Exploring the Mechanisms Influencing Graduate Students’ Adoption of Generative AI: Insights from the Technology Acceptance Model

Abstract

The rapid development of Generative Artificial Intelligence (GenAI) in graduate education has changed human–AI interaction within knowledge-intensive environments, leading to important questions about user-side cognitive adaptation in probabilistic AI systems. While many studies focus on ethical implications, limited attention has been paid to the cognitive mechanisms underlying graduate students’ adoption of GenAI. Drawing on the Technology Acceptance Model (TAM), this study explores the cognitive and interactional mechanisms shaping graduate students’ adoption and usage of GenAI. Using thematic analysis of in-depth interviews with 20 graduate students from diverse academic backgrounds, the study identifies seven interrelated constructs: perceived usefulness, perceived ease of use, external environment, risk perception, attitude, behavioral intention, and interaction subjectivity. This study demonstrates that the adoption of GenAI is not merely a result of perceived efficiency but is shaped by cognitive calibration between trust and risk evaluation. Moreover, interaction subjectivity emerges as a metacognitive factor that determines whether engagement results in human–AI collaboration or passive automation. By integrating external environment, risk perception, and interaction subjectivity, this study provides a cognitively grounded framework for understanding human–AI adoption and interaction dynamics. Practically, the findings provide design-relevant insights for developing GenAI systems that support calibrated trust, uncertainty awareness, and adaptive cognitive participation.

1. Introduction

Unlike earlier deterministic systems, GenAI operates as a probabilistic cognitive computational system, which generates outputs based on large-scale pattern learning rather than rule-based logic. Such systems introduce epistemic uncertainty into human–machine interaction, as outputs are plausible and not inherently verifiable. Therefore, understanding the cognitive mechanisms of users’ engagement becomes increasingly important.

From a cognitive computing perspective, adoption of GenAI is not merely a behavioral choice but a process of cognitive calibration under uncertainty. This calibration process reflects that interaction outcomes depend not only on algorithmic capability but also on human regulation and metacognitive engagement. While GenAI has influenced knowledge-intensive environments, higher education institutions worldwide have been forced to grapple with both the opportunities and the risks of integrating GenAI into teaching and research [1]. For graduate education in particular, the stakes are high. Graduate students, positioned at the forefront of knowledge production and innovation, face a central dilemma: whether GenAI adoption supports the long-term development of independent research capacities or merely delivers short-term productivity gains at the expense of scholarly agency.

As a supportive instrument, GenAI can enhance academic research and interdisciplinary innovation by providing access to knowledge, generating new ideas, assisting in academic writing, and even supporting data analysis. These affordances are particularly appealing to graduate students, who often face intense productivity pressures and heavy cognitive loads [2]. For students writing in a second language, GenAI also provides linguistic support that can help overcome barriers to publication and scholarly communication [3]. However, concerns remain regarding its opacity, the reliability of outputs, and its potential to undermine critical thinking, academic integrity, and independent learning [4]. Previous research has identified issues such as hallucinated references, plagiarism, and over-reliance on machine-generated text, all of which raise questions about how students can engage responsibly with these tools [5].

Despite the growing body of literature, much existing scholarship has focused on the technological potential, ethical principles, and legal implications of GenAI. While valuable, these perspectives often neglect cognitive mechanisms through which users regulate interaction with probabilistic AI systems. Graduate students’ adoption of GenAI is particularly important to examine because their attitudes, intentions, and practices directly shape their academic performance, research capabilities, and future professional roles [6]. Moreover, graduate education is a context marked by unique tensions: the pressure to produce original research while maintaining academic integrity, the challenge of navigating disciplinary norms, and the opportunity to engage with cutting-edge technologies. Understanding how graduate students negotiate these tensions in the face of GenAI adoption is therefore both timely and necessary. To our knowledge, graduate students’ adoption of GenAI has implications that extend beyond immediate research efficiency to the long-term development of academic practices and research training.

The outputs of GenAI are interpreted and operationalized through human cognition. However, few studies have qualitatively examined the cognitive mechanisms underlying graduate students’ adoption of GenAI. Examining how graduate students cognitively negotiate in GenAI interaction contributes to a deeper understanding of human–AI co-adaptation processes, which are central to cognitive computing research.

To address this gap, this study draws on the Technology Acceptance Model (TAM) as a theory-informed analytical framework to investigate the cognitive mechanisms influencing graduate students’ adoption of GenAI. Through thematic analysis of in-depth interviews with 20 graduate students, this study explores not only why GenAI is adopted, but how distinct cognitive configurations shape patterns of human–AI interaction. By integrating perceived usefulness, perceived ease of use, external environment, risk perception, and interaction subjectivity into TAM, this research extends the model toward a cognitively grounded interpretation of AI adoption dynamics and offers insights relevant to the governance and design of AI-mediated research environments.

