AI for Education: A Living Catalog of Emerging Practice

What is the Living Catalog?

The Living Catalog of Generative AI for Education is an evolving resource designed to support both researchers and educators in exploring how AI is being used in teaching and learning. Developed by the Center for Learning Sciences at EPFL, the catalog maps a range of applications of generative AI in education.

Our goal is to point users to relevant scientific literature and examples of real-world practice. By doing so, we aim to provide both a starting point and practical guidance for those seeking to implement generative AI in pedagogically sound ways.

The SAMR Framework

The SAMR framework (Puentedura, 2009; Belkina et al., 2025) provides a useful lens for interpreting the spectrum of AI applications in education.

🔄 Substitution AI performs traditional tasks without changing their fundamental nature.

📈 Augmentation AI improves the efficiency or quality of existing processes.

🔧 Modification AI introduces substantial changes to how processes are carried out.

🚀 Redefinition AI enables entirely new educational experiences that were previously unattainable.

How the Catalog is Built

🎯 Expert-Defined Practices We identify key trends in the educational use of generative AI, grounded in research, theory and practice.

➡️

📄 Evidence from the Field We use open-access scientific literature to bring evidence-based insights from real-world AI applications in educational settings.

➡️

🤖 AI-Powered Distillation We use large language models to synthesize vast amounts of information, making complex research data more accessible and usable.

➡️

⚙️ Automatic Updating The pipeline from querying academic databases to extracting structured data is designed for a high degree of automation, systematicity and traceability.

Warning

Critical Use Encouraged – We do not verify the quality of the referenced studies. Users are encouraged to read the original articles and apply their own critical judgment.

---

AI for Education

Explore each practice in its dedicated space, featuring a Synthesis with key applications and a filterable Table of real-world implementations. Select a practice below.

Synthesis

Applications

Generative AI is being widely adopted across disciplines to substitute, augment, modify, and redefine various facets of feedback provision in education. At the substitution level, it primarily reduces teacher workload by supporting qualitative feedback delivery for assignments and exercises, while augmentation can enhance feedback quality through alignment with established pedagogical frameworks. Modification transforms feedback administration through automation, enabling immediate delivery and possible integration with learning management systems. Redefinition more profoundly reimagines feedback provision, encompassing innovative approaches such as feedback within roleplay scenarios and simulation-based learning environments.

Substitution: Generating Feedback

Given generative AI’s sophisticated linguistic capabilities, the most extensive application has been in providing feedback on written assignments, including for example essays, reports, and translation tasks, where AI systems offer linguistic corrections, structural improvements, and guidance on content quality and argumentation [2], [6], [8], [19], [32], [40], [60], [81], [84], [95], [97], [98], [111], [124], [128], [129], [132], [167], [170], [192], [194], [196], [206], [213], [218], [220], [224], [226], [133], [142], [23], [7], [33], [22], [160], [93], [5], [11], [45], [88].

Feedback on programming assignments represents a natural extension of these linguistic abilities, as programming languages themselves consist of syntax, semantics, and structural rules. Generative AI can identify syntactic or logic errors, and provide feedback on code efficiency, style, readability and documentation [12], [49], [50], [53], [54], [70], [61], [62], [63], [64], [76], [77], [83], [157], [160], [166], [180], [191].

Beyond purely linguistic feedback on natural or programming languages, AI can be employed to provide feedback that draws upon knowledge domains. This involves evaluating students’ understanding of subject matter, assessing the quality of their reasoning and problem-solving approaches, and their disciplinary performance [36], [91], [99], [117], [122], [28], [71], [46], [48], [3], [110], [195], [161], [56], [15], [101], [18], [82]

Augmentation: Enhanced Feedback Design

Generative AI can enhance educational feedback in multiple ways: alignment with specific learning objectives, adherence to established pedagogical frameworks, application of evidence-based feedback principles, incorporation of domain-specific knowledge of errors, implementation of thoughtful design principles, and feedback delivery through diverse modalities and styles [19], [168], [85], [9], [109], [230], [3], [8], [27], [172], [197], [65], [63], [82], [171], [73].

Modification: Automatic Integrated Feedback

The affordance of generative AI to generate adapted content on the fly enables modification of feedback delivery in several ways, including automation of feedback provision and it’s integration with learning platforms, and realtime or immediate delivery of feedback that is personalized or adapted to individual learner needs by representing a dynamic student state. These applications are prevalent in the computer science domain [157], [166], [53], [69], [63], [228], [85], [74], [190], [66], [67], [79], and in the educational field more generally [173], [225], [15], [176], [3], [28], [156], [170], [82], [171], [222]. Realtime delivery of feedback is particularly relevant for language learning, where students can get feedback on speaking performance during realtime language practice [149], [108], [9], [137], [24], [230].

Redefinition: Interactive Feedback and Simulations

Finally, generative AI can redefine feedback integration by creating roleplay and simulation-based learning environments where feedback is either embedded within the interactive experience or delivered immediately afterward. These applications are most prevalent in the medical field [158], [25], [26], [43], [89], [94], [115], [118], [238], [29], [99], with some exceptions in physics [59], [202] and teacher education [13]

Measures and Outcomes

Feedback Quality

A first important outcome is the actual quality of AI-generated feedback itself, and how it compares to feedback created by human experts. Several studies have argued that AI-generated feedback can be effective as a support tool for teachers [22], [160], [5], [73], [220] and/or be comparable to or exceed human feedback [155], [11], [23], [53], [28], [3], [61], [84], [122]. However, other studies have reported mixed results, noting that AI-generated feedback can be ineffective or less effective in some areas [12], [98], [191], [192], [13] and/or low comparability with human feedback [41], [6], [86], [18], or inconsistent and unreliable [156], [62], [76], [50].

