Pedagogical Applications of Generative AI in Higher Education: A Systematic Review of the Field

Research on GenAI in education is predominantly concentrated in China (including Hong Kong, which is administered as a Special Administrative Region with a high degree of autonomy), with a combined total of 11 publications highlighting a strong regional focus on GenAI-enhanced learning. The U.S. follows with eight publications. Beyond these regions, contributions span multiple continents, including Europe (Finland, Germany, Netherlands, Serbia, Sweden, UK), Asia (Singapore, Taiwan, Thailand, Turkey), Africa (South Africa, Ghana), North America (Canada), South America (Brazil), and Oceania (Australia).

China's national strategy, exemplified by the Next Generation Artificial Intelligence Development Plan (2017), has prioritized AI innovation, directing substantial funding toward educational applications and encouraging universities to align their research with national objectives. This approach has accelerated empirical studies on GenAI in higher education (Knox, 2023; State Council of China, 2017). Additionally, Chinese higher education institutions, including those in Hong Kong, place a strong emphasis on STEM disciplines, facilitating the rapid integration of GenAI tools in fields such as coding and engineering. These policy and cultural factors contribute to China and Hong Kong's current dominance in applied GenAI research publications in education Figure 2.

Full size image

Figure 3 illustrates the distribution of GenAI tools, highlighting ChatGPT as the dominant choice by a substantial margin. Nevertheless, customized GPT-based applications are gaining traction, indicating a growing interest in AI tools tailored to specific educational contexts. Notable examples include AnatomyGPT for anatomy education (Collins et al., 2024), AI Learning Companion Systems (LCS) designed to enhance students'self-efficacy in information literacy (Hu et al., 2024), and retrieval-augmented generation (RAG) chatbots, which blend retrieval and generative capabilities for more sophisticated AI interactions (Guo et al., 2024). Additionally, beyond text-based outputs, tools for image and video generation are emerging, expanding the scope of GenAI applications (Cummings et al., 2024; Koh et al., 2024; Netland et al., 2025; Tsao & Nogues, 2024 Furthermore, writing support tools like Wordtune, Rytr, Wordtune, and specialized platforms such as Rytr assist students in improving their academic writing (Cummings et al., 2024; Koh et al., 2024; Tsao and Nogues, 2024). GenAI is also being increasingly integrated into game-based learning environments to facilitate adaptive learning as an embedded feature (Song et al., 2024).

Full size image

Despite the diversity of available tools, ChatGPT continues to dominate due to its ease of use (especially version 3.5), its effectiveness in text-based academic tasks (e.g., writing, critical thinking), and its general suitability for higher education contexts (Kasneci et al., 2023). However, specialized GPT tools, like AnatomyGPT, have emerged to overcome specific limitations associated with general-purpose platforms such as ChatGPT, Google Gemini, and Claude. These specialized GPTs provide tailored capabilities for discipline-specific teaching and learning needs. However, tools involving visual media—such as image and video generation applications—often face higher barriers, including substantial technical requirements, limited applicability in non-visual disciplines, and significant ethical concerns around misinformation and copyright (Bender, 2021). Hybrid AI applications, such as virtual reality (VR) combined with AI (Muengsan & Chatwattana, 2024), face even greater hurdles, demanding interdisciplinary collaboration and substantial investment in resources.

Overall, the current trend suggests a preference for GenAI tools that prioritize simplicity, accessibility, and alignment with core educational tasks, such as writing and critical thinking (Kasneci et al., 2023). Nonetheless, the exploration of more specialized, immersive, and hybrid AI experiences highlights an important avenue for future development—one that promises richer, more interactive learning opportunities, but requires careful attention to practical, ethical, and collaborative considerations.

Figure 4 illustrates the distribution of GenAI application studies by academic discipline. To categorize these 37 studies, each was coded based on its primary educational context. Studies with disciplinary overlaps—such as those combining Education with ESL or Computer Science—were classified according to their predominant pedagogical focus. Note that the category"Education"broadly includes studies addressing the process of improving or expanding educational systems, practices, and outcomes, including general or liberal education, teacher preparation, and critical skill development (e.g., creativity, critical thinking, learner autonomy, and prompt literacy).