This study makes two primary contributions. First, from a theoretical perspective, it extends the TAM by incorporating external environment, risk perception, and interaction subjectivity into the analysis of graduate students’ adoption of GenAI. These additional factors provide a more comprehensive understanding of how attitudes and behavioral intentions are shaped when users interact with probabilistic AI systems. Specifically, this study contributes to cognitive computing research by illustrating how users calibrate risk evaluation, interaction strategies and other factors in human–AI interaction contexts. Second, from a practical perspective, before the interaction process can be translated into measurable indicators, their conceptual boundaries and interaction dynamics must be clarified. By identifying graduate students’ cognitive biases, attitudes, and challenges, this study provides empirically grounded insights for future design and evaluation of human-centered cognitive systems in academic environments.

To guide the analysis, the study addresses the following research questions:

RQ1: What factors influence graduate students’ acceptance and use of Generative AI?

RQ2: How do these factors interact to shape the mechanisms of adoption?

2. Literature Review

2.1. Generative AI in Higher Education: Opportunities and Concerns

Since late 2022, GenAI systems such as ChatGPT have become widely accessible, offering powerful language generation, summarization, and coding assistance that can support literature review, idea generation, and academic writing. Kasneci et al. [1] highlight that such affordances have created new possibilities for teaching and learning, particularly in higher education. For graduate students, who are expected to master complex disciplinary knowledge and produce original research outputs, these affordances are particularly attractive given the high cognitive load and productivity pressures of graduate training. Studies have shown that GenAI can enhance efficiency in synthesizing literature, generating new perspectives, and assisting in second-language academic writing [7].

At the same time, concerns about reliability and ethics have grown. Dwivedi et al. [2] note that the opacity of model mechanisms complicates the evaluation of outputs, while hallucinated references and fabricated facts remain frequent. Empirical studies also warn that reliance on AI-generated content can raise issues of plagiarism and authorship disputes, which may in turn weaken students’ ability to develop independent reasoning and critical thinking [8].

Adoption is further shaped by disciplinary contexts. In STEM fields, GenAI is often framed as a productivity enhancer, valued for coding support, data analysis, and rapid problem-solving. In contrast, the humanities and social sciences show more caution. For example, Lund et al. [9] found that many faculty members in social sciences worried about originality and interpretive nuance, echoing student concerns that over-reliance on GenAI might erode the individual voice in academic writing. These contrasts indicate that GenAI is not a neutral tool but one mediated by the epistemic cultures of different fields. International organizations have underscored that GenAI should be integrated in education in ways that support, rather than replace, human creativity, stressing the importance of a human-centered and responsible approach [10].

2.2. Technology Acceptance Model and Its Extensions in Educational Contexts

The TAM posits that perceived usefulness (PU) and perceived ease of use (PEOU) shape attitudes and behavioral intention toward technology adoption [11]. Its influence across information systems and education research is substantial, as it provides a simple yet predictive account of technology adoption behaviors. Later iterations, such as TAM2, TAM3, and the Unified Theory of Acceptance and Use of Technology (UTAUT), add variables including social influence, facilitating conditions, and experience-based factors [12,13,14]. More recently, Zacharis and Nikolopoulou [15] confirmed the relevance of UTAUT2 in e-learning contexts, demonstrating that learning value plays an essential role in predicting intention.

TAM3 emphasizes determinants of perceived ease of use, such as computer self-efficacy and perceived enjoyment [13], while UTAUT2 extends the framework with constructs like hedonic motivation, price value, and habit [16]. In higher education, TAM has been widely applied to investigate the adoption of digital technologies. Teo [17], for instance, demonstrated that TAM could predict teachers’ technology adoption in classrooms, but also argued that cultural contexts required adaptation of the model. Such findings suggest that while TAM offers predictive power, it may need contextual modifications to account for institutional and cultural dimensions.

More recently, with the emergence of artificial intelligence tools in education, scholars have begun testing TAM-related frameworks in this new domain. Zhang et al. [6] found that acceptance of AI among pre-service teachers varied significantly across groups, indicating that factors beyond perceived usefulness and ease of use, such as ethical concerns and trust, also matter. These insights point to the need for an extended TAM framework when studying graduate students’ adoption of GenAI, where transparency, accountability, and quality of human–AI interaction become critical considerations.

2.3. External Environment and Social Influence on Technology Adoption

TAM and its extensions recognize that adoption is not only an individual decision but also socially embedded. For GenAI, the external environment influences adoption through two primary channels. First, media exposure plays a powerful role. High-profile coverage of ChatGPT has shaped both enthusiasm and anxiety, framing it alternately as a “revolutionary assistant” and as a “threat to education”. Such narratives strongly affect students’ initial expectations and willingness to try the technology [4]. Second, interpersonal networks, including supervisors, colleagues, and academic communities, function as key vehicles of influence. Supervisors’ attitudes often legitimize or constrain student usage [17], while peer networks serve as informal training grounds where prompting skills and critical experiences are shared [18].

