Towards a Sustainable and Ethical Integration of AI Chatbots in Higher Education

Abstract

This paper examines students’ perceptions of factors influencing normative support for the integration of AI Chatbots in universities, providing an empirical basis for developing institutional policies and implementation strategies in higher education. Framed within the sustainability perspective, the study examines how ethical, cognitive, and perceptual factors shape the long-term adoption of AI technologies in academic environments. Our study employs a structural model comprising 10 constructs, 46 items, and 9 hypotheses, tested on a sample of 408 economics students from Timisoara. The research identifies AI literacy as the most influential factor in the formal integration of these technologies in universities. The following factors have a direct impact: teacher perception, student perception, and cognitive risks (reliance on AI Chatbots and avoidance of intellectual effort). Use for personalized learning is a factor with a significant direct effect on positive perceptions and intentions to use AI Chatbots among students. Academic integrity risks, as well as limitations on accuracy and reliability, have no significant impact. AI Chatbots represent an essential opportunity to transform higher education. However, their positive impact is realized only through responsible formal integration, grounded in ethical policies, adequate digital education, and the adaptation of pedagogical practices. Universities must regard AI as a strategic ally for teachers and students, while keeping human interaction, critical thinking, and academic integrity at the centre of the educational process. The study argues that students’ perceptions are that universities must approach AI Integration as a strategic component of sustainable educational ecosystems, aligning innovation with long-term academic integrity and the objectives of sustainable development, particularly Sustainable Development Goal 4 (Quality Education).

1. Introduction

Sustainable digital transformation has gained increasing attention in recent years, reflecting the need to align technological innovation with long-term economic and sustainable goals (SDG 4). Thus, digital technologies can support sustainable development goals by increasing institutional adaptability across all sectors, including higher education.

A 2025 study of DataCamp says that “69% of leaders believe AI literacy is important for their teams’ daily tasks; 60% of leaders believe their organization has an AI literacy skill gap; 50% of organizations are planning employee training to increase GenAI adoption; 71% of organizations regularly use GenAI in at least one business function, up from 65% in 2024” [1]. A prior goal of universities has been to prepare students for the required work, and they are increasingly exploring the formal integration of AI. Also, AI literacy is being extended to pre-university education [2]. AI in higher education brings opportunities such as automating administrative processes, personalized learning, and monitoring student progress [3,4] but also risks related to equity, confidentiality, and academic autonomy [5]. A recent study shows expanded investment in AI tools to optimize institutional efficiency and student attention [6]. Depersonalization of the educational relationship [7], the destabilization of traditional teacher–student hierarchies [8], or the algorithmic risks of discriminating against groups of students [9] are concerning examples regarding the use of AI. The use of conversational AI raises acute academic integrity issues, including the risk of AI-assisted plagiarism and the difficulty of distinguishing between original and automatically generated text [10,11].

The literature suggests that AI is not a threat, but as a stimulus for pedagogical and institutional restructuring [5,12]. Sustainable integration requires clear policies, investments in teacher training, and the development of students’ ethical literacy. In this context, the article presents a study on integrating popular AI Chatbots (e.g., ChatGPT, Gemini, DeepSeek, Microsoft Copilot, Claude, Grok, Perplexity, etc.) in universities, focusing on the teacher–student relationship, the ethical risks generated by AI Chatbots, and the educational policy directions necessary for a balanced and responsible adoption. Thus, our study aims to provide answers to the following research questions:

RQ1: 

Under what normative conditions does AI become institutionally legitimate, in the students’ perceptions?

RQ2: 

What are the risks of the AI Chatbots integration?

Our study aims to show that in higher education, sustainable digital transformation involves more than the adoption of artificial intelligence tools. It involves developing AI literacy, adopting pedagogical practices, and implementing governance mechanisms that ensure innovation strengthens institutional resilience rather than undermines academic values.

This paper is structured in six sections. Section 1 introduces the topic. Section 2 provides a general background based on a literature review. Section 3 describes the methodology and data collection. Section 4 presents the results, and Section 5 presents discussions. Section 6 presents the conclusions, limitations, and directions for future research.

2. Theoretical Framework

2.1. Background to AI Integration in Universities

The broader concept of digital sustainability clarifies that sustainable digital transformation requires a long-term vision that integrates ethical safeguards and digital competencies. In this context, universities are complex ecosystems in which technological innovation must coexist with academic norms and equitable access to learning resources.

“Coursera was launched in 2012 with a mission to provide universal access to world-class learning. Today, it is one of the largest online learning platforms worldwide, with 197 million registered learners. Coursera partners with over 375 leading universities and industry partners to offer a broad catalogue of content” [13]. The Coursera course categories are: Artificial Intelligence, Business, Data Science, Information Technology, Computer Science, Healthcare, Physical Science and Engineering, Personal Development, Social Sciences, Language Learning, and Arts and Humanities. Currently, Coursera, like other online learning platforms, uses an AI Assistant (i.e., an AI Chatbot) in each course—this means that it is known that today, students use AI Chatbots in the learning process, on the one hand, and that they encourage this, on the other hand. Microsoft has integrated Copilot into the entire Office suite. This means that everyone who writes technical content in Microsoft Word will soon use the built-in Copilot, if they don’t already [14]. Google has integrated Gemini into Google Drive applications [15]. Moodle is increasingly incorporating AI solutions, including an AI Chatbot [16]. In this context, AI integration in universities contributes to long-term institutional and societal sustainability.

Integrating artificial intelligence into universities offers meaningful opportunities and significant challenges, reshaping teaching, learning, and university administration [17,18]. This change needs coherent academic and ethical guidelines for AI’s transparent, equitable, and trustworthy use to advance invention without compromising essential instructional codes. According to [17], the ethical integration of AI in education refers to the responsible design, deployment, and use of AI systems in educational contexts in ways that carefully regulate automation, promote equity and transparency, ensure meaningful human oversight, and protect learners’ autonomy, well-being, and fundamental rights, particularly those of vulnerable groups. Transparency in AI-assisted assessment, AI literacy among educators and students, and proactive mitigation of algorithmic bias are essential to ensure the responsible and inclusive use of AI.

One visible effect is the reconfiguration of the teacher–student relationship: teachers shift from being the primary source of knowledge to the moderator of learning and designer of the educational experience [3,6]. At the same time, AI Chatbots provide students with quick access to resources and explanations, increasing their autonomy, but at the risk of encouraging shallow thinking and reliance on automated solutions [19,20,21].