Feedback Perception

Beyond the objective quality of AI-generated feedback, students’ subjective experience of this feedback represents a crucial factor in determining its ultimate educational impact. Research reveals that students’ pre-existing attitudes toward AI technology influence their perception of feedback quality beyond the actual content of the feedback itself [169], [168]. Interestingly, students often cannot reliably distinguish between AI-generated and human-generated feedback, frequently rating AI feedback quite favorably when they are unaware of its source [83], [46]. However, student preferences vary considerably when the source is disclosed. Some studies report favorable attitudes toward AI feedback or relatively even preferences between AI and human-generated responses [97] [93], [151]. Conversely, other research indicates that students prefer human feedback and express skepticism toward AI-generated responses [52], [4]. This perception deficit can be partially addressed through the design of natural, flexible interaction experiences[57].

Engagement

Several studies have investigated the impact of AI-generated feedback on student engagement, with many reporting positive effects across cognitive, affective, and behavioral dimensions [141], [194], [44], [222], [176], [15], [59]. This enhanced engagement with feedback has been associated with improved subsequent revision behaviors [66], [72], [167], [32]. However, research also reveals potential drawbacks. AI-generated feedback can overwhelm students cognitively and prove time-consuming to process effectively [7], [15], [40]. Additionally, excessive usage patterns have been linked to decreased performance outcomes [157]. More broadly, some studies suggest that student engagement with feedback remains generally low regardless of source, showing little improvement when AI-generated feedback is provided compared to human feedback [109], [81].

Educational Outcomes

Research examining the actual educational impact of AI-generated feedback has mostly painted an encouraging picture. In writing and language learning contexts, AI feedback has been associated with improved skills and performance [95], [196], [45], [114], [128], [193], [226], [154], [140], [142], [184], [108], [171]. Similar positive outcomes extend beyond language domains, with studies documenting enhanced learning performance across various educational disciplines [89], [201], [85], [15] as well as improved academic achievement [222], [172], [79], [97], sometimes even suggesting that AI feedback may help reduce disparities in learning outcomes [176]. These benefits are often mediated by increased student engagement [235], [123], [59]. However, these positive findings are not universal. Some studies have explicitly reported a lack of significant learning gains from AI-generated feedback [56], [157].

Beyond domain-specific learning outcomes, AI-generated feedback has been linked to various psychological and metacognitive benefits. These include increased motivation [154], [85], [94], reduced anxiety [42], [144], enhanced autonomy and perceived competence or self-efficacy [44], [165], [219], [137], and improvements in self-regulated learning [9], [172].

Limitations and Recommendations

Enhancing AI Feedback

A critical challenge is countering generic AI-generated feedback that lacks both domain-specificity and pedagogical effectiveness. To address this limitation, researchers have explored several promising approaches. Some studies implement retrieval-arugmented generation (RAG) to ensure that feedback is contextually appropriate by drawing from relevant educational resources and domain-specific knowledge bases[48], [49], [230]. Other research has focused on fine-tuning language models using specialized datasets tailored to particular academic disciplines, or employing reinforcement learning techniques to guide models toward producing feedback that aligns with educator objectives [207], [230], [3]. These approaches aim to move beyond one-size-fits-all feedback toward more specialized, contextually appropriate responses.

Multi-step AI systems can further enhance feedback quality by decomposing the complex feedback generation process into distinct, specialized stages. Rather than attempting to evaluate student work and generate pedagogical guidance simultaneously, these systems can separate submission assessment from instructional feedback delivery. For instance, one component might focus exclusively on evaluating content quality and another on writing quality, while a subsequent stage incorporates pedagogical principles to craft appropriate feedback. A third validation layer can then assess the generated feedback against established quality criteria to ensure accuracy and educational effectiveness. This modular approach can be implemented through various architectural strategies. Multi-component systems may combine LLM-based modules with traditional rule-based components, capitalizing on the strengths of each approach. Alternatively, multi-agent frameworks can deploy several instances of a language model, each guided by distinct system prompts that define specific roles in the feedback pipeline. Some implementations strategically match different LLMs to particular tasks based on their relative strengths [110], [60], [68], [163], [53], [99], [230].

Designing Effective Human-AI Interaction

Studies consistently concur that hybrid human-AI workflows, where teachers curate and moderate LLM output, are preferable to fully autonomous tutors. These approaches combine scalable automation with pedagogical expertise, offering complementary strengths where teachers can for example address higher-order concerns while AI handles surface-level feedback [173], [230], [167], [161].

Furthermore, effective design requires balancing educational goals with practical usability while navigating inherent tensions between learning effectiveness and user satisfaction. Key principles include promoting transparency about AI limitations, maintaining instructor control, and creating natural, flexible interaction experiences [63], [57]. Systems must also leverage generative AI’s capabilities to promote inclusion and ensure equitable access for diverse learner populations [225], [208].