These dynamics resonate with UTAUT’s “social influence” construct [12], but they are particularly salient in graduate education, where hierarchical relationships with supervisors and collaborative peer cultures structure daily academic practices. Prior studies show that positive endorsements from trusted mentors can significantly accelerate adoption, while negative or cautious views may restrain it [19]. Thus, adoption decisions are embedded in a broader ecology of media narratives and academic community norms [6]. Together, these findings underscore that GenAI adoption cannot be understood solely as an individual cognitive process but must also be analyzed within its social and institutional context.

2.4. Risk Perception as a Determinant of Technology Adoption

While PU often dominates the adoption of productivity tools, GenAI’s distinctive risks require explicit consideration. Risk perception encompasses epistemic risks (hallucinations, unverifiable outputs), ethical and legal risks (plagiarism, authorship disputes, data privacy), and developmental risks (cognitive dependency) [2]. Cotton et al. [5] further warn that students’ reliance on machine-generated text can erode independent reasoning and weaken critical thinking, amplifying developmental risks.

Cross-cultural research suggests that perceptions of risk are shaped by policy and cultural environments. In Western contexts, debates around GenAI often center on ethics and transparency, with concerns about fairness, accountability, and bias [4]. Universities in Europe and North America have responded by issuing guidelines focused on academic integrity and disclosure [10]. In China, by contrast, policy discourse emphasizes regulation and governance, as reflected in the 2023 Interim Measures for the Administration of Generative AI Services [20]. Chinese universities often highlight compliance and state guidelines as central to acceptable use. These differences suggest that while students everywhere confront epistemic risks, their framing of ethical and legal risks is mediated by broader cultural and institutional contexts.

From a behavioral standpoint, risk perception does not always prevent use. Studies suggest an asymmetrical effect: when usefulness is perceived as high, especially under time pressure, students tend to downplay risks [6]. This is particularly true for graduate students, whose workload incentivizes efficiency. Nevertheless, neglecting risks can undermine academic integrity and professional development in the long term. Explicitly modeling risk perception within TAM helps capture these dynamics and explains variation in adoption behaviors.

2.5. Interaction Subjectivity and Human Agency in GenAI Use

A growing body of research emphasizes that effective GenAI use depends not only on attitudes and intention but also on interaction subjectivity, the degree of agency students exercise when engaging with the technology. This aligns with critical perspectives on digital learning, which stress that technology should augment rather than replace human judgment [4]. Building on critical theory and revised Bloom’s taxonomy [21], interaction subjectivity highlights the cognitive and ethical quality of usage. It comprises two dimensions: (1) Critical use: verifying accuracy, recognizing biases, and applying ethical standards when evaluating AI-generated outputs. (2) Exploratory practice: iterative experimentation with prompts, seeking to understand model behavior, and integrating disciplinary knowledge to improve outcomes.

For graduate students, whose research demands originality, interaction subjectivity is indispensable. Low subjectivity risks superficial reliance, outsourcing writing or analysis without deeper engagement [5]. High subjectivity, in contrast, can transform GenAI into a partner for inquiry, strengthening critical thinking and creativity. This construct links adoption behavior with epistemic outcomes, filling a gap left by classical TAM [1]. By foregrounding human agency, it also resonates with international calls for human–AI use in education [10].

2.6. Human–AI Interaction and Cognitive Adoption

Unlike deterministic systems that implement rule-based mappings, outputs of GenAI remain statistically plausible yet not inherently verifiable. Contemporary research about uncertainty in deep neural networks distinguishes between model uncertainty (arising from limited knowledge, stochastic training, and distributional gaps) and data uncertainty (stemming from measurement noise and inherent ambiguity) [22]. Accordingly, reliability in AI systems depends not only on accuracy but on users’ calibration [23].

Research on human creative agency suggests that users’ perceptions of whether AI complements or competes with their capabilities mediate their engagement and regulatory stance [24]. This aligns with human-centered AI perspectives proposing that sustainable human control is crucial in high-automation contexts.

GenAI systems should thus be conceptualized as components within a broader cognition system, in which interaction subjectivity is not peripheral but structurally crucial to AI-mediated cognition.

2.7. Theoretical Gaps in Existing Research

The reviewed literature suggests that while TAM remains a robust framework, the unique context of GenAI in graduate education necessitates theoretical expansion. Specifically: (1) External environment captures how media and interpersonal networks shape early awareness and norms. (2) Risk perception reflects epistemic, ethical, and legal uncertainties central to GenAI. (3) Interaction subjectivity explains how the depth and quality of human–AI engagement determine whether adoption leads to augmentation or dependency.