Implementing AI in universities is often undertaken without explicit ethical risk assessment policies [22]. There are multinational regulatory frameworks, such as [17,18], but their implementation is separate and voluntary [12]. Existing trends focus primarily on cybersecurity and data protection, overlooking pedagogical and epistemological dimensions [23]. Educators’ lack of involvement exacerbates these deficiencies [4].

Nowadays, the evolution of explicit procedures and guidelines for the use of AI in universities is a priority [24,25]. Ignoring or refusing AI is not a viable solution [25]. It is essential to develop a set of adjusted rules that categorize AI applications by risk and impose rigorous conditions on protection, transparency, rights, and responsibility [26,27,28,29].

For reliable management, it suggests an ethical observatory of AI in education. This observatory will investigate appropriate initiatives, deliver platforms for information sharing, and design strategic goals and ethical codes [24,25]. Policies must also handle difficulties such as individual privacy and the risk of discrimination [30].

Research reveals the threats of using AI Chatbots to develop high-quality argumentative texts without origins, promoting plagiarism and inappropriate attribution [25,31,32,33,34]. Universities must clearly distinguish between legitimate AI use and academic fraud [29].

Data privacy issues are essential. Policies must regulate how teacher and student data are collected and used through AI systems [24,25] to prevent information manipulation and privacy intrusion [35]. For example, ChatGPT was banned in Italy for privacy reasons [35].

AI Chatbots have captured the attention of universities, as modern components with a transformative role in education [36,37,38]. The specialized literature, however, notes a notable lack of robust empirical studies investigating the real impact of AI Chatbots on teaching and learning processes [25,39,40]. The absence of a solid evidence base makes it difficult to develop informed education policies and to deeply understand the implications of using AI in education [24,26]. For this reason, education policies should encourage greater research to provide a solid evidence base.

2.2. AI Literacy in Education. Changing Roles in Educational Activities

AI Chatbots have become a constant partner in learning, and artificial intelligence is profoundly transforming society and education [31,41]. In this context, using AI is insufficient; understanding it becomes imperative [30,42].

AI literacy is considered the new standard of digital competence. It is more than a technical skill: it represents a framework of thought through which students and teachers can critically evaluate results, recognize ethical limits and implications, and collaborate responsibly with generative AI systems [43,44,45]. This literacy involves the development of fundamental knowledge, skills, and values, including prompt engineering, critical fact-checking, and using AI as a starting point rather than a final authority [40,46]. Reference [40] says that AI literacy is essential for educational integration because: it’s inevitable—AI is already transforming education; knowledge gaps create resistance—lack of literacy breeds fear and rejection; AI requires new competencies—traditional tech skills are insufficient; AI is fundamentally different—it requires specialized understanding. The [30] main argument is “AI literacy is not just an educational objective but a vital life skill for the twenty-first century”. Reference [47] shows that AI literacy is a significant predictor of adoption. Quantitative empirical evidence [48] indicates that AI literacy—specifically, practical AI knowledge combined with data literacy—is not only beneficial but also essential and measurable for successful AI integration in education.

Hypothesis 1: 

GenAI literacy significantly predicts perceived normative support regarding institutional AI integration in higher education settings.

The rapid development of AI Chatbots is profoundly changing education and necessitating a redefinition of teachers’ roles [49,50]. AI can automate organizational and repetitive jobs, letting teachers concentrate on mentoring, examination, and monitoring the learning process [31,51].

This transformation of the teaching process requires developing new skills, including critically evaluating AI-generated content for accuracy, originality, and relevance [52,53]. The ability to effectively interrogate AI systems and to train students to formulate prompts and understand model limitations is becoming increasingly important, transforming them from passive receivers into active investigators [54,55].

Understanding and prior knowledge of AI technologies positively influence the perception of their usefulness [54]. Universities inspire the investigation into and adoption of AI through sustained initiatives within educational administration [31]. University professors favor the use of AI Chatbots in the educational process [46,56].

Most teachers agree that the effective use of AI Chatbots requires adequate training [27,57]. University professors believe AI Chatbots should be a complementary tool, not a substitute for interpersonal relationships in education [56,58].

Teachers are gatekeepers—their literacy determines adoption success [40,59] shows that when academic staff perceive AI Chatbots as enjoyable, human-like, beneficial, and socially endorsed, they develop positive emotions and greater willingness to integrate them into their teaching and learning practices.

Hypothesis 2: 

Teachers’ favorable perceptions of AI Chatbot use positively predict perceived normative support toward institutional AI integration in higher education institutions.

AI Chatbots adapt well to traditional assessments, but they remain vulnerable to emerging technologies, necessitating the transformation of assessment methods [49,60]. The specialized literature proposes a reform oriented towards higher-level thinking and authentic tasks, shifting the emphasis from questions “What?” to questions “Why?” and “How?” [54,61,62]. Knowledge testing can be helpful when using active pedagogical methods [54] or oral evaluations [62,63].

2.3. AI Chatbots as Educational Tools

AI Chatbots are a rapidly evolving technology reshaping society and education. They simulate human conversations and open new directions for learning and research [64,65]. Provides quick answers to theoretical questions, generates ideas for essays, provides feedback on drafts, and supports exam preparation [27,53]. They contribute to literature reviews, data analyses, and hypothesis formulation in research.

The capacity to process and examine information at a higher level, by quickly synthesizing large volumes of text, saving time and supporting focus on critical interpretation [66,67]. They function as partners in the creative process, providing alternative perspectives, stimulating critical thinking and creativity [65,68].

Reference [69] demonstrates that when AI Chatbots possess characteristics like ease of use, usefulness, human-like interaction, reliability, and personalization capabilities, they significantly and positively influence teachers’ perceptions and willingness to adopt them for teaching and learning purposes.

Hypothesis 3: 

The characteristics of AI Chatbots positively influence teachers’ perception of using AI Chatbots.

AI Chatbots are increasingly present tools in learning and documentation processes [61,70]. They can cultivate critical thinking by delivering information and interacting with users [30,46]. Chatbots provide quick, context-aware access to complex information, reducing the time spent searching and allowing focus on interpretation [31,68]. Interacting with AI requires users to develop critical evaluation skills, such as verifying information with reliable sources and being aware of the risk of errors or biases [25,31]. AI Chatbots facilitate exploration of multiple perspectives, challenging users to approach problems from various angles [39,56,67].