Full size image

The 37 empirical studies represent a diverse range of academic disciplines, with Education, Writing and English Language Learning (including ESL/EFL), and Computer Science emerging as the most prominent. Education accounts for 15 studies, reflecting its broad applicability in pedagogical innovation and learning enhancement. For example, Yang et al. (2024) examined student agency facilitated by GenAI, while Van den Berg and du Plessis (2023) explored ChatGPT’s use in lesson planning within teacher education. Writing and English Language Learning followed closely with 13 studies, divided into Writing/Composition (5 studies)—as exemplified by Bedington et al.’s (2024) research on AI-supported writing processes—and ESL/EFL (8 studies), such as Moorhouse et al. (2024), who investigated first-language (L1) use in second-language (L2) classrooms, highlighting GenAI’s role in linguistic and compositional skill development. Computer Science, comprising 5 studies, emphasizes technical innovation and practical applications, including Zhong and Kim (2024) utilization of ChatGPT for R programming in business analytics and Allen et al.’s (2024) development of the specialized"Q-Module-Bot."Other disciplines, such as Business (2 studies), Biology (1 study), and Medicine (1 study), are comparatively underrepresented.

The dominance of Education, Writing and English Language Learning, and Computer Science reflects GenAI’s strong alignment with pedagogical needs, technical suitability, and practical demands within these fields. Education’s prominence can be attributed to its versatility, using GenAI to foster essential cross-disciplinary skills like creativity, critical thinking, and digital literacy (Selwyn, 2022). Moreover, its application in teacher education aligns closely with contemporary educational priorities, including technological fluency and pedagogical advancement (Prensky, 2012; U.S. Department of Education, 2023).

Similarly, the prominence of Writing/Composition (Cummings et al., 2024; Farazouli et al., 2024; Nguyen et al., 2024; Tsao and Nogues, 2024) and ESL/EFL studies (Escalante et al., 2024; Guo & Li, 2024; Guo et al., 2024; Waluyo & Kusumastuti, 2024; Wiboolyasarin et al., 2024) stems from GenAI’s inherent strengths in text-based tasks, including grammar correction, writing support, and conversational practice (Hwang and Chang, 2023). This emphasis aligns closely with Crompton and Burke’s (2023) systematic review findings, reaffirming ChatGPT’s central role in language acquisition and writing support within higher education contexts.

Computer Science’s representation (Allen et al., 2024; Groothuijsen et al., 2024; Song et al., 2024; Yilmaz and Yilmaz, 2023) aligns with previous findings by Zawacki-Richter et al. (2019) on pre-GenAI programs, reflecting the discipline’s natural synergy with AI capabilities in programming, data analysis, and technical problem-solving, as well as institutional readiness to adopt innovative technological solutions. In contrast, fields such as Biology (Collins et al., 2024), Business (Milić et al., 2024; Netland et al. 2025), and Medicine (Hudon et al., 2024) remain underexplored, largely due to challenges in adapting general-purpose GenAI models to specialized or niche contexts.

Overall, the existing literature prioritizes text-centric and technically aligned disciplines, emphasizing ChatGPT’s prevalent use. Moving forward, there is a clear need for future research to expand the exploration of domain-specific GenAI applications, extending beyond the current dominance of ChatGPT to uncover richer, more diverse possibilities for GenAI-supported learning.

Automated feedback and assessment are a cornerstone of GenAI’s pedagogical utility. The nine codes related to these functions can be grouped into three key themes: immediacy, personalization, and continuous or formative evaluation. For example, Collins et al. (2024) demonstrated the power of this approach with AnatomyGPT, a customized AI tool for anatomical sciences education. AnatomyGPT assesses students'mastery using National Board of Medical Examiners sample items, and provides immediate, detailed feedback complete with rationales and citations. Similarly, Hudon et al. (2024) showed that ChatGPT can generate medical education assessments comparable to expert-designed Script Concordance Tests, highlighting the tool's reliability in creating valid evaluation instruments. Tang et al. (2024) extended these capabilities by harnessing ChatGPT to evaluate critical thinking skills in online peer feedback, automating the assessment process, and delivering immediate insights that guide student improvement. Meanwhile, Txirides et al. (Tzirides et al., 2024) employed GenAI to offer rapid feedback on writing tasks, enabling students to iteratively refine their work in real time. Additionally, Beddington et al. (2024) and Du et al. (2024) used ChatGPT to generate formative feedback, such as executive summaries and counterarguments, that students then use to polish their compositions. These studies illustrate how GenAI not only streamlines the assessment process but also enriches learning by delivering timely, personalized feedback. This dual role helps reduce the burden on instructors while enhancing student engagement and supporting continuous improvement.