Despite valuable insights, few studies have examined how users cognitively negotiate these dynamics in real-world interaction contexts. TAM was developed in contexts where technologies functioned as relatively stable and deterministic tools. However, GenAI generates novel outputs and introduces epistemic uncertainty. In such environments, perceived usefulness alone may not adequately capture the cognitive negotiation. Moreover, the recent literature acknowledges concerns regarding ethics and reliability; personal subjectivity is often treated as a peripheral variable in cognitive evaluation, without sufficient attention paid to its interactive relationships with other variables.

To address this gap, the present study employs in-depth qualitative interviews and thematic analysis to identify and construct salient factors and articulate their mechanisms within an extended TAM framework.

3. Methodology

3.1. Research Design

This study adopts a qualitative and theory-informed research design to investigate the mechanisms influencing graduate students’ adoption of GenAI. Since the technology is still emerging and prior empirical evidence remains limited, qualitative methods are particularly suitable for capturing graduate students’ nuanced perceptions, experiences, and behaviors.

The TAM was employed as an analytical scaffold. It provided a conceptual lens for the design of the interview protocol as well as the coding scheme. While TAM traditionally operates within deductive and quantitative paradigms, this study employed it in an abductive manner to guide initial coding categories, remaining open to inductively emergent constructs.

Building on prior studies and the exploratory nature of this research, this hybrid strategy aligns with theory-informed thematic analysis approaches, in which pre-existing theoretical constructs are iteratively refined through empirical data.

Figure 1 presents the extended TAM framework adopted in this study, which integrates both the traditional constructs of TAM and the additional dimensions emerging from the GenAI context. This framework is used as the basis for qualitative analysis.

3.2. Data Collection

Data were collected through one-on-one semi-structured interviews with 20 graduate students in Chinese universities between February and May 2024. A maximum variation sampling strategy was adopted to ensure diversity across degree levels and disciplines. The sample included 12 master’s students and 8 doctoral students, representing both STEM fields (n = 13) and the humanities and social sciences (n = 7). All participants were enrolled at universities classified as “Double First-Class” higher education institutions in China, ensuring that they were engaged in research-intensive academic environments.

Data collection continued until no substantively new categories emerged from interviews. After the eighteenth interview, no additional conceptual dimensions were identified; two further interviews were conducted to confirm saturation stability.

Each interview lasted between 40 and 60 min. Interviews were conducted either in person or online, depending on participants’ availability, and were recorded with their informed consent. The recordings were transcribed verbatim and anonymized to remove identifying information.

Table 1. Demographic Information of Participants.

Participant ID	Gender	Year of Study
A01	Male	Master’s, Year 3
A02	Male	Ph.D., Year 2
A03	Male	Ph.D., Year 4
B01	Female	Master’s, Year 2
A04	Male	Master’s, Year 3
B02	Female	Master’s, Year 1
A05	Male	Master’s, Year 2
A06	Male	Master’s, Year 3
B03	Female	Ph.D., Year 4
A07	Female	Master’s, Year 2
A08	Male	Master’s, Year 2
B04	Female	Master’s, Year 3
A09	Female	Master’s, Year 1
B05	Female	Ph.D., Year 2
A10	Male	Ph.D., Year 1
A11	Male	Ph.D., Year 2
A12	Male	Master’s, Year 1
A13	Male	Ph.D., Year 3
A14	Male	Master’s, Year 2
A15	Male	Ph.D., Year 2

presents the demographic profile of the participants.

3.3. Data Analysis

The interview transcripts were analyzed using NVivo 14 software, following the six phases of thematic analysis proposed by Braun and Clarke [25]. These phases were: (1) familiarization with the data, (2) generating initial codes, (3) searching for themes, (4) reviewing themes, (5) defining and naming themes, and (6) writing the report.

Grounded theory coding strategies were used to add depth and systematicity to the process [26]. In the first stage, open coding was applied to identify meaningful text segments, which produced 21 preliminary themes. In the second stage, axial coding grouped and connected these themes into seven sub-themes: external environment, perceived usefulness, perceived ease of use, risk perception, attitude, behavioral intention, and interaction subjectivity. In the final stage, selective coding refined and integrated the sub-themes into a comprehensive model of adoption mechanisms [27,28].

To enhance the rigor of the analysis, members of this research team all have backgrounds in higher education research and educational technology. Also, they did not have supervisory relationships with all participants. An initial coding framework was independently applied by two researchers to 30% of the transcripts. Intercoder agreement was calculated as percentage agreement across coded segments, yielding an agreement rate of 85.7% at the level of major thematic nodes. After reconciliation discussions, the coding schema was refined and applied to the full dataset. An audit trail of coding decisions was kept to ensure transparency and consistency throughout the process.