However, preventing over-reliance on AI tools remains crucial to avoid diminishing critical thinking and creativity. Both educators and students require comprehensive training in AI literacy and feedback literacy to ensure responsible and effective use of these technologies. These efforts should operate within regulatory frameworks that prioritize responsible use and support ethically grounded pedagogical implementations rather than purely technical solutions [198], [203], [159], [223], [232], [214], [174], [153].

System Implementation Recommendations

Successful deployment requires attention to system scalability and performance optimization to handle educational workloads effectively [54], [20], [228], [171]. Furthermore, privacy compliance represents a critical concern, as educational data requires special protection measures [230]. To address privacy concerns while maintaining functionality, institutions should consider implementing open-source LLMs that allow for local deployment and greater data control [70], [77], [160].

Conclusion and Future Directions

Generative AI can transform educational feedback by enabling immediate feedback that is aligned with pedagogical principles, and supporting entirely new formats such as feedback within simulations and interactive feedback. Future work should focus on advancing feedback quality through refined approaches such as retrieval-augmented generation and multi-step systems, with consistent expert validation as an essential component. Research should also emphasize AI’s potential for creating innovative feedback delivery methods that incorporate simulation and real-time interaction.

Table

Warning

Note on data verification: Not all papers in this table have been double-checked by a human expert. Papers highlighted (marked with a reference number in the Ref column) have been cited in the synthesis and verified. Please consult original publications and apply your own critical judgment.

Download CSV

Synthesis

Applications

Generative AI is being widely adopted across disciplines to substitute, augment, modify and redefine various facets of assessment design in education. At the substitution level, it is primarily used for generating assessment items such as multiple-choice questions (MCQs), exercises, and practice materials. Augmentation involves enhancing assessments through alignment with learning objectives and integration of pedagogical frameworks. Modification changes how assessments are administered, through automatic question generation and adaptive or personalized assessments. Redefinition changes the nature of assessments more deeply and encompasses innovative approaches like interactive assessments and simulations.

Substitution: Creating Assessment Items & Exercises

Creating Multiple-Choice Questions (MCQs) represents the most prevalent application. Studies span multiple domains, but the medical domain is particularly well represented. Studies have documented the use of generative AI to produce medical questions 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19 and clinical cases 20, 21, 22; computer science questions 23, 24, 25, 26, 17, 27, 28 and exercises 29 30, 31, 32; language questions 33, 34 and exercises 35; mathematics and physics questions 36, 37, 38, 39; business and law questions 40, 41 and exercises 42; and K-12 curriculum-based questions 43, 44, 45.

Augmentation: Enhanced Assessment Design

Generative AI can be used to augment traditional assessment design processes, supporting question refinement 46, 16, 12, and alignment with learning objectives 26, 46, 47, 48, 49, or integration of pedagogical design frameworks such as Bloom’s taxonomy and Mayer’s Multimedia Theory.

Bloom’s Taxonomy serves as a common design framework to align AI-generated questions with appropriate cognitive levels. It is used both to guide question generation 58, 59, 60, 25, 41, 26, 46, 37, 30, 43, 61, 62, 39 and to evaluate cognitive demand of AI-generated questions 2, 40, 3, 8, 34, 63, 6. A critical finding emerges that AI tends to default to lower-order cognitive levels unless carefully engineered through specific prompting strategies 3.

Mayer’s Multimedia Theory offers another lens for augmentation by integrating multimodal elements into assessments 64. Even when not explicitly referenced, many studies implicitly apply Mayer’s principles by incorporating images, diagrams, and other visual representations to support assessments, which aligns with Mayer’s principles of dual coding 65, 66, 38.

Modification: Automatic & Adaptive Assessments

Generative AI also enables more profound modifications in assessment by automating question generation, allowing for on-demand delivery in contexts such as online learning and interactive quiz platforms 50, 51, 28. It also supports personalized assessment content 35, 30 and adaptive question selection 52.

Redefinition: Interactive Assessments & Simulations

Finally, GenAI is redefining assessment design through innovative approaches that go beyond traditional formats, by creating interactive assessments in the medical field, such as oral examination dialogues 53, 54 or patient simulations 55. Beyond the medical field, GenAI is also being used to create interactive assessments in other domains, such as intercultural communication 56, where it simulates communication tasks, and in cybersecurity, where it has been used to simulate crisis scenarios 57.

Measures & Outcomes

Several studies evaluated AI-generated items using measures such as difficulty indices, reliability coefficients, content validity, discrimination parameters, and alignment with learning objectives. 1, 67, 2, 25, 68, 3, 4, 8, 17, 19, 41, 6, 7, 9, 69, 70, 34, 10, 39, 20, 61, 26, 46, 59.

Across studies, there were mixed findings for the psychometric quality of AI-generated MCQs. There were reports that these items tended to be easier than those written by human experts, often reflecting lower cognitive demand and, in some cases, factual inaccuracies and validity concerns 40, 3, 30, 13, 70, 34, 1, 14. Despite this, some studies argued that discrimination indices show promise, with several studies demonstrating that AI-generated items can effectively differentiate between high and low-performing students 1, 3, 6, 10, 67.

Limitations & Recommendations

The primary endeavor is the production of high-quality AI-generated educational content. Output quality is highly dependent on phrasing and model version, necessitating clear, detailed, structured prompts 71, 41, 14. Expert validation procedures are critical to ensure content quality 4, 17, 1, 58, 3, 73. Advanced implementations employ retrieval-augmented generation to improve quality and domain-specific pertinence 72, 17, 58, 46, or multi-agent frameworks that simulate collaborative expert processes, incorporating generation of items, checking of items (e.g., plausibility of distractors), and refining of items 43, 74, 59.