Reference [71] demonstrates that the cognitive benefits (enhanced learning, knowledge acquisition, problem-solving) and pedagogical benefits (personalized learning, efficient teaching, continuous support) significantly and positively influence educators’ perceptions of AI chatbot adoption in higher education.

Hypothesis 4: 

The cognitive and pedagogical benefits of AI Chatbots positively influence teachers’ perception of using AI Chatbots.

2.4. Students’ Perspective on AI Chatbots

AI Chatbots are considered useful tools because they can streamline processes and increase productivity [56,65]. Students perceive AI Chatbots as valuable and practical tools, especially for generating new ideas, saving time, and improving learning efficiency [72].

According to recent studies, one year after the introduction of ChatGPT, most students are comfortable with adopting GenAI technology and developing habitual user behaviour [30,37]. Several factors influence students’ intention to adopt and use ChatGPT:

Students are more likely to use a chatbot if they perceive it as applicable in counselling. Thus, AI is believed to help students complete their learning tasks more quickly and more productively [56,64]. Students may be willing to adopt AI counselling Chatbots if they find them easy to use [72]. Studies show that performance expectations motivate students to use and accept new technologies [64].

The primary goal of personalization is to provide answers that understand users’ requirements and contexts. AI Chatbots can provide individualized recommendations to students, increasing collaboration and communication, and improving learning outcomes. Studies indicate a significant positive impact of personalization on students’ intention to use AI chatbots [34,72]. AI Chatbots have been shown to increase student engagement and learning outcomes by simulating human conversations [31,39,73].

Reference [74] demonstrates that, in the higher education context, emotional engagement (enjoyment) and trust are paramount for the adoption of AI chatbots. The research provides actionable insights for educational institutions and developers seeking to enhance AI integration in learning environments, emphasizing the need to prioritize student engagement and security to facilitate successful adoption.

Hypothesis 5: 

Students’ positive perceptions of AI Chatbots are positively associated with stronger perceived normative support for institutional AI integration.

AI Chatbots are for explanations but also generation of questions for self-assessment, being a valuable tool for students in the process of understanding concepts by providing detailed descriptions and logical structuring of information automatic generation of tests for the preparation of assessments [27,55], precise answers to complex questions from various fields of science [70,75], facilitating language learning [68] but the possibility of continuing the dialogue on a topic, which allows for the deepening of knowledge [68].

One of the most valuable functions of AI Chatbots is to adapt the learning process to each student’s needs, thereby generating personalized learning experiences [31,58]. Therefore, AI Chatbots analyze the students’ pace, style, and level of knowledge to provide a personalized educational path [76], improve engagement and support self-directed learning [37,41], and provide immediate, personalized, and scalable feedback [51]. Activities adapt according to the strengths or difficulties encountered by each student [52,76].

AI Chatbots function independently of time and space, ideal for asynchronous online learning and adult learners [38,70], act as an interactive assistant [27], with instructions in real time [44,51], and can simulate human conversations [38,73]; the possibility of receiving step-by-step assistance at any time creates a sense of permanent support [45], AI Chatbots can deliver dedicated help for students with distinctive needs, encouraging a more inexpensive education [30].

AI chatbots can serve as valuable student allies, complement traditional learning methods, and facilitate student preparation, thereby supporting learning and revision [52,65,77]. Currently, AI chatbots provide support for writing code, essays, poems, and other forms of written content [78]. Moreover, they support academic research by analyzing information [31] and assisting in structuring research papers [45,65]. Thus, AI Chatbots provide detailed and tailored feedback. Actively contributing to the development of competences in writing academic papers [45,79], the automatic identification of weaknesses and the recommendation of strategies to improve scientific papers [45], the automatic correction of written texts [79], but also the provision of a framework for reflection and self-regulation, through constant feedback throughout the academic writing process [45].

Ref. [80] emphasizes that how students engage with AI tools matters more than mere access or awareness—suggesting that assisted and personalized learning contexts (which provide structure and purpose for AI use) would be particularly effective in developing positive perceptions and meaningful engagement.

Hypothesis 6: 

Using AI Chatbots for assisted and personalized student learning positively influences students’ positive perception of using AI Chatbots.

2.5. Risks and Limitations of AI Chatbots

Although AI Chatbots are powerful and useful, they also pose risks that require special attention [65,81]. Excessive use can decrease critical thinking, creativity, and problem-solving ability [53,82]. It fosters passive learning, centered on the reproduction of information [53,68].

AI-generated responses can be superficial, incomplete, or contradictory [25,61], and are susceptible to the phenomenon of “AI hallucination” [61,81]. Opinions divide—some investigators claim that outsourcing the innovative process can alter individual creativity [81], while others emphasize the possibility of AI helping creativity [65].

AI Chatbots cannot reproduce the complexity of interpersonal relationships and a teacher’s empathy [37,83]. Technology should complement, not replace, the authentic interaction between teachers and students [37,83]. Human contact remains essential in socio-emotional development and the formation of critical thinking through dialogue and debate [31,83].

Hypothesis 7: 

Perceived cognitive risks and limitations of AI Chatbots influence the perceived normative support for institutional AI integration.

The use of AI Chatbots has raised concerns about academic integrity [37,41]. This reality necessitates reassessing educational practices and institutional policies [70,84]. Risks of plagiarism and academic fraud include automated generation of high-quality content [46]; passing professional exams [60,61]; circumvention of the learning process [41,45]; and a subtle form of plagiarism [41,83]. Amplification of “contract cheating” [64]: the phenomenon whereby students outsource the realization of work is amplified by the use of AI Chatbots, as it becomes increasingly difficult to differentiate between original and automatically generated content [64].

Hypothesis 8: 

Perceived academic integrity risks of AI Chatbots negatively influence perceived normative support for institutional AI integration.

The authors have reported significant concerns about the information provided by GenAI [31,65]. The phenomenon of artificial “hallucinations” and systematic inaccuracies includes the generation of false but plausible information [61,81], lack of self-knowledge and verification [53,84], algorithmic biases and systemic biases [35,52], and uniform presentation of correct and incorrect information [57,70]. Temporal and informational limitations: knowledge frozen in time—absence of real-time connectivity [53,84]; limitations in accessing recent academic literature [84].