Learning support is the second foundational application of GenAI. The 19 codes related to learning support can be grouped into three types: immediacy/efficiency, affective support, and cognitive support. Immediacy and efficiency are key advantages of GenAI due to its ability to provide rapid, targeted assistance. For example, Allen et al. (2024) demonstrated this with Q-Module-Bot—a Q&A bot tailored for biology education that delivers immediate, module-specific responses to student queries. Similarly, Zhong and Kim (2024) extended this approach to business education by using ChatGPT to generate R code that simplifies logistic regression for students with limited programming skills. Van den Berg and du Plessis (2023) further amplified this support by employing ChatGPT to create lesson plans, worksheets, and visual presentations for teacher education.

Beyond efficiency, GenAI tools provide significant affective support by fostering an encouraging and patient learning environment. Hu et al. (2024) illustrated this with an AI Learning Companion that offers tailored assistance, helping students navigate course material independently. Studies also reveal that the social presence of GenAI tools, being consistently encouraging and patient, enhances students’ overall learning experiences. GenAI also plays a crucial role in cognitive support, acting as a scaffold for brainstorming, practice, and real-world scenario applications. For instance, Yilmaz and Yilmaz (2023) showcased how ChatGPT assists in programming education by providing clear code examples and detailed explanations, which reinforce key computational concepts and facilitate deeper understanding. Together, these examples underscore GenAI's role as a versatile scaffold—bridging gaps in understanding and empowering students with instant, context-specific support across various dimensions of learning.

The third pedagogical application of GenAI—development of critical skills—encompasses 19 codes targeting higher-order competencies, including creativity, critical thinking, learner autonomy, and emergent prompt literacy. This theme is distinguished by its focus on metacognitive growth and adaptive skill-building, rather than immediate task support. For example, codes such as"divergent thinking"and"reflective learning"were grouped here because they prioritize fostering intellectual independence and problem-solving agility, while"prompt literacy"reflects the growing necessity for learners to strategically engage with AI as a collaborator. Unlike transactional supports (e.g., cognitive aids), these codes emphasize foundational capacities that enable learners to navigate complex, evolving educational landscapes, aligning with GenAI’s transformative potential to cultivate lifelong, self-directed learners.

Creativity flourishes as GenAI inspires novel outputs across disciplines. Bedington et al. (2024) demonstrated this in a professional writing course where ChatGPT generates social media posts and summaries, prompting students to refine drafts into polished work. Tsao and Nogues (2024) extended this to creative writing, integrating multimodal tools like Midjourney and Stable Diffusion with ChatGPT to support storytelling and graphic narratives. Similarly, Essel et al. (2024) highlighted AI-assisted brainstorming in research methods courses, enabling students to generate original research ideas. Huang et al. (2024) leveraged ChatGPT’s generative capabilities to enhance creative ideation in a product design course. Muengsan and Chatwattana (2024) further illustrated this in game-based learning, where GenAI designed educational content, fostering imaginative engagement. These examples showcase GenAI’s role as a creative scaffold.

Critical thinking is strengthened as students analyze and refine GenAI outputs. Tzirides et al. (2024) demonstrated this through students evaluating AI-generated feedback for accuracy and bias, enhancing analytical skills. In teacher education, Van den Berg and du Plessis (2023) had instructors critique ChatGPT-generated lesson plans, such as an ESL prepositions lesson, adapting them to specific classroom needs, promoting reflective judgment. Tang et al. (2024) used ChatGPT to assess critical thinking in peer feedback, requiring students to scrutinize AI evaluations. Pinochet et al. (2023) took this further, encouraging ethical critiques of AI in collaborative tasks and fostering deeper critical perspectives and active evaluation.

Learning autonomy is supported by GenAI’s immediacy and efficiency, reducing instructor dependence. Yang et al. (2024) enhanced student agency in a postgraduate course, where learners independently explored tasks using chat logs and journals. Students who took an active role in their learning (e.g., resourceful and reflective approaches) were more likely to benefit from GenAI, while those who were passive/receptive or resistant may not fully leverage its potential. Hu et al. (2024) introduced an AI Learning Companion that provides tailored, autonomous support, while Allen et al. (2024) presented Q-Module-Bot, enabling students to seek answers independently. Guo et al. (2024) offered EFL students AI-generated exercises for self-paced learning. These studies have underscored GenAI’s potential to empower learning autonomy and learner agency.