3.4. Ethical Considerations

The study was conducted in accordance with ethical guidelines for research involving human participants. All participants were informed of the study’s purpose, the voluntary nature of participation, and their right to withdraw at any point without consequences. Written informed consent was obtained prior to data collection. To safeguard privacy, all transcripts were anonymized, and data were stored securely with restricted access to the research team. The research design and procedures were reviewed and approved by the Institutional Ethics Committee of the host university.

4. Findings

Thematic analysis of 20 interview transcripts generated seven sub-themes that together explain graduate students’ adoption of GenAI.

Table 2. Sub-themes and examples of candidate themes.

No.	Sub-Theme	Definition
1	External Environment	The overall climate and conditions shaped by external information sources and social influences during the formation of attitudes toward GenAI.
2	Risk Perception	The negative impacts and uncertainties that GenAI may bring to users, society, ethics, and technology itself.
3	Perceived Usefulness	The extent to which graduate students perceive GenAI as capable of improving their work or task performance.
4	Perceived Ease of Use	The extent to which graduate students consider GenAI simple, intuitive, and low-cost to learn and operate.
5	Attitude	The overall affective and evaluative orientation of graduate students toward GenAI.
6	Behavioral Intention	The plans and intentions of graduate students to use GenAI in the future.
7	Interaction Subjectivity	The degree of initiative and agency graduate students demonstrate when interacting with GenAI.

summarizes the sub-themes and examples of candidate themes identified during the coding process.

4.1. Technology Perceptions and Attitudes Toward GenAI

Graduate students’ acceptance of GenAI is closely tied to their perceptions of its technical attributes and their resulting attitudes. Across the interviews, two broad orientations emerged: acceptance and resistance. Resistance did not mean outright rejection but rather a tendency to minimize use, for example “only using it when absolutely necessary” (只有在十分必要的时候才去使用). Some participants noted that they would “consciously reduce the frequency of use and avoid becoming overly dependent on it” (会有意识地减少使用频率与依赖性). These different orientations were rooted in how students perceived usefulness, ease of use, and potential risks.

Perceived ease of use was generally high. Most participants described GenAI as simple to operate, requiring no specialized background in computer science. As one student explained, “I can use it just like chatting, by typing basic prompts” (只要像正常聊天一样输入简单的指令就可以使用了). The threshold for using the technology was also considered low; as another participant noted, “you don’t need to have a background in machine learning or large language models to use it smoothly” (不需要拥有机器学习或是大语言模型等计算机科学的学科背景知识也可以顺畅地使用它).

Perceived usefulness was also consistently strong. Students reported that GenAI helped them in different stages of research. Some interviewees highlighted its ability to “quickly search for and provide conceptual and foundational information, helping to build an initial understanding of a research topic” (快速搜索和提供概念性, 基础性的信息, 帮助构建起对研究主题的初步理解). Others emphasized its role in stimulating intellectual exploration, noting that it could “generate new research ideas and broaden my research perspectives” (提供新的科研创意, 拓宽研究思路). In addition, many interviewees valued its practical functions in text-related tasks, such as “polishing languages” (润色语言). Ease of use reinforced the usefulness by lowering the effort required to access these benefits.

At the same time, students recognized potential risks. They frequently mentioned hallucinations, or the generation of plausible but false information, particularly when tasks involved cross-disciplinary expertise. Concerns also included privacy, over-reliance on the tool, and “the possibility of diminished cognitive abilities” (认知能力削弱). Despite these risks, the interviews revealed that perceived usefulness often outweighed caution, which explains why most students maintained a generally positive attitude toward GenAI.

4.2. External Environment: Media Exposure and Interpersonal Networks

Initial awareness and attitudes were strongly influenced by external cues. Media exposure was the most common entry point. Many participants first learned about ChatGPT from online news or social media. One doctoral student recalled, “I saw a report in early 2023 and immediately tried it out”.

Interpersonal networks also played a central role. Supervisors and peers introduced GenAI, shared experiences, or demonstrated practical uses. For instance, one master’s student said, “My advisor encouraged us to try it, so I used it for my coursework” (我的导师鼓励我们使用, 我就在写课程作业的时候使用了几次). Peer groups often served as informal training grounds where prompting strategies were exchanged. As one participant explained, “My friends often share useful prompts in our group chat. If someone finds a good way to phrase a question, they’ll post it, and the rest of us try it out” (我们经常在群聊里分享有效的提示词, 他们发现了有效的提问方式之后我们其他人也会试一试). These influences created a sense of legitimacy that shaped subsequent adoption.