Superficiality of answers and lack of analytical depth: vague and insubstantial answers [57,85]; the absence of deep semantic understanding [45,84]; issues of relevance and specificity [65,86]; quantitative limitations [57,68]. Among the critical elements are the difficulty of distinguishing between correct and incorrect [27,70,75], the transparency regarding training data and how AI generates content [43,45], attribution and referencing issues [66], and the quality of responses [63].

Hypothesis 9: 

Perceived limitations in accuracy and reliability significantly reduce perceived normative support for institutional AI integration.

2.6. Longitudinal Studies on the Long-Term Impact of Artificial Intelligence Chatbots on Students’ Educational Experiences

Longitudinal studies play a critical role in deeply understanding the long-term impact of AI Chatbots on students’ educational experiences. This research enables monitoring of the evolution of attitudes, behaviors, and learning outcomes over an extended period, providing a more complete and realistic picture of the problem [72,83].

Several arguments, highlighted in the literature, support the need for such studies. Thus, from a technological standpoint, AI chatbots are considered a relatively recent innovation, as the public launch of ChatGPT occurred in November 2022, and their many applications are still being explored [30,65]. Most current studies provide only current perspectives and do not capture adaptation processes or evolutionary changes in dynamics [72,87]. AI technologies have growing capabilities and applicability [73,81]. Longitudinal studies help to understand how technological transformations affect educational approaches [46,72]. Short-term results do not necessarily reflect a real impact [42,87]. For this reason, long-term studies can distinguish between temporary effects and sustainable benefits. An important issue is the transformation of students’ intentions to use AI Chatbots into actual use behaviors, and the monitoring of changes in usage patterns as users gain experience [64,72].

Long-term analysis of how AI Chatbots influence students’ engagement and motivation in learning leads to identifying both positive effects and potential risks of technological addiction [27,31]. As the use of AI Chatbots is increasing in the university system, this leads to the need to conduct longitudinal studies that can capture and monitor challenges related to academic integrity, directly contributing to the development of policies for the ethical and responsible use of AI [72,84].

The teacher-student relationship and teaching roles may change as AI is integrated into teaching (e.g., generating materials [50,53] or providing personalized feedback [37,45]). Thus, changes in the dynamics of educational relationships may occur [73], necessitating longitudinal investigations.

Longitudinal studies are fundamental to a rigorous and comprehensive understanding of how AI Chatbots influence the educational process. Only through a long-term perspective can well-founded educational policies and effective pedagogical practices be articulated and adapted to the challenges and opportunities posed by new technologies [2,25,46].

To test the research Hypotheses H1–H9, we build the model shown in Figure 1.

3. Materials and Methods

Based on the specialized literature presented in the references, a 75-item survey was used with a 5-point Likert scale. Between March 2025 and April 2025, we distributed the questionnaire online to students of the Faculty of Economics and Business Administration at the West University of Timisoara, and we received 408 valid responses out of 429 (see

Table 1. Respondents ‘data.

Characteristics	Category	Frequency	%
Gender	Female	228	55.88
	Male	180	44.12
Students	First-year	90	22.06
	Second-year	140	34.31
	Third year	178	43.63
Age	18	8	1.96
	19	85	20.83
	20	123	30.15
	21	144	35.29
	22	33	8.09
	23	7	1.72
	over 23	8	1.96
AI chatbot using	ChatGPT	408	100.00
	Gemini	74	18.14
	DeepSeek	46	11.27
	Microsoft Copilot	40	9.80
	Claude	20	4.90
	Grok	14	3.43
	Perplexity	8	1.96

). For reasons related to model reliability and validity, 29 items were eliminated, leaving 46 items in the final model. We constructed a model with ten constructs (Figure 2 and Appendix A): AII—perceived normative institutional AI Integration: although AII refers to institutional transformation, it is measured at the perceptual level through students’ evaluations of institutional responsibilities and governance expectations.; AIL—Artificial intelligence literacy; PR—Perception and intention to use AI Chatbots among higher education professors; FT—AI Chatbot features; BEN—Cognitive and pedagogical benefits of AI Chatbots in developing critical thinking and creativity; ST—Positive perceptions and intention to use AI Chatbots among students; AIU—AI Chatbots using for assisted and personalized student Learning; RL—Perceived cognitive risks and limitations of AI Chatbots; AIR—perceived academic integrity risks; and LAR—perceived AI chatbots’ limitations of accuracy and reliability.

When we started the study, we searched the Web of Science for papers on AI Integration in Education and selected well-rated journals. Even though many of these papers were more theoretical, the journals’ reputation helped us move towards reading with recognition within the academic environment. The authors are university teachers, which means we are directly involved in this process and understand very well what specialized literature presents, but we emphasize relevant and recognized literature in the academic environment. We consider the presented references the foundation of our study, and we selected many items that resonated with our general perception. We are convinced that the selection is not necessarily optimal, but we still consider it relevant. 29 items were eliminated. In almost all cases, items were eliminated due to reliability concerns (outer loadings were below 0.68). For example, the item “Chatbot AI had a positive impact on critical thinking skills, exposing different points of view and challenging them to analyze and evaluate” used in AIU items, had a 0.473 value for outer loading. We added the item “AI Chatbots can generate lesson plans, teaching materials, and updated course content” to FT. It had an outer loading value of 0.455 for FT and 0.696 for AIR. Here, we acknowledge an interpretive gap between what we intended to ask and what the students understood. Perhaps the term ‘teaching materials generation’ seemed more closely related to AIR than to FT. Thus, we dropped this item.

We analyzed the model using Partial Least Squares Structural Equation Modeling (PLS-SEM) and SmartPLS 4. We assessed the reliability and validity of the research model. PLS-SEM is a variance-based statistical technique used to analyze complex relationships between observed variables and latent constructs. This is a widely used statistical method in research. Theoretical support can be found in many sources, including [88]. In the next section, we present the numerical results obtained using PLS-SEM.

4. Results

In PLS-SEM, outer loadings indicate how well each item reflects the construct to which it belongs. Common rules [88]: good, reliable indicator (≥0.70); conditionally acceptable (0.40–0.69) is kept if the AVE and CR are good; is eliminated (<0.40). In

Table 2. Outer loadings, Cronbach’s alpha, Composite reliability rho_a, Composite reliability CR, and Average variance extracted (AVE).