Conclusion and Future Directions

Generative AI can transform assessment design by enabling rapid and automatic item generation, aligning assessments with learning objectives and pedagogical principles, introducing adaptive delivery and personalization, and supporting entirely new formats such as simulations and interactive tasks. Future work should focus on advancing the quality of AI-generated assessments, for example through refined prompting strategies, retrieval-augmented generation, multi-agent systems, and the consistent use of expert validation as an essential step in assessment development. Research should also expand beyond the current emphasis on multiple-choice questions to explore AI’s potential for creating innovative assessments that incorporate multimodality and interactivity, thereby fully realizing its educational value in assessment design.

Table

Warning

Note on data verification: Not all papers in this table have been double-checked by a human expert. Papers highlighted (marked with a reference number in the Ref column) have been cited in the synthesis and verified. Please consult original publications and apply your own critical judgment.

Download CSV

Synthesis

Applications

Generative AI is being widely explored as a foundational element for educational tutoring systems, leveraging its conversational abilities through applications spanning from basic question-answering to immersive simulation-based learning. At the substitution level, the most prevalent application involves deploying LLM-based chatbots as conversational tutors that answer student questions and provide explanations across domains, often enhanced with retrieval-augmented generation (RAG) to ground responses in course materials. Augmentation emerges when these tutors implement explicit pedagogical frameworks, most notably Socratic tutoring methods that guide students through questioning rather than providing direct answers. Modification is achieved through intelligent tutoring systems with concrete student modeling and adaptive mechanisms that systematically adjust to individual learner needs. Redefinition encompasses the creation of entirely new learning experiences, including simulations, roleplay scenarios, and immersive environments, thereby fully leveraging LLMs’ conversational capabilities.

Substitution: Basic Q&A and Content Delivery

The largest category of implementations involves deploying generative AI as conversational tutors that answer student questions, provide explanations, and deliver course content. These systems function primarily as intelligent information sources, supplementing human tutors by taking on a basic question-answering role. Several studies have reviewed educational applications and design considerations of such chatbots [35], [102], [111], [6], [10], [61], [23].

Many such applications have been developed in computer science education [3], [8], [28], [48], [51], [52], [57], [70], [72], [78], [79], [95], [99], [134], [180], as well as other STEM fields [41], [47], [137], [146], [167], [58], and other domains more broadly [8], [76], [93], [177],[13], [119], [92], [152].

In some cases, studies investigate how students use ChatGPT, and other commercially available tools, as tutors for problem-solving and concept explanation [91], [107], [130], [139], [151], [65].

Augmentation: Pedagogical Framework Implementation

Some applications move beyond basic question-answering by embedding specific pedagogical strategies into the AI tutor’s interaction design. Essentially, tutors should stimulate reflection rather than direct answers [115], [31], [132], [16], [96].

This is evident in Socratic tutoring, where AI systems are designed to guide students through questioning, prompting reflection, and encouraging self-explanation rather than providing direct answers [32], [64], [121], [75], [181], [116], [138], [62], [14].

Some other studies augment their tutors with pedagogical best practices, pedagogical frameworks, and scaffolding according to learning theories [29], [128], [136], [142], [145].

Modification: Adaptive Intelligent Tutoring Systems

A technically sophisticated set of implementations features explicit modeling of students and learning trajectories, and systematic adaptation mechanisms that adjust tutoring to individual learner states. Some studies have reviewed these applications [112], [21], as well as design considerations [176].

Practical implementations include knowledge tracing and learning path scheduling [14], [73], [123], [124], [145], [27], [135], [129], [85], [87], [45], [56], or even engagement tracking through neurophysiological signals [131].

Redefinition: Roleplaying & Simulated Learning Environments

A small number of studies fully leverage GenAI conversational abilities, to enable new forms of learning experiences that were previously impossible or impractical without LLM capabilities. For example, students have the opportunity to teach an LLM-based agent as a learning strategy [153], or to observe multi-agent debates that argue different positions regarding a same topic, favoring critical thinking skills [74]. Finally, there are also immersive simulations [55], [113], [158], that can be enhanced with augmented [39] or virtual reality [157].

Measures & Outcomes

Perception & Engagement

Student perceptions are generally positive [85]; they appreciate the personalized assistance, round-the-clock accessibility and non-judgmental environment that GenAI tutors provide, allowing them to practice and ask questions freely without fear of social evaluation [26], [49], [152], [119], [167], [121].

Moreover, reduced waiting time and frustration, as well as a supportive or empathetic chatbot tone help manage students’ negative emotions [75], [109], [86]. However, emotional support gaps in chatbot behavior can also lead to stress responses in some cases[4].

Despite an overall positive appreciation of GenAI tutors, students also found that these chatbots were lacking in some areas [152], such as reliability and accuracy of AI responses [49], [167], and pedadogical gaps, entailing a view of GenAI tutors as a supplementary tool rather than a replacement for traditional instruction [121], [49], [167].