Prompt literacy is emerging as a critical skill in the AI era, as students increasingly refine their ability to craft effective queries for AI systems. Yilmaz and Yilmaz (2023) underscored its importance in programming, demonstrating how precise prompts enable ChatGPT to generate accurate code. Similarly, Guo et al. (2024) explored its application in EFL instruction, where tailored prompts enhance language learning exercises. Bedington et al. (2024) further advanced the concept by introducing the “rhetoric of prompting,” a framework for refining AI-generated outputs. Meanwhile, Zhong and Kim (2024) highlighted its role in R code generation, showing how prompt literacy can deepen analytics education.

Although these critical skills have been studied individually, future research should examine their interconnections to maximize GenAI’s impact. Creativity sparks ideas, critical thinking refines them, autonomy drives exploration, and prompt literacy unlocks GenAI’s full potential—together fostering a more cohesive and transformative skill development approach.

The 42 codes, which coalesced into three overarching themes, have been central to pedagogical adoptions of GenAI in its initial two years of use within teaching and learning. They draw on GenAI’s core strengths—natural language processing, adaptability, and scalability (Kasneci et al., 2023), while simultaneously addressing key institutional priorities such as improving efficiency, enhancing student engagement, and fostering skill development. This synergy positions GenAI as an essential tool in the evolving landscape of higher education (Partnership for 21 st Century Learning, 2019). In addition, these applications align with broader educational imperatives frequently discussed in scholarly literature.

The most commonly cited challenges in integrating GenAI into teaching and learning revolve around technical limitations, usability barriers, and scalability constraints. These challenges are well-documented in empirical studies. Technical issues, such as unreliable query handling and system glitches, affect tools like Q-Module-Bot (Allen et al., 2024) and Fermat (Cummings et al., 2024), while self-made RAG chatbots often generate inaccurate content (Guo et al., 2024). Similarly, the AI Learning Companion’s dependence on ChatGPT’s inconsistent accuracy (Hu et al., 2024) underscores the need for more robust technical infrastructure. Usability barriers further hinder adoption, with 50% of users finding Q-Module-Bot’s interface non-intuitive (Allen et al., 2024), while Fermat’s confusing spatial canvas frustrates students (Cummings et al., 2024). Additionally, students struggle with the steep learning curve of crafting effective prompts (Guo et al., 2024), emphasizing the need for user-friendly design and better training (Hu et al., 2024). Sustainability remains another major concern, particularly regarding scalability and resource constraints. The Q-Module-Bot’s reliance on local CSV storage limits its expansion (Allen et al., 2024), and highly customized RAG chatbots restrict broader applications (Guo et al., 2024). Furthermore, resource integration challenges seen in the AI Learning Companion (Hu et al., 2024) and tools like Elicit and Wordtune (Cummings et al., 2024) highlighted the need for scalable, resource-efficient solutions to support GenAI’s long-term viability in education.

These issues mirror the early adoption phases of previous technologies, such as classroom computers and virtual learning environments, where initial glitches, user unfamiliarity, and scalability constraints required iterative refinement and investment (Rogers, 2003). Addressing these challenges demands a multifaceted approach, including robust technical development, user-centered design, and scalable infrastructure (Sillitoe, 2017). Without such efforts, the potential of GenAI to transform education will remain slow to be realized.

Concerns about GenAI’s reliability—particularly misinformation, bias, and unreliable feedback—persist across recent research. Escalante et al. (2024) found that GPT-4 occasionally generates inaccurate or harmful content, emphasizing the need for stringent oversight. Farazouli et al. (2024) highlighted ChatGPT’s tendency toward repetitive and incoherent responses, undermining its educational utility. Van den Berg and du Plessis (2023) linked inaccuracies to outdated training data and the absence of real-time information, while Hudon et al. (2024) stressed the need for AI-generated medical content to align with expert-validated knowledge, especially in fields demanding high precision.

Ethical challenges—including bias, privacy risks, and academic integrity threats—further complicate GenAI’s adoption. Bedington et al. (2024) and Habib et al. (2024) documented biases in ChatGPT that misrepresented human experiences or fabricated information, requiring users to evaluate outputs critically. Escalante et al. (2024) exposed weaknesses in safeguards against misuse, while Van den Berg and du Plessis (2023) raised concerns about copyright infringement and plagiarism in AI-generated texts. Smerdon (2024) and Waluyo and Kusumastuti (2024) pointed to academic integrity risks, noting that AI enables students to outsource assignments, challenging traditional notions of authorship and intellectual ownership.