Construct	Item	Outer Loading	Cronbach’s Alpha	rho_a	CR	AVE
AII	AII1	0.802	0.849	0.850	0.888	0.570
	AII2	0.743
	AII3	0.769
	AII4	0.700
	AII5	0.754
	AII6	0.759
AIL	AIL1	0.775	0.819	0.819	0.873	0.580
	AIL2	0.801
	AIL3	0.800
	AIL4	0.687
	AIL5	0.738
	AIR1	0.775	0.830	0.835	0.887	0.662
	AIR2	0.821
	AIR3	0.841
	AIR4	0.816
	AIU1	0.884	0.835	0.862	0.883	0.604
	AIU2	0.722
	AIU3	0.704
	AIU4	0.825
	AIU5	0.735
	BEN1	0.810	0.743	0.746	0.854	0.662
	BEN2	0.866
	BEN3	0.762
	FT1	0.769	0.848	0.852	0.887	0.567
	FT2	0.732
	FT3	0.783
	FT4	0.780
	FT5	0.726
	FT6	0.726
	LAR1	0.782	0.765	0.769	0.850	0.587
	LAR2	0.733
	LAR3	0.736
	LAR4	0.810
	PR1	0.702	0.713	0.716	0.822	0.536
	PR2	0.792
	PR3	0.713
	PR4	0.719
	RL1	0.772	0.875	0.876	0.905	0.615
	RL2	0.809
	RL3	0.794
	RL4	0.796
	RL5	0.785
	RL6	0.747
	ST1	0.896	0.786	0.800	0.877	0.705
	ST2	0.890
	ST3	0.722

, all loadings are ≥0.687, indicating no major problematic indicators. According to the PLS-SEM literature [88], for internal reliability, it is necessary that: Cronbach’s alpha ≥ 0.70, rho_a ≥ 0.70; composite Reliability (CR) between 0.70 and 0.95 (good internal consistency, no redundancy). The results in Table 2 indicate that all constructs exhibit adequate internal reliability, with Cronbach’s alpha and composite reliability (rho_a and CR) exceeding the recommended threshold of 0.70. Also, the AVE values exceed the 0.50 threshold for all constructs, confirming the measurement model’s convergent validity.

In PLS-SEM, for discriminant validity, we study Fornell–Larcker criterion analysis and the heterotrait–monotrait ratio (HTMT) matrix. The Fornell–Larcker criterion states that, to ensure discriminant validity, the square root of the AVE (diagonal values) must exceed all correlations of that construct with other constructs (the values in the row and column). For the heterotrait–monotrait ratio (HTMT) matrix, the values less than 0.85 mean good discriminating validity. The values in

Table 3. Fornell–Larcker standard study.

	AII	AlL	AlR	AlU	BEN	FT	LAR	PR	RL	ST
AII	0.755
AlL	0.703	0.761
AlR	0.465	0.351	0.813
AlU	0.305	0.269	0.269	0.777
BEN	0.283	0.298	0.168	0.433	0.814
FT	0.431	0.389	0.205	0.543	0.517	0.753
LAR	0.362	0.320	0.546	0.148	0.158	0.234	0.766
PR	0.642	0.529	0.364	0.371	0.403	0.384	0.278	0.732
RL	0.572	0.446	0.633	0.240	0.124	0.270	0.495	0.384	0.784
ST	0.674	0.600	0.402	0.472	0.253	0.513	0.235	0.562	0.567	0.840

and

Table 4. Heterotrait–monotrait ratio matrix.

	AII	AlL	AlR	AlU	BEN	FT	LAR	PR	RL	ST
AII
AlL	0.822
AlR	0.548	0.419
AlU	0.353	0.312	0.314
BEN	0.356	0.380	0.213	0.547
FT	0.495	0.462	0.242	0.626	0.650
LAR	0.447	0.398	0.686	0.179	0.232	0.283
PR	0.812	0.668	0.464	0.463	0.558	0.470	0.377
RL	0.661	0.517	0.745	0.265	0.203	0.308	0.608	0.473
ST	0.821	0.730	0.492	0.567	0.329	0.620	0.306	0.728	0.677

confirm the discriminant validity, as shown in Figure 1 and Figure 2.

Structural model assessment measures include the coefficient of determination (R2), the variance inflation factor (VIF), the statistical significance and relevance of path coefficients, and the predictive relevance (Q²).

Usual benchmarks in PLS-SEM [89] for R²: ≥0.75—substantial; ≥0.50—moderate; ≥0.25—weak. The structural model (Figure 2) shows strong explanatory power for AII (R² = 0.67) and modest explanatory power for PR (R² = 0.205) and ST (R² = 0.223), indicating that the model explains AII well, while professors’ and students’ intentions/perceptions are only partially captured and likely depend on additional determinants not included in the model.

The rule for path coefficients is: >0.30—strong effect; >0.20—moderate effect; >0.10—weak effect. Results indicate that AIL has a moderate to strong positive effect on AII (β = 0.348), while PR (β = 0.256) and ST (β = 0.205) also exhibit meaningful contributions (

Table 5. Path coefficient effects. (D—direct, I—indirect, T—total).

	D	I	T	D	T	D	T
	AII	AII	AII	PR	PR	ST	ST
AII	0.348		0.348
AIR	0.056		0.056
AlU		0.097	0.097			0.472	0.472
BEN		0.071	0.071	0.278	0.278
FT		0.062	0.062	0.240	0.240
LAR	0.024		0.024
PR	0.256		0.256
RL	0.154		0.154
ST	0.205		0.205

). Higher AII is associated with higher AIL. Substantively, this supports the logic that in institutions that integrate AI more strongly (policies, redesign, support), the users are more literate/competent in AI use. When professors are more positive about AI Chatbots and more willing to use them, universities tend to show higher levels of AII. Students’ trust/satisfaction/stance toward AI chatbots improves AII—but the effect isn’t large, other determinants likely matter.

The path coefficient from RL to AII (β = 0.154) indicates a small but meaningful positive effect, suggesting that concerns regarding learning-related risks—such as reduced critical thinking, creativity, and over-reliance on AI chatbots—moderately increase support for structured and responsible AI integration in higher education. While these risks do not strongly drive integration strategies, they underscore the need for pedagogical guidance and institutional frameworks. AIR shows a negligible positive effect on AII in universities (β = 0.056), indicating that concerns about cheating and plagiarism are not a substantial direct driver of AII. Ethical risks are recognized, but they are not, in themselves, sufficient to trigger major structural changes. The path coefficient from LAR to AII is extremely weak (β = 0.024), indicating a negligible influence of perceived limitations and accuracy risks of AI chatbots on AII.