How students percieve GenAI tutors and how they feel during chatbot interaction impacts engagement with these tools [114]. Several studies report levels of student engagement that are high [3], [85], [92], [167], and increased compared to baseline [119], [4], [75]. One study even found that students were more engaged and motivated with GenAI tutors compared to in-class active learning [29]. Another study used EEG methods and to objectively show the increase of engagement when interacting with a GenAI tutor compared to a non-adaptive chatbot [131]. Finally, gamification elements (i.e., progress visualization, loss aversion) can further increase engagement and persistence [4].

Educational Outcomes

A central concern is whether GenAI tutors meaningfully improve learning outcomes. The evidence is mixed: some studies find no notable benefit [121], [131], while others report gains relative to baseline or control conditions [137], [78], [92], or show improvements over traditional instruction [29], [4], [101]. These positive findings are tempered, however, by weak or absent long-term retention data, with available data not necessarily supportive of significant long-term retention [4].

Beyond whether AI tutoring helps, how students use it appears to matter considerably. Moderate, strategically guided engagement, such as Socratic dialogue, tends to align with better outcomes, whereas over-reliance risks undermining independent learning [31], [152], [64] and, in some cases, produces significantly worse results than studying without AI assistance [96].

Self-Regulation and Prompting

The interactive nature of GenAI tutoring means that learning outcomes depend on students’ ability to formulate effective queries and engage productively in dialogue. This dynamic can cut both ways. On one hand, the ease of obtaining direct answers can encourage passive reliance, reducing students’ independent problem-solving effort and eroding critical thinking [96]. Broader evidence confirms a gap between how students actually prompt GenAI tutors and what pedagogical goals would require: students tend toward passive, instructivist interactions rather than active, constructivist, or metacognitive ones, pointing to a clear need to steer students toward more purposeful prompting that fosters self-regulation rather than dependency [58], [88], [152], [32].

Research on what effective educational prompting looks like has begun to identify key components and practices, to support both the quality of AI outputs and the development of students’ cognitive and metacognitive skills [9]. Implementation choices matter here as well: how AI tutors are designed and deployed can actively scaffold metacognitive processes and self-regulation [161], [126], for example helping students to pay selective attention to their weak areas, analyze their errors deliberately, adapt their study sessions, and formulate better learning objectives [4]. Making prompting requirements explicit, for example by asking students to provide context, structure, and specificity, encourages goal-setting and self-monitoring behavior. Gradually, these self-regulatory routines become embedded into prompting behavior itself and strengthen students’ sense of self-efficacy [75].

Limitations and Recommendations

Enhancing AI Tutor Behavior

A fundamental challenge is ensuring that GenAI tutors produce accurate, reliable responses [130], [69] while also meeting specific pedagogical needs [125]. Several approaches have been developed to address both concerns.

Retrieval-augmented generation (RAG) has been widely adopted to ground responses in verified course materials, reducing the risk of factually unsupported or irrelevant outputs [3], [51], [56], [57], [117], [124], [134], [45].

Fine-tuning on domain-specific educational data offers another path to improved accuracy and pedagogical alignment [123], [145], [128], [118]. For instance, training on student learning interaction data, either real [127] or synthetic [143], [144], can produce tutors that guide reasoning rather than simply deliver answers [[127]]. A important caveat here is the student data paradox: models fine-tuned on learner-generated data may absorb misconceptions, leading to degraded performance [148].

Multi-agent architectures have emerged as a promising approach for enhancing tutor quality, with specialized agents handling distinct aspects of the tutoring process. This leverages the complementary strengths of different models for different tasks, such as content generation, pedagogical strategy selection, response validation, and different learning tasks [73], [123], [129], [43], [87], [146], [45].

Human-AI collaboration encompasses a range of practices: educator-driven authoring tools that allow teachers to configure AI behavior without programming expertise [58], [174], and human-in-the-loop validation of content and pedagogical reasoning [45], [47]. More broadly, studies advocate for hybrid approaches in which AI tutors complement rather than replace human instructors. In this model, AI handles routine questions and practice tasks while teachers maintain oversight of AI-generated content and attend to higher-order concerns and students’ socio-emotional needs [4], [23], [133].

Ethical and Privacy Considerations

Ethical concerns are raised across the literature, including issues of data privacy when student interactions are logged by commercial AI services, academic integrity risks when students use AI to bypass rather than support learning, and equity concerns regarding differential access to AI tools [38], [23], [12], [96], [25], [120].

Conclusion and Future Directions

Generative AI is being deployed across education as conversational tutors, with applications ranging from basic question-answering to adaptive intelligent tutoring systems and immersive learning simulations. The largest body of evidence concerns substitution-level implementations, such as chatbots that answer student questions. Results seem generally positive for engagement and satisfaction, though long-term learning outcomes are unclear and depend critically on pedagogically aligned usage.

Table

Warning

Note on data verification: Not all papers in this table have been double-checked by a human expert. Papers highlighted (marked with a reference number in the Ref column) have been cited in the synthesis and verified. Please consult original publications and apply your own critical judgment.

Download CSV

Synthesis

Generative AI-for-Learning Literacy in Students

While learning literacy and GenAI literacy have mostly been treated as separate topics in educational research, GenAI fundamentally changes how students engage with learning. This calls for either reconceptualizing learning literacy in light of GenAI technologies or expanding GenAI literacy to encompass learning-specific competencies. Given that effective GenAI use for learning requires both foundational GenAI literacy skills (e.g., critical evaluation, prompting, responsible use) and knowledge of how learning works, including how GenAI can positively or negatively impact learning and how it can be used strategically to enhance rather than offload learning processes, we have conducted a literature mapping that situates GenAI-for-learning as a specific competency within the broader domain of AI literacy.