Addressing these challenges requires comprehensive training for educators and students to recognize and mitigate GenAI’s limitations, alongside strong ethical frameworks to guide responsible use (Bond et al., 2024; Francis et al., 2025; Yan et al., 2024). The U.S. National Institute of Standards and Technology’s AI Risk Management Framework (2023) provides structured approaches for detecting and correcting biases and inaccuracies, offering a model for educational applications. Similarly, the U.S. Department of Education (2023) emphasizes equitable implementation and ongoing professional development to ensure GenAI enhances learning without exacerbating disparities or enabling misconduct. While ethical concerns have long been part of educational technology, their scale and complexity have expanded with GenAI, making proactive measures more critical than ever.

As discussed earlier, automated feedback and assessment are the primary pedagogical applications, but they also present significant challenges. Bedington et al. (2024) observed that while ChatGPT could summarize drafts, it missed critical points students deemed essential, necessitating human revision to ensure feedback addressed rhetorical intent. Escalante et al. (2024) suggested that while AI can generate feedback on writing, its output is not inherently reliable and requires scrutiny to ensure accuracy and appropriateness in an educational context. These findings suggest that while GenAI provides accessible, scalable feedback, studies suggest that human oversight is necessary to ensure nuanced understanding and support. This balance is particularly crucial in areas requiring ethical judgment or deep critical analysis, as noted in Farazouli’s (2024) work on AI in teacher assessment practices.

Cheating risks and assessment validity were recurring concerns. Pinochet et al. (2023) warned of students outsourcing assignments to ChatGPT, threatening traditional evaluation methods. Bedington et al. (2024) critiqued AI detection tools for generating false positives and unfairly penalizing students. Waluyo and Kusumastuti (2024) reported difficulties in assessing originality in AI-assisted English writing as students might use AI as “an easy way to finish assignments quickly,” bypassing deep engagement, a concern echoed by teachers’ call for “judicious” use to maintain integrity (p. 8). Such issues demand rethinking assessment design, feedback, and academic integrity policies.

These challenges—AI feedback’s lack of depth, obscured assessment of student work, and the need for human oversight—underscore a critical pedagogical tension: GenAI’s efficiency can enhance scalability but risks diluting the personalized, critical engagement central to higher education. Addressing this requires innovative assessment strategies and instructor training to integrate AI effectively while preserving educational rigor.

A lack of AI literacy among students, coupled with over-reliance, emerged as a significant challenge to effectively leveraging GenAI in higher education. Knoth et al. (2024) emphasized that non-expert users—those without formal AI training—struggled with unsystematic and trial-and-error approaches to crafting prompts for large language models, often overgeneralizing expectations from human interactions, which hampered their ability to elicit high-quality outputs. Similarly, Song et al. (2024) addressed this literacy gap in their development of LearningverseVR, noting the difficulty of prompt writing and the risk of"prompt word attacks"that could destabilize AI responses, necessitating a nested prompt engineering design to reduce user burden and prevent misuse (p. 4). Together, these studies underscore how inadequate AI literacy undermines GenAI’s educational potential, requiring structured support to enhance prompt engineering proficiency among students and educators.

Compounding this literacy deficit, overreliance on GenAI threatens potentially the development of foundational skills, particularly in critical and creative domains. Gao et al. (2024) observed that business students overly dependent on ChatGPT saw their independent problem-solving abilities diminish, a pattern echoed in programming education by Groothuijsen et al. (2024), where AI support reduced peer collaboration and weakened students’ capacity to tackle challenges independently. Xie et al. (2024) found that excessive AI interaction reduced social presence and learning autonomy, while Yilmaz and Yilmaz (2023) found AI tools did not mitigate motivation gaps in programming courses, leaving less active students sidelined in group tasks. Faculty concerns amplify these findings. Essel et al. (2024) reported adverse cognitive effects from students outsourcing critical thinking, and Habib et al. (2024), alongside Tsao and Nogues (2024), warned that AI’s generic outputs might foster cognitive fixation, stifling creative agency.

Collectively, these studies highlight a tension between GenAI’s efficiency and the risk of overreliance, which can undermine deep learning if not thoughtfully managed. While GenAI offers rapid support, unchecked dependence may diminish critical thinking, creativity, and autonomy—core pillars of higher education. Addressing this challenge requires more than technical solutions; it demands a pedagogical shift. The National AI Initiative (2021), a U.S. policy framework, advocates embedding AI literacy into curricula, suggesting that courses on prompt design and AI limitations could help mitigate overreliance. Similarly, EDUCAUSE (2023) underscores the importance of faculty development programs that model balanced AI integration, ensuring tools like ChatGPT serve as supplements rather than substitutes for intellectual effort. Without such strategies, GenAI’s potential may be compromised, fostering dependency instead of enhancing the very skills it aims to support.