AIU strongly predicts ST (β = 0.472), suggesting that AI usage plays a key role in shaping ST. When students experience or perceive themselves using AI Chatbots in these practical, supportive ways, they tend to develop more positive attitudes and stronger intentions to use them.

The results indicate that for the indirect relationships AIU → AII, BEN → AII, and FT → AII, the magnitudes of the path coefficients are small (β = 0.097, β = 0.071, and β = 0.062, respectively), suggesting limited explanatory power. These findings imply that AIU, BEN, and FT contribute only marginally to explaining the variance in AII. Rather than acting as key determinants, these constructs appear to play complementary roles within the overall explanatory model, with weak effects.

To explore the R² statistical implications, we used a bootstrapping technique in SmartPLS with 5000 iterations. Hypotheses 1–7 are supported (these hypotheses are true)—the values in

Table 6. Path coefficients. Direct effects.

	Path Coefficients	0.025	0.975	T Statistics	p Values	Remark
AlL -> AII	0.348	0.269	0.430	8.495	0.000	H1 is supported
PR -> AII	0.256	0.162	0.356	5.191	0.000	H2 is supported
FT -> PR	0.240	0.156	0.338	5.095	0.000	H3 is supported
BEN -> PR	0.278	0.172	0.378	5.289	0.000	H4 is supported
ST -> AII	0.205	0.112	0.290	4.487	0.000	H5 is supported
AlU -> ST	0.472	0.375	0.572	9.493	0.000	H6 is supported
RL -> AII	0.154	0.068	0.241	3.511	0.000	H7 is supported
AlR -> AII	0.056	−0.024	0.139	1.334	0.182	H8 is not supported
LAR -> AII	0.024	−0.065	0.115	0.524	0.601	H9 is not supported

are statistically significant; Hypotheses 8 and 9 are not supported (these hypotheses are false). In Table 6, we present the path coefficients (values greater than 0.100 are significant), confidence intervals, t-statistics, p-values, and remarks on each hypothesis.

In addition to establishing discriminant validity using the HTMT criterion, we assessed common method variance using the full collinearity VIF approach [89]. All construct-level VIF values were below the conservative threshold of 3.3 (

Table 7. The full collinearity of VIFs at the construct level.

Construct	Full Collinearity VIF
AII	2.636
AIL	2.649
PR	2.723
ST	2.908
BEN	2.869
FT	2.873
AIU	2.967
RL	2.911
AIR	2.931
LAR	3.007

), suggesting that collinearity is not problematic and that common method bias is unlikely to threaten the validity of the results. Taken together, these diagnostics provide further support for the empirical distinctiveness of the constructs, beyond potential measurement-method effects.

In

Table 8. f² values.

	AII	PR	ST
AlL	0.206
AlR	0.005
AlU			0.287
BEN		0.071
FT		0.053
LAR	0.001
PR	0.122
RL	0.032
ST	0.059

, we present the effect size f² (≥0.02 is small; ≥0.15 is medium; ≥0.35 is large), which indicates the proportion of variance in one factor accounted for by another factor. f² tells how much an individual path contributes to the R² of the dependent variable. The real effects are: AIL → AII, with f² = 0.206—medium, solid effect and AIU → ST, with f² = 0.287—medium, solid effect. Effects are small for relationships: PR → AII: 0.122; BEN → PR: 0.071; FT → PR: 0.053; RL → AII: 0.032; ST → AII: 0.059. Negligible/problematic effects are AIR → AII: 0.005; LAR → AII: 0.001—these are dead paths, which cause a high SRMR. The impact on AII is almost non-existent; removing them from the model would probably not change R² significantly. If we were to refer to a model like “what are the factors that significantly influence AII”, these pathways would have to be eliminated to improve the SRMR, thereby outlining a better model. As the literature indicates, there are many concerns regarding AIR and LAR. In our study, we retained these constructions precisely to examine how they influence AII, and the finding that they do not is part of our results. SRMR (the main indicator of model fit) is 0.076, which falls below the accepted threshold in PLS-SEM: SRMR < 0.08. Following Henseler’s bootstrapped model fit assessment, the discrepancy measure d_ULS = 6.191 and d_G = 2.089 of the estimated model exceeded the 95% quantile of the bootstrap distribution. NFI = 0.627—in PLS-SEM, values between 0.60 and 0.80 are frequently encountered in complex models. However, in variance-based SEM, global fit measures are considered supplementary diagnostics. In our case, the SRMR value (0.075) remains below the recommended cutoff of 0.08, suggesting an acceptable overall model fit. Combined with the satisfactory reliability, convergent and discriminant validity of the constructs, the absence of problematic construct-level collinearity, and the substantial explained variance in AII (R² ≈ 0.67), the model can be regarded as empirically adequate for explaining the hypothesized relationships.

Guidelines [84] for Q² are: high prediction ≥0.5; ≥0.2 moderate prediction; ≥0.10—weak prediction. The PLS predictive assessment indicates strong predictive relevance for AII (Q² = 0.535), demonstrating robust out-of-sample predictive performance for institutional AI integration. In contrast, PR (Q² = 0.192) and ST (Q² = 0.212) show modest but positive predictive relevance, suggesting acceptable predictive performance while indicating the presence of additional unexplored predictors.

Descriptive statistics are in Appendix B.

5. Discussion

This paper examines the factors influencing the normative support for the integration of AI Chatbots in universities. AI literacy is the most critical factor influencing AII (with a path coefficient of 0.385; see Figure 2 and Table 5) and exerts a direct effect. AI literacy refers to the fact that to integrate AI chatbots into education effectively, it is essential to organize training sessions for students and teachers. Reference [47] confirms that AI Literacy is a foundational enabler of AI adoption by influencing performance expectancy, effort expectancy, and social influence. Reference [48] shows that AI Knowledge is the strongest predictor of actual AI adoption. AIL → AII, with f² = 0.206, means that removing AIL from the model would produce a substantial decrease in R² for AII. AIL has one of the model’s real structural effects: AI literacy is a catalyst for institutional reform.