4.3. Behavioral Intention: From Initial Use to Long-Term Habits

Although attitudes mattered, actual adoption was more closely linked to behavioral intention. Initial intentions were often situational, such as using GenAI for inspiration when facing difficulties with assignments. Positive experiences during these first trials frequently led to stronger long-term intentions. One participant explained, “I started using it because I had no idea how to approach an assignment for one of my courses. I thought maybe it could help me generate some ideas, and it actually turned out to be really helpful” (习惯用它是因为当时我对一门课程作业的题目毫无头绪, 我想到可以使用它来帮我找灵感, 结果真的派上了用场).

Over time, behavioral intention evolved into habits and stable motivations. Some students integrated GenAI into daily routines, consulting it regularly for literature summaries or language editing. For others, the main motivation was efficiency: the ability to save time and broaden thinking. Trust in the technology also mattered, though it was secondary to usefulness. Even those with doubts about reliability still used GenAI consistently because it “helped with repetitive tasks and freed them to focus on creative aspects of research” (它可以自动处理很多重复性的任务, 让我能够专注于更有创意和分析性的工作, 这对我来说是最大的吸引力).

4.4. Interaction Subjectivity: Critical and Exploratory Practices

Interaction subjectivity captured how actively students engaged with GenAI. Two dimensions were observed: critical use and exploratory practice. Critical use involved verifying information, rejecting unreliable outputs, and applying ethical standards. Some students described “when dealing with interdisciplinary topics, I checked outputs against multiple sources” (当涉及跨学科内容时会参考不同信息源), while others highlighted giving feedback to the system when outputs violated academic or ethical norms. Less critical users tended to accept outputs at face value, prioritizing convenience. For example, one interviewee said “In most cases, the results it generates are generally reliable, so I don’t feel the need to spend extra time verifying everything” (生成的结果大体上应该都是可靠的, 不需要费时费力去专门查证).

Exploratory practice involved experimenting with prompts, iterating on strategies, and integrating disciplinary knowledge to improve results. Students with high exploratory engagement reported better outcomes, such as refined ideas. Some participants “keep adjusting the prompts and asking follow-up questions, the ideas become much clearer and more specific” (我不断优化提示词, 给出的结果就越来越清晰和具体). By contrast, those with low subjectivity limited GenAI use to repetitive tasks, often without reflection. Overall, interaction subjectivity determined whether GenAI use fostered deep learning and creativity or became a superficial shortcut.

4.5. Mechanisms of Influence

The findings suggest that graduate students’ adoption of GenAI is shaped by the interplay of seven interrelated factors: PU, PEOU, risk perception, attitude, external environment, behavioral intention, and interaction subjectivity.

The external environment including institutional policies, accessibility of technological resources, and peer or media discourse awakens students’ cognitive awareness and provides contextual cues, thereby framing their initial encounter with artificial intelligence. These contextual influences then feed into students’ perceptions, which constitute the cognitive foundation of acceptance. Specifically, PU reflects whether students believe GenAI can enhance their academic efficiency and outcomes, while PEOU captures the perceived level of effort required to integrate GenAI into learning and research routines. In contrast, risk perception highlights students’ concerns regarding privacy, reliability, and academic integrity, which may counterbalance positive evaluations.

Together, these perceptions shape students’ attitudes toward GenAI. A favorable attitude strengthens the likelihood of developing a strong behavioral intention to use the technology, whereas negative or ambivalent attitudes weaken it. Once formed, behavioral intention serves as the bridge between psychological acceptance and actual practice. Early adoption experiences feed back into the system: positive outcomes can reinforce motivation and habitual use, while negative experiences can discourage further engagement.

Crucially, interaction subjectivity moderates this entire process. It reflects how individual agency, peer influence, and the social context of human–AI interaction determine the depth and quality of engagement. For some students, supportive networks and collaborative practices amplify adoption and lead to meaningful academic integration; for others, skepticism or critical peer attitudes may weaken the translation of intention into sustained engagement.

Figure 2 illustrates this integrated model. It depicts how the external environment initiates awareness, how perceptions of usefulness, ease of use, and risk shape attitudes, how attitudes influence behavioral intention, and how interaction subjectivity moderates the outcomes of adoption. By synthesizing these relationships, the model offers a holistic account of the mechanisms through which graduate students adopt and engage with GenAI in higher education.

5. Discussion

From the perspective of cognition, this study investigated the factors influencing graduate students’ adoption of GenAI and the mechanisms through which these factors interact. Drawing on the TAM and extending it with external environment, risk perception, and interaction subjectivity, the analysis produced several important insights.

5.1. Cognitive Mechanisms in GenAI Adoption

This study demonstrates that graduate students’ adoption of GenAI is best understood as a cognitive calibration process rather than a simple decision. The findings indicate that under human–AI interaction contexts, adoption of GenAI consists of dynamic evaluations of usefulness, effort, and risk.