This competency is increasingly important for educational outcomes [63], [152], particularly as institutions deploy educational chatbots that risk disproportionately benefiting students with higher GenAI literacy [199]. Students using these tools must be guided to actively direct their learning rather than passively relying on AI [53], [64], [168]. Fostering GenAI literacy, especially as applied to learning activities, across all students is therefore essential for educational equity [48].

Expanding AI Literacy in Education

Beyond education, scholars have proposed various frameworks defining AI literacy competencies [Pinski and Benlian 2024], [Almatrafi 2024], [Annapureddy 2025]. These frameworks provide the building blocks for AI literacy defined in the context of education, proposed by several prominent actors including UNESCO.

[UNESCO’s framework] organizes competencies across four dimensions: Human-centered mindset, Ethics of AI, AI techniques and applications, and AI system design, with three progression levels: Understand, Apply, and Create. [The Digital Education Council] uses five dimensions: Understanding AI and Data; Critical Thinking and Judgment; Ethical and Responsible Use; Human-Centricity, Emotional Intelligence, and Creativity; and Domain Expertise. This final layer pertains to specialized AI competencies relevant to specific fields. [The AI Literacy in Teaching and Learning (ALTL) framework] structures student competencies around Technical Understanding, Evaluative Skills, Practical Application, and Ethical Considerations.

Recent scholarship has further advanced framework development. The ABCE framework [200] proposes Affective (e.g., motivation, self-efficacy), Behavioral (e.g., commitment, collaboration), Cognitive (e.g., basic understanding, problem-solving), and Ethical (responsible use) dimensions. Another four-dimensional model [201] integrates cognitive, metacognitive, affective, and social dimensions. The dynamic AI literacy model (DAILM) extends foundational (universal) AI literacy with domain-specific components [112]. Other frameworks propose additional distinctions: one model [202] applies the ABC (Attitude, Behavior, Cognition) framework to Generic AI literacy (universal understanding), Domain-specific AI literacy (professional application), and AI ethics literacy (ethical awareness). Another [203] distinguishes AI knowledge from AI competency, applying this distinction across five interconnected components: Technology, Societal Impact, Ethics, Collaboration with AI, and Self-reflection.

Compared to AI literacy, GenAI literacy places greater emphasis on integrating critical and humanistic perspectives with technical and instrumental ones [6], while also reflecting an “AI as tools” perspective that positions prompt engineering as a core skill [11]. GenAI literacy frameworks have accordingly expanded upon AI literacy frameworks by adding competencies for prompt writing and managing ongoing dialogue [47]. The 3wAI Framework structures GenAI literacy around three dimensions: Know What (theoretical understanding), Know How (practical skills), and Know Why (critical perspective) [188]. Domain-specific adaptations have also emerged; in engineering education, competencies extend beyond foundational GenAI literacy to include specialized technical proficiency and domain-specific problem-solving [2].

Competency Domains in the Literature

Fundamental knowledge encompasses understanding how GenAI systems work, their technical foundations, capabilities, and limitations [3], [55], [59], [74], [90], [100], [158], [165], [188], [233].

Effective application involves the skillful incorporation of AI tools to solve various tasks [8], [16], [23], [34], [47], [74], [137], [138], [158], [162], [165].

Prompting skills refer to crafting effective prompts, formulating queries, and interacting strategically with AI systems [13], [14], [27], [28], [37], [72], [100], [142].

Critical thinking encompass critically evaluating, assessing, and validating AI-generated content for accuracy, bias, and quality [6], [8], [16], [23], [25], [34], [36], [40], [41], [47], [57], [74], [87], [93], [100], [131], [175], [183], [242].

Ethical awareness emphasizes responsible use, data privacy, and understanding societal implications [16], [23], [24], [47], [59], [90], [92], [112], [156].

Academic integrity involves maintaining authenticity of work and appropriate attribution when using AI [24], [90], [101], [126], [140], [147].

Metacognitive skills involve awareness and regulation of one’s own thinking when using AI [23], [25], [40], [81], [169].

AI-for-learning literacy specifically emphasizes understanding how learning works and using GenAI strategically to enhance, support, and deepen learning rather than offloading or replacing authentic learning [1], [15], [17], [44], [53], [58], [100], [114], [147], [164], [168].

Domain-specific competencies apply to particular fields such as clinical domains [33], writing and language learning [34], [204], [205], or coding [206].

Measuring AI Literacy

Several studies have designed survey items measuring GenAI familiarity, understanding, confidence, perceptions, and readiness to use GenAI tools in general [3], [16], [36], [64], [93], [156], [208], [209], [210], [211], [212], and as a learning tool specifically [213]. Some studies use performance tasks or observation settings requiring demonstrated skills in specific GenAI tasks [214].

Validated Instruments

Several validated instruments to measure AI literacy have been developed and used across studies, including self-report questionnaires and performance-based tests.

Self-report questionnaires include:

• A 12-item self-report instrument [Wang 2023] adapted in several studies to measure generative AI literacy [216], [217], [219].

• A self-report questionnaire assessing students’ understanding and capabilities related to AI, to measure GenAI literacy, adapted from Long & Magerko 2020 [152]

• A 32-item self-reported AI Literacy Questionnaire (AILQ) developed and validated to measure students’ literacy development across four dimensions (affective, behavioral, cognitive, ethical) [220], [221].