The following factors influence direct AII:

PR (professors’ perceptions on using AI Chatbots in universities)—with the path coefficient of 0.227, see Figure 2 and Table 5. University professors see AI Chatbots as a promising tool for streamlining and enriching the educational process, provided their use complements classical pedagogical approaches. Our study confirms that this positively impacts AII [59] shows that teachers’ positive perceptions influence the integration of AI chatbots in universities. Key positive perception factors are hedonic motivation (the strongest predictor), positive emotions, and performance expectancy. PR exerts a small-to-moderate effect on AII (f² = 0.122), suggesting that positive academic attitudes toward AI Chatbots are associated with tendencies toward institutional reform. However, the effect is weaker than that of AI literacy, suggesting that favorable perceptions alone are insufficient to drive structural transformation without deeper conceptual and pedagogical grounding. The effect of BEN on professors’ perceptions (PR) is small but meaningful (f² = 0.071). This indicates that while the perceived potential of AI Chatbots to stimulate critical thinking and broaden perspectives contributes to positive attitudes, it does not constitute the primary explanatory factor. Professors’ perceptions appear to be shaped by a broader pedagogical and strategic framework rather than by cognitive enhancement alone. The effect of FT on PR is small (f² = 0.053). This suggests that while awareness of AI Chatbots’ capabilities contributes to positive attitudes, technological functionality alone does not strongly shape professors’ perceptions. AI acceptance is not purely technological. Teachers’ perceptions are mediated by pedagogical considerations. Functionality is necessary, but not sufficient, for educational legitimacy. Attitudes appear to be influenced more by pedagogical value than by technical performance.

ST (positive perceptions and intention to use AI Chatbots among students)—with a path coefficient of 0.215, see Figure 2 and Table 5. Students are already familiar with AI Chatbots and find them especially useful for generating new ideas, saving time, and increasing learning efficiency. The effect of ST on AII is small (f² = 0.059), indicating that while students’ familiarity, perceived usefulness, and intention to use AI Chatbots influence AII, they are not the primary drivers of AII. This shows that AII Integration is not primarily determined by student openness, but rather by institutional and pedagogical factors (AIL and PR in our model). Reference [70] shows that students’ positive perceptions of AI chatbots influence their integration into universities, and that perceived enjoyment is the strongest predictor. One of the main advantages is the ability to offer a personalized learning experience. Reference [90] demonstrates, through PLS-SEM analysis, that the use of AI chatbots as assisted and personalized learning tools (operationalized through self-directed learning with technology and perceived usefulness) positively influences students’ perception of the use of chatbots, reflected by intention to use and actual use. Figure 2 shows that AIU exerts a substantial effect on ST (f² = 0.287), indicating that practical, pedagogically meaningful engagement with AI Chatbots significantly enhances students’ familiarity, perceived usefulness, and intention to adopt these tools. This finding suggests that experiential interaction, rather than abstract attitudes alone, plays a central role in shaping student endorsement of AI technologies in higher education.

The informal use of AI Chatbots among students and teachers is already massive [30,72]. This use brings important benefits [63,69] but also challenges that require a strategic approach [41,64]. For effective integration, higher education institutions need to develop clear policies [41,46], provide training programs for students and teachers [53,64], redefine assessments [60,81], and promote a culture of ethical and responsible use [58,64].

In addition to these positive factors, the study also reveals a negative factor that can positively influence AIII: the cognitive limitations generated by AI Chatbots (addiction and avoidance of intellectual effort), even if they are real problems, can act as a driver of responsible integration [67,68]. Consequently, universities need to actively manage these risks by developing AI skills. The effect of RL on AII is small (f² = 0.032). This suggests that RL is not the primary driver of institutional transformation. AII appears to be shaped more by pedagogical and strategic considerations than by RL.

Our paper shows that academic integrity risks [37,64] and limitations of accuracy and reliability [81,82] do not influence AII. However, universities must balance innovation with adherence to educational and ethical standards through clear policies, the redesign of assessments, and the promotion of AI literacy. The effect of AIR on AII is negligible (f² = 0.005). This finding indicates that concerns related to plagiarism, cheating, and source attribution do not substantially explain institutional AI integration strategies. AII is not triggered by the fear of fraud, but by the need for pedagogical and institutional adaptation. The effect of LAR on AII is negligible (f² = 0.001). This finding indicates that concerns related to hallucinations, outdated information, and response vagueness do not meaningfully explain AII. AII is not determined by technological imperfections, but by pedagogical and institutional maturity. AIR and LAR were not significant predictors in this sample. Several contextual and methodological explanations may account for these non-significant results. First, the relative homogeneity of the sample (economics students from the same institution) may have limited variance in perceptions. Second, measurement limitations or ceiling effects may have reduced the ability to detect significant relationships. Finally, it is possible that perceived benefits of AI integration outweigh perceived risks in this educational context.

This study advances the literature on AI adoption in higher education by reconceptualizing institutional AI integration as a legitimacy formation process rather than a purely technological adoption outcome. By positioning perceived normative support as a central institutional mechanism, the model extends traditional acceptance frameworks (e.g., TAM—Technology Acceptance Model [91] and UTAUT—Unified Theory of Acceptance and Use of Technology [92]) toward an institutional and climate-based perspective grounded in legitimacy and organizational theory. Furthermore, the study integrates cognitive drivers (e.g., literacy and perceived benefits) and ethical–technical inhibitors (e.g., academic integrity risks and reliability limitations) within a unified explanatory framework, offering a more balanced and realistic account of AI integration dynamics. In doing so, it contributes to the emerging discourse on responsible AI governance by demonstrating that institutional transformation depends on socially constructed normative support rather than on individual attitudes or technological capabilities alone.

The model’s context is institutional AI integration in higher education. The model integrates multiple theoretical layers: micro level—TAM, UTAUT, IS Success Model (Information Systems Success [93]); meso level—Organizational climate theory [94], Institutional theory [95]; and macro level—Institutional legitimacy [96] and Responsible AI governance [97].

When AI integration is viewed as socially supported within the university context—by leadership, peers, and academic communities—it gains legitimacy, reducing uncertainty and resistance while fostering collective alignment. In this sense, perceived normative support functions as a proximal institutional mechanism that translates individual evaluations into organizational-level integration processes. Therefore, higher levels of perceived normative support are expected to facilitate institutional AI integration in higher education settings, as shown in other works [98,99,100].