Both PU and PEOU function as signals that reduce perceived interaction costs and increase willingness to integrate AI into research workflows. GenAI was valued for its ability to support different stages of the research process, from building an initial understanding of a topic to generating new ideas and refining academic writing. Its conversational interface and low entry barrier further reinforced perceptions of usefulness. These findings are consistent with earlier studies showing that when digital tools are seen as both effective and easy to use, adoption becomes more likely [11,17].

At the same time, the study highlights determinants that extend beyond the classical TAM. Risk perception emerged as a counterbalancing cognitive factor. Students were concerned about hallucinations, data leakage, and potential over-reliance on GenAI. However, many expressed a willingness to accept these risks because of the clear time-saving and efficiency benefits. This trade-off illustrates a pragmatic orientation: usefulness was prioritized, while risks were acknowledged but often discounted. Similar dynamics have been observed in recent studies, which show that younger cohorts frequently downplay AI-related risks when efficiency benefits are salient [2,5].

The external environment also played a crucial role. Media coverage introduced the technology, often framing it as revolutionary, while interpersonal networks such as peers and supervisors either legitimized or questioned its value. Supervisors’ attitudes in particular influenced students’ perceptions of whether using GenAI was appropriate in academic contexts. These findings resonate with the “social influence” construct in extended acceptance models such as UTAUT [15], but they highlight its specific relevance in graduate education, where academic norms and hierarchies strongly shape student behavior.

The findings further suggest disciplinary and cohort differences in adoption. STEM students tended to emphasize efficiency, scalability, and technical assistance, integrating GenAI into coding, statistical analysis, and information processing. By contrast, students in the humanities and social sciences were more cautious, concerned with originality, argumentation, and the preservation of scholarly voice. This echoes prior work showing that epistemic norms shape how technologies are evaluated [7]. Similarly, doctoral students tended to use GenAI in more selective and strategic ways, for example to test research ideas or cross-check interpretations, whereas master’s students often relied on it for routine support such as text polishing or drafting outlines. These contrasts indicate that adoption cannot be explained by universal constructs alone but is mediated by disciplinary cultures and academic maturity.

Another distinctive contribution of this study lies in identifying interaction subjectivity as a determinant of adoption. Students who engaged critically and exploratively achieved deeper learning outcomes. By contrast, those who used GenAI mainly for repetitive or superficial tasks risked dependency and diminished critical thinking. This dimension extends TAM by showing that adoption is not just about intention to use, but also about a process of cognitive alignment between user expectations, perceived system affordances, and evolving interaction experiences.

5.2. Interaction Subjectivity and Human–AI Co-Adaptation

The findings suggest that adoption unfolds as a dynamic process rather than as a one-off decision. External environment operates at the entry stage, shaping awareness and expectations. Perceived usefulness, ease of use, and risk perception form the cognitive foundation for acceptance or resistance. Attitudes, informed by these perceptions, guide behavioral intentions.

Behavioral intention itself was shown to be dynamic. Initial use was often situational, driven by immediate needs such as completing coursework. If these early experiences were positive, they reinforced trust and efficiency, leading to long-term integration of GenAI into research routines. This transition from short-term to long-term intention highlights the importance of early encounters in shaping positive cognition.

Interaction subjectivity plays a moderating role in this process. It determines whether intention translates into superficial or meaningful adoption. Students with higher levels of subjectivity treated GenAI as a partner for inquiry, integrating its outputs into critical analysis and original thinking. Those with lower subjectivity often relied on default prompts or simple templates, using GenAI as a shortcut rather than as a catalyst for learning. These findings echo educational theories emphasizing that technology can either enhance or undermine learning, depending on how actively users engage with it [21].

On a cognitive level, this distinction is crucial. It indicates that effective human–AI collaboration depends not solely on algorithmic capability but on user-side engagement patterns. Interaction subjectivity functions as a cognitive modulation variable that shapes whether AI use enhances adaptive reasoning or reinforces superficial automation.

The integrated model (Figure 2) captures this interplay. External cues spark awareness, cognitive perceptions shape attitudes, and attitudes influence intentions. Interaction subjectivity then moderates the pathway from intention to actual use, determining whether GenAI enhances academic practices or fosters dependency. This dynamic view shows that adoption is not a linear outcome but a negotiated process shaped by personal, disciplinary, and cultural factors.

Importantly, the moderating role of interaction subjectivity suggests that sustainable adoption depends on the quality of human–AI engagement, rather than on use itself. Graduate students who engage critically and exploratively with GenAI are more likely to integrate it as a tool that supports reflective thinking, originality, and responsible knowledge production. By contrast, low interaction subjectivity risks transforming GenAI into a shortcut that undermines learning processes and academic integrity. These findings extend TAM by shifting attention from intention to interaction quality. Adoption should therefore be conceptualized not merely as frequency of use but as the configuration of cognitive engagement during use.

