectancy has a positive effect while predicting the adoption of AI among ChatGPT users.

Therefore, our empirical analysis demonstrates a positive relationship between perceived benefits from GenAI and the adoption of GenAI. Hence, we accept our first hypothesis that there is a significant positive relationship between perceived benefits and the adoption of GenAI. If so, it signals that there is a need for governance in GenAI because the information is sorted automatically by machine learning and big data analytics.

Simultaneously, based on our empirical result, we could reject the second hypothesis that there is a significant positive relationship between perceived compatibility and the adoption of GenAI.

Once we accept the first hypothesis, we could consider perceived benefits as a conviction for the governance of GenAI, as once the end-users of GenAI realize their perceived benefits from GenAI, the necessity of governing GenAI will gain more emphasis. Hence, regarding our first research question, we could conclude that the perceived benefits from GenAI are one of the convictions to constitute the framework of the GenAI governance. Now, a rational question is how to govern GenAI. In the following sub-section, we discuss this query, which is also the second research question of this study: what are the key themes and models in GenAI governance, especially in the education sector? To answer this, we shed analytical light on the key themes of the GenAI governance model.

Contemporary Key Themes in the Governance of GenAI, Especially in the Education Sector

The World Bank (2024) addresses five areas of challenges that pose a threat to the governance of AI. These are national security and challenging dilemmas, regulatory challenges and innovation dynamics, regulatory fragmentation and arbitration, the adequacy of existing regulations, regulatory capture, and inclusive governance. In a recent survey of 1,300 U.S. household heads about GenAI, they expressed their concern about the risks of data breaches and data abuse and overwhelming support for the regulation of GenAI. These results reflect that, in advanced countries, both regulators and the public are concerned about the governance of GenAI. However, in the case of developing countries, the governance of GenAI is not yet formulated, although they are adopting AI strategies at the national level.

For instance, some low- and middle-income countries, including Egypt, introduced a national AI strategy in 2021; India and Brazil introduced their national AI strategies in 2018 and 2021, respectively. The Republic of Rwanda will approve its national AI policy in 2023. One of the missing points in all those so-called national AI policies is that there is no specific governance system proposed to monitor and guide GenAI in the policies, as we mentioned earlier. We discussed earlier in Section 2 that developed countries use several approaches to govern GenAI, but it is still missing. It is still developing countries' context as well as how to govern GenAI, especially when it is used in the education sector. While discussing the threats posed, the company Ernst & Young has identified the usage of GenAI, which includes design risks, performance risks, algorithmic risks, and data risks.

Regarding GenAI, especially for the education sector, data risks are a major concern, and there are two key challenges regarding data risks, that is, “misinformation” created by GenAI from data and “misuse” of GenAI data in the academic world (2Ms). In other words, we could say 2Ms are the prime challenges of GenAI, especially in the education sector. The term “misinformation” here refers to the information or data used by GenAI that is not true, and the term “misuse” refers to using GenAI while violating the guidelines for using GenAI published by the competent authorities. It is urgent to address these 2Ms because, in our previous section, we had empirical evidence that students are opting for GenAI because they perceive a benefit from it.

As long as the end-users are getting perceived benefits from GenAI, it is quite difficult to restrain them from using GenAI; instead, it is wise to free the GenAI data from the 2Ms. We could solve this problem within the lens of agency theory.

Agency theory points out that if a perceived inequity exists, agents are likely to engage in self-interested behavior. When the agent engages in self-interested behavior, information asymmetries are created where the principal is unable to properly monitor the agent’s behavior. Indeed, agency theory is based on two basic assumptions. First, information asymmetry exists between the principal and the agent. Second, the principal cannot monitor agents fully. The principal-agent problem could be resolved by establishing an effective governance mechanism that monitors and limits an agent’s self-serving behavior. We could align the agency theory concept with the GenAI governance framework. For instance, in the case of GenAI, we could assume that the national regulator of GenAI of a specific country is a principal or owner, and they need to regulate and monitor the GenAI output.

In contrast, GenAI could be considered an agent or manager who will provide information to the end-users, including the regulator (principal). As GenAI works as a machine learning process, the prime challenge is to confirm the accuracy of the information, which could create an agency problem among GenAI, regulators, and users. Such agency-related problems could be minimized in two ways: either the regulator filters the information through an information confirmation center or GenAI would establish self-censorship within itself. Establishing self-censorship inside GenAI itself is difficult, as we have to input real and true information into GenAI. Hence, the first option is to establish an information center to feed the information into GenAI, which could be the best option. We may call it a “verification center” within the GenAI governance framework (see Figure 5).

We suggest that a state regulator (principal) monitors the central database, and GenAI (an agent) could provide information only from the data provided by the central database (Figure 5). Once the data has been filtered through the database and fed up with GenAI, the output will go to the end-users. If we can follow this process, the level of misinformation or distortion of data could be reduced. It is presumed that it could help build trust and reduce misinformation in GenAI.