This model is a hybrid framework integrating Technology Adoption Models (TAM/UTAUT) with Institutional Theory to explain both individual AI use and perceived institutional transformation in higher education: TAM/UTAUT (BEN → perceived usefulness; ST → behavioral intention; AIU → actual use; FT → performance expectancy); Risk theory (RL; AIR; LAR); and Institutional Theory (AII—normative pressure; AIL -institutional capacity building; PR—normative influence of professors). This model extends technology adoption theories by incorporating institutional-level normative expectations, thus bridging individual adoption mechanisms with perceived organizational transformation processes (see

Table 9. Conceptual Positioning of the Model.

Dimension	Hypotheses	Description	Theoretical Grounding	Observation
Cognitive and Perceptual Drivers
Competence Dimension	H1—GenAI literacy	cognitive capability, critical AI understanding, and informed evaluation capacity.	Knowledge-based view [101]; IS capability frameworks [93]; Digital competence theory [102].	This hypothesis suggests that the legitimacy of AI integration is partially rooted in informed literacy rather than blind enthusiasm.
Teachers’ Perceptual Dimension	H2—Teachers’ favorable perceptions H3—AI Chatbot characteristics H4—Cognitive and pedagogical benefits	These hypotheses establish a structural chain: AI Chatbot characteristics → Teacher perception → Normative support	TAM (Perceived Usefulness) [91] UTAUT (Performance Expectancy) [92] IS Success Model [93]	Unlike traditional adoption studies, the outcome is not behavioral intention, but institutional legitimacy perception.
Students’ Perceptual Dimension	H5—Students’ positive perceptions H6—AI-supported personalized learning	These hypotheses establish a structural chain: AI-assisted learning → Student perception → Normative support	Student-centered learning theory [103], Educational technology acceptance models [92], Diffusion of innovation logic [104]	By incorporating both teachers and students, the model captures institutional legitimacy as a socially distributed construct.
Ethical and Cognitive Inhibitors—the model includes both drivers and barriers.
	H7—Cognitive risks	deskilling, overreliance, superficial learning	Risk–benefit adoption frameworks [105] Technology resistance models [106], Technology trust theory [107], Institutional legitimacy theory [96]
	H8—Academic integrity risks	plagiarism, ghostwriting, assessment validity erosion
	H9—Accuracy and reliability limitations	Hallucinations, factual inaccuracies, instability
Central Institutional Mechanism: Perceived Normative Support
	The pivotal construct of the model is: Perceived Normative Support for Institutional AI Integration	It functions as a legitimacy mechanism, a perceived normative climate, and a social filter for technological adoption.	Theory of Planned Behavior (subjective norm component) [108], Institutional theory (normative pressures) [95], Organizational climate theory [108], Legitimacy theory [109].	By positioning normative support as central, the model extends beyond TAM/UTAUT and moves into the domain of Institutional AI governance and organizational transformation.

).

6. Conclusions

This paper indicates that the success of AII depends mainly on the academic community’s AI literacy and teachers’ and students’ positive perceptions of AI. The finding that cognitive risks positively affect AII implies that universities must develop strategies to mitigate them.

The study presented in this paper shifts focus from individual use to perceived institutional integration; integrates dimensions of governance and responsibility; introduces a normative-institutional perspective on AI in higher education; and extends classic adoption models towards organizational transformation.

6.1. Theoretical and Practical Implications for Higher Education Institutions

Our paper shows the importance of prioritizing artificial intelligence (AI) literacy through developing training programs for teachers and students, investing in educational resources for AI skills, and creating centers of excellence in educational AI. To address the need to develop institutional policies, it is necessary to clearly define the ethical use of AI-based chatbots, establish guidelines for academic integrity, and develop assessment criteria adapted to the AI era. Thus, educational processes will be redesigned based on adapting teaching methods to human-AI collaboration, reforming assessment systems, and promoting critical thinking and creativity.

The results of this paper highlight practical implications for teachers: AI literacy and adapting pedagogical methods. In addition, the redefinition of the teaching role will be done by moving from transmitting information to mentoring, focusing on developing critical thinking, and promoting interactive and personalized learning. Regarding practical implications for students, the emphasis is on responsible use, developing information verification skills, understanding the limits of AI chatbots, and maintaining a balance between AI assistance and personal effort.

6.2. Limitations of the Paper and Future Research Directions

Our study relies on a sample consisting exclusively of 408 economics students from a single university. While the sample size is adequate, its homogeneity introduces notable limitations regarding the representativeness of the findings. Students enrolled in business and economics programs could be more digitally literate and more receptive to technological innovations, which may positively bias attitudes toward AI-based tools such as chatbots. Moreover, the absence of participants from disciplines in which ethical, pedagogical, or epistemological concerns about AI are more pronounced (e.g., the humanities, law, social sciences, teacher education) could constrain the scope of the conclusions. As a result, the findings may differ for the university student population, even though our study results suggest otherwise. The authors acknowledge this limitation. Future research should use more diverse, heterogeneous samples across institutions and academic disciplines to enhance the external validity of findings. Comparative and mixed-method approaches would provide deeper insights into disciplinary differences and contextual factors shaping students’ attitudes toward AI Chatbots.

AI technologies, particularly generative AI tools, are evolving at an unprecedented pace, accompanied by fluctuating public discourse and institutional responses. As a result, student perceptions of AI use in higher education remain dynamic rather than fully stabilized. While this snapshot provides insight into prevailing trends at an early, formative stage of AI adoption, it may not fully reflect enduring or long-term attitudes toward AI integration. Perceptions of usefulness, risk, and ethical implications are likely to shift as students gain more experience with AI tools and as institutional guidelines mature. Future research would benefit from longitudinal or repeated-measures designs to examine how students’ perceptions and usage patterns develop over time. Such approaches would allow for the identification of stable trends, transitional phases, and potential shifts in attitudes as AI technologies become more embedded in educational practice.

Higher education institutions must invest in robust AI literacy programs, clear ethical-use policies, and examination systems adapted to the AI era to ensure the adequate and ethical integration of AI. The future of education lies in a hybrid system that integrates AI with fundamental human and academic values.

In conclusion, the ethical integration of AI-based chatbots in higher education contributes to sustainable digital transformation by enhancing educational quality. AI literacy and personalized learning promote equitable access and engagement, while governance frameworks and ethical policies ensure responsible adoption of AI. By integrating AI into sustainable institutional practices, universities can align technological innovations with long-term educational sustainability and SDG 4 (Quality Education). Thus, AI becomes a strategic tool that strengthens both learning outcomes and the broader role of universities as long-term sustainable ecosystems.