• A self-reported Meta AI Literacy Scale (MAILS) [222] used across multiple studies [158], [223], [224].

• For GAI-for-learning literacy specifically, the validated GenAI-LLs instrument consists of four factors and 29 items [229], [230].

Performance-based assessments include:

• SAIL4ALL-12 scale [225], [28], [226]

• Generative AI Literacy Assessment Test (GLAT) [48], [199], [227]

• A set of five different validated performance-based instruments: AI Concepts for Problem-solving test, Metacognitive Strategies survey, Empowerment survey, AI Ethics survey, and Literacy in Problem-solving survey [201].

• An 18-item performance-based GenAI literacy assessment [228]

Interventions to Foster AI Literacy

Instructional Approaches

[1] propose to adapt instruction across three levels: core knowledge (lectures, foundational texts, structured practice, guided discussions); leveraging for learning and productivity (performance tasks, workshops, building proficiency); and expansion (mentoring, resources, learning communities). [165] propose a Universal Module for AI Literacy that can be woven into existing curricula through scaffolded, discipline-contextualized learning activities.

Implementation strategies for GenAI literacy have been documented across multiple sources. Literature reviews have identified approaches for teaching specific components such as prompting, tool use, and ethics [11]. Research has explored both instructor strategies for integrating GenAI into teaching practices and student preferences for acquiring AI literacy [93], [154], while pedagogical frameworks have provided concrete implementation examples [23]. For engineering education specifically, frameworks address AI both as a subject matter and engineering tool, and as a learning tool [2].

Targeted Interventions

Targeted interventions outside formal curricula can take various forms. Student-facing GenAI guides can provide basic support [231], while AI disclosure forms can serve as pedagogical tools that foster transparency, ethical engagement, and self-regulation [232]. Workshops and standalone lectures covering GenAI concepts, models, prompting strategies, and ethical considerations have demonstrated significant improvements in students’ self-perceived understanding [233], [3]. Specific competencies can also be targeted by specialized interventions. The “DeBiasMe” intervention enhances awareness of cognitive biases while empowering user agency in AI interactions [40]. Three-session clinics focused on prompt engineering have increased AI literacy while reducing technology anxiety [28]. Training can also be embedded within learning management systems through microcourses that cover GenAI mechanics, limitations, and efficient and ethical use [90].

Curricular Integration

Embedding GenAI literacy within regular course activities, while scaffolding it’s use, helps develop skills in prompting and critical evaluation, promotes active rather than passive AI use, and reduces inappropriate GenAI usage. Reflective journaling often accompanies this approach. Successful implementations span various domains such as programming and computer science [25], [206], [234], design thinking [92], economics [235], writing assignments [114], [205], [236], [237], public health [238], and rhetorical analysis [239].

Larger-scale curricular revisions have also been implemented. An 18-session program involving guided training, reflective tasks, and applied practice for GenAI in academic learning led to improvements in cognitive and behavioral dimensions, with students demonstrating enhanced control over learning processes, more strategic AI tool use, and increased academic self-regulation [53]. A 10-week undergraduate module included experimentation with GenAI models, practical sessions and discussions, leading to self-reported improvements in understanding and competence [171]. A 10-week EAP (English for Academic Purposes) module with integration of AI improved students’ confidence in AI tool use and their ability to critically evaluate AI-generated content [16].

K-12 Education

Several tools and approaches have been developed for exploring AI with K-12 students, promoting understanding of AI concepts [241], [55], [158], [74], [203]. Literature reviews have further identified classroom activities for developing key AI competencies in secondary education [47]. A 12-lesson AI literacy workshop over two months, with learning activities aligned with the AILQ instrument, achieved significant improvements in AI literacy [200]. A 14-hour AI literacy course for senior secondary students, targeting the competence of problem-solving through project-based learning integrating the use of AI, improved students’ ability to use AI concepts for real-life problems, metacognitive strategies, and understanding of ethical boundaries of AI use [240].

Teacher Education

A 5-week out-of-curriculum intervention for preservice teachers combined theoretical grounding with practical, problem-based learning activities focusing on various AI tools, effectively facilitating understanding and application in educational contexts [214]. Another program engaged 35 teachers in experiential activities developing AI literacy, alongside a collaborative assignment to co-author an open-access textbook, Teaching and Creating With Generative Artificial Intelligence. To support equitable and inclusive access to educational benefits offered by AI, the Student Artificial Intelligence Literacy (SAIL) framework was developed, facilitating student AI literacy through curriculum engagement and three types of interactions: cognitive, socio-emotional, and instructor-guided [138].

Gaps and Limitations

Several gaps emerge from this literature mapping:

1. Literature complexity: AI literacy is a complicated and multifaceted construct with numerous frameworks proposing different dimensions and components.

2. GenAI-for-learning: Only a few studies explicitly acknowledge this dimension. Whether it should be treated as a separate literacy or as part of general AI literacy remains unresolved. While it can be argued that it is a domain-specific competency (i.e., GenAI applied to learning), it is transversal skill that affects all disciplines.

3. Limited measurement instruments: The existing measurement instruments have undergone validation efforts, but have not yet been widely tested. Data on criterion validity are absent. Only one validated instrument (GenAI-LLs) specifically targets GAI-for-learning literacy, using self-report measures.