5.3. Implications for Cognitive Computing and AI-Mediated System Design

This study offers a cognitively grounded model of GenAI, which contributes to cognitive computing research by illustrating how users calibrate usefulness, risk perception and interaction strategies when engaging with probabilistic AI systems. These findings suggest that perceived usefulness functions as a dominant evaluative cue, with users prioritizing efficiency signals when engaging with AI systems. Visualized performance feedback and uncertainty within the interface might help users to cultivate more rational cognition, rather than a singular pursuit of usefulness which leads to uncritical reliance.

Risk perception emerges as an equally important factor with implications for AI system design. Participants’ concerns about hallucination, data reliability, and over-dependence highlight the value of incorporating explainability features, which make reasoning processes more transparent and invite verification. Such considerations may encourage reflective engagement and sustained cognitive oversight.

Interaction subjectivity further indicates that the quality of human–AI collaboration depends not only on quality of algorithm but on the depth of user engagement. Interface structures that prompt comparison, revision, or iterative refinement may help scaffold active cognitive participation, strengthening adaptive human–AI co-evolution rather than passive automation.

Overall, these insights suggest that the effectiveness of human-centered cognitive systems depends not only on algorithmic performance but also on how interface design and interaction structures support users’ reflective engagement and cognitive regulation during human–AI interaction.

6. Conclusions

6.1. Main Findings

This study examined the mechanisms influencing graduate students’ adoption of GenAI through an extended TAM. Based on qualitative interviews and thematic analysis, the findings identified seven determinants: perceived usefulness, perceived ease of use, risk perception, attitude, external environment, behavioral intention, and interaction subjectivity. The results indicate that usefulness and ease of use remain the important drivers of adoption, but they are insufficient to explain the complexity of graduate students’ behaviors. Risk awareness, media exposure, social influence, and above all the quality of students’ engagement with the technology also play crucial roles. Adoption is not a one-off decision but a process of cognitive calibration, shifting from initial trials to long-term integration, and it is moderated by how actively students interact with GenAI.

6.2. Theoretical Contributions to Cognition of GenAI

Classical TAM assumes relatively stable technological affordances. However, GenAI systems generate novel and uncertain outputs. By framing adoption as cognitive calibration under epistemic uncertainty, this study extends TAM from a belief-intention model toward a dynamic cognitive adaptation framework. This study indicates that the integration of risk perception and personal subjectivity moves beyond efficiency-centered models of acceptance. In GenAI contexts, adoption involves simultaneous evaluation of performance and epistemic risks. Moreover, the integrated model proposed in this study shifts the analytical focus from “whether users adopt” to “how users cognitively co-adapt with AI systems”.

6.3. Implications for Human–AI Interaction Design

In conclusion, this study offers several implications for the design of cognitively adaptive human–AI systems. Understanding these mechanisms is essential for ensuring that GenAI becomes a catalyst for responsible innovation rather than a shortcut that weakens independent scholarship. These findings underscore that effective human–AI collaboration depends not only on algorithmic capability but also on users’ subjectivity and patterns of cognitive engagement. Designing systems that invite reflection, verification, and cognitive control may enhance adaptive co-evolution between human reasoning and generative models, thereby supporting the development of more effective human-centered cognitive computing systems.

7. Contributions and Limitations

7.1. Theoretical Contributions

This research reframes the Technology Acceptance Model (TAM) as a framework for understanding the cognitive adoption of GenAI in graduate education. Rather than treating adoption as a purely individual decision, the framework considers it through three interconnected lenses. The adoption is situated within the external environment, revealing that such decisions are embedded in media narratives, supervisor–student relationships, and peer networks rather than being purely individual. Ultimately, this framework introduces interaction subjectivity as a new dimension, where the quality of engagement determines whether adoption enhances or undermines academic practice.

7.2. Practical Contributions

Beyond theoretical insights, this study provides cognition-oriented implications for the governance of human–AI interactions. By identifying graduate students’ cognitive biases, usage patterns, and challenges in engagement with GenAI, the findings inform strategies for more adaptive and calibrated interaction with GenAI systems. These strategies include enhancing epistemic transparency, designing systematic training programs that strengthen critical and reflective engagement, and fostering supportive environments that strengthen adaptive collaboration between human reasoning and GenAI.

7.3. Limitations and Directions for Future Research

Despite its contributions, this study has several limitations. The sample was limited to 20 graduate students from research-intensive “Double First-Class” universities in China. This may restrict generalizability as students in highly resourced academic environments may have greater exposure to GenAI tools and stronger technological competencies than those in less research-intensive institutions.

Another limitation is that self-reported interviews may not fully reflect actual usage behaviors. Particularly, discrepancies may exist between stated cognitive strategies and real-time engagement practices. Future studies should therefore adopt mixed methods, combining interviews with log data, surveys, or classroom observations to capture both perceptions and practices.