Note. GenAI = generative artificial intelligence.

This concept supports the Infocomm Media Development Authority (IMDA) of the Singapore Ministry of Communication and Information, which proposed a GenAI governance model in 2024, where several steps in data validation are advocated (which in our model we refer to as a central database/verification center) and describe a total of nine dimensions of the GenAI governance framework. This framework is considered the most contemporary and comprehensive GenAI governance framework. We summarize the nine dimensions for modeling the AI governance framework for GenAI in Table 10.

Note. AI = artificial intelligence; R & D = research and development. Adapted from Model AI governance framework for generative AI: Fostering a trusted ecosystem, by Infocomm Media Development Authority (IMDA) and AI Verify Foundation, May 30, 2024, (the linked source wp​-content/​uploads/​2024/​05/​Model​-AI​-Governance​-Framework​-for​-Generative​-AI​-May​-2024​-1​-1​.pdf). Copyright 2024 by IMDA and AI Verify Foundation.

According to the second dimension, to ensure good quality and representative data, IMDA and AI Verify Foundation (2024) also suggest establishing trustworthy use of personal data, balancing contribute to the data dimension to ensure data quality and trustworthiness in the GenAI governance framework.

In addition, the hourglass model of AI governance coined by Mäntymäki et al. (2022) advocates eight components in the operational governance of the AI system layer. These are AI systems, algorithms, data operations, risk and impacts, transparency, explainability, and contestability, accountability and ownership, development and operations, and compliance. Mäntymäki et al. (2022) pointed out that we must ensure that AI system data is sourced, used, and monitored in alignment with the organization’s strategic goals and values. In their model, the data operations component has four sub-components, which are shown in Table 11. They discussed each sub-component’s measurement sticks for data operations in their model.

Note. C = component; T = Task. Adapted from Putting AI ethics into practice: The hourglass model of organizational AI governance, by M. Mäntymäki, M. Minkkinen, T. Birkstedt, and M. Viljanen, 2022, arXiv (the linked source).

What is more, while describing the AI system life cycle, the Organisation for Economic Co-operation and Development (2019) referred to data verification and validation as a second phase of the AI system life cycle. However, neither they nor Mäntymäki et al. (2022) offered any specific model for data validation or verification. Our proposed model could contribute to understanding data operations better, verifying and validating the data in the GenAI governance debate.

CONCLUSION.

The current study aims to analyze the governance of GenAI from a contemporary and institutional perspective. We have raised two research questions in this study. First, we attempt to investigate what convictions frame GenAI governance. We developed two hypotheses to explore this query and conducted an empirical investigation. Relying on PLS path modeling, the findings emphasize that perceived benefits contribute to the adoption of GenAI by end-users. Hence, we conclude that ‘perceived benefits’ is one of the convictions for GenAI governance. Our second research question focuses on exploring what themes and models are used in GenAI governance. To answer this query, we have discussed contemporary models of GenAI governance and offered a new model for GenAI governance.

Our findings have at least two implications. First, our findings provide insights into why we need to govern GenAI. We argue that since the end-users are benefiting from GenAI, we need to govern it, mainly aiming to reduce misinformation and misuse of GenAI, especially in the sector. Although several academic institutions are restricted from using GenAI for academic purposes, a good number of universities are now allowed to use GenAI for academic purposes. Hence, we must make it clear and transparent to the end-users of GenAI to what extent they will depend on the use of GenAI for academic purposes such as doing research, preparing reports, and so on. At this level, policymakers should consider the issue of the growing number of end-users of GenAI since it provides incentives to the users (perceived benefits). Hence, it is urgent to provide GenAI with a transparent information channel.

Second, the study focused on existing GenAI governance models and considered an agency problem that exists in the GenAI governance framework. We observe that existing models cover mostly regulatory and ethical issues. However, the quality of the data is not measured in the existing models, that is, how to ensure the quality of the data in GenAI is underrated. We propose a new model that helps control the quality of data in GenAI. The study argues that once the regulator provides a database to GenAI or establishes a verification center to verify the information in GenAI, end-users will get more accurate information while using GenAI.

The limitations of the study could be sampling biases and industry biases. While the empirical investigation uses only undergraduate students in the population, it uses a small sample size. Hence, it is possible that the findings do not fully represent the overall picture of the GenAI governance framework. Future research could use more representative sampling techniques and conduct longitudinal research. Future studies could also be focused on how to establish an information verification center that will monitor the information center, whether the information center is a public good or not.

COMPETING INTERESTS

The author of this publication declares there are no competing interests.

FUNDING

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Funding for this research was covered by the author of the article.

PROCESS DATES

APPENDIX.