4. Few intervention studies: Few studies report interventions targeting GenAI-for-learning literacy, or even GenAI literacy more broadly.

5. Limited outcome data: Almost no data exist on the uptake and outcomes of literacy interventions, limiting our understanding of what works in practice.

Table

Warning

Note on data verification: Not all papers in this table have been double-checked by a human expert. Papers highlighted (marked with a reference number in the Ref column) have been cited in the synthesis and verified. Please consult original publications and apply your own critical judgment.

Download CSV

About the Catalog Creation

This catalog is built through a systematic, multi-stage process including data retrieval and processing, combining AI-powered analysis and essential human oversight. Our goal is to create a dynamic resource that is regularly updated.

Important

We do not evaluate the scientific quality, methodological rigor, or validity of the studies referenced in this catalog. Our role is to point to relevant literature and synthesize available information—not to certify its credibility. Users must consult the original publications and apply their own critical judgment when interpreting the findings.

Select a topic below to learn more about our methodology.

Our process begins by casting a wide net to capture all potentially relevant research.

• Exhaustive Search: For each emerging practice, we develop tailored search queries to retrieve references from three major academic databases: Web of Science, Scopus, and Semantic Scholar. This initial step typically yields thousands of references.

• Full-Text Filtering: We then narrow this pool to focus exclusively on open-access articles. This ensures we can analyze the complete full-text content, which is essential for in-depth analysis. This filtering step typically reduces the list to about one-third of the original references, for which we are able to collect an open-access full text.

Once we have a collection of full-text articles, we leverage Large Language Models (LLMs) to perform a detailed, multi-step analysis.

• Relevance Screening: An LLM reviews each article abstract to determine its relevance. Only studies where the emerging practice is a primary focus of the research article are selected for further analysis.

• Variable Extraction & Coding: These selected studies are then further analyzed by an LLM, by extracting key variables of interest, such as the description of the application, the pedagogical frameworks used, and the reported outcomes. This yields a comprehensive, filterable application table that can be explored in the catalog.

• Synthesis creation: Finally, we create a narrative synthesis based on the structured data in the application table, by combining AI categorization with human insights. The synthesis identifies and categorizes the main types of applications and outcomes and links to DOIs of relevant papers.

A note on our tools: We utilize the OpenAI API for these tasks. As per the data processing agreement, none of the data submitted is used for training their models, ensuring the confidentiality of the research analyzed.

Our workflow is a collaborative system where human expertise is crucial for ensuring the quality and accuracy of the final output. The human-in-the-loop has several critical tasks:

• Schema Definition: Defining and refining the coding schemas that instruct the AI on what information to extract and how to categorize it.
• Verification: Comparing AI-coded data against human-coded samples to validate accuracy and consistency.
• Synthesis: Co-creation of research insights with AI.

The research pipeline, from querying academic databases to extracting structured data, is designed for a high degree of automation. This enables rapid updates in a continuously evolving field while ensuring systematicity through consistent procedures for collecting, filtering, and analyzing research papers. Each step is traceable via documented parameters (e.g., research queries, analysis prompts) and preserved historical data, and the workflow is reusable across different contexts.

Phase	Description
Phase 1 – Retrieval	Automatically retrieves an up-to-date collection of references matching a research query from Web of Science, Scopus, and Semantic Scholar, then either creates a new dataset or updates an existing one.
Phase 2 – Filtering & Full Text	Automatically filters for relevant papers and retrieves their full texts from publishers or open-access repositories.
Phase 3 – LLM Analysis	Applies LLM-powered extraction to transform full texts into structured information.

🚨 Decision checkpoints between these phases allow for crucial human oversight and quality control.

This catalog is a living project, not a static report. Our methodology is in a state of continuous development. We view the entire workflow as an iterative process to discover the most effective ways to derive optimal results from LLMs.

We are constantly experimenting with and refining key aspects of the process, including:

• The formulation of our prompts to guide the AI.
• The design and granularity of the coding schemas.
• The methods for pre-processing and presenting full texts to the AI.
• The specific requirements and structure for the final synthesis.

This commitment to iterative improvement means that as we refine our techniques and as AI models evolve, the quality, depth, and reliability of this catalog will continuously increase over time.

---

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

How to Cite This Catalog

If you use this catalog in your research or work, please use the citation information below. Click on a style to expand it and copy the citation to your clipboard.

APA 7th ed.

Uittenhove, K. (2025). AI for Education: A Living Catalog of Emerging Practice (Version v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.15633017

MLA 9th ed.

Uittenhove, Kim. AI for Education: A Living Catalog of Emerging Practice. Version v1.0.0, Zenodo, 10 June 2025, doi:10.5281/zenodo.15633017.

Chicago 17th ed.

Uittenhove, Kim. 2025. "AI for Education: A Living Catalog of Emerging Practice." Zenodo. https://doi.org/10.5281/zenodo.15633017.

Harvard

Uittenhove, K. (2025) 'AI for Education: A Living Catalog of Emerging Practice', Zenodo. doi: 10.5281/zenodo.15633017.

Vancouver

1. Uittenhove K. AI for Education: A Living Catalog of Emerging Practice [Internet]. Zenodo; 2025 Jun 10. Available from: https://doi.org/10.5281/zenodo.15633017

IEEE

[1] K. Uittenhove, "AI for Education: A Living Catalog of Emerging Practice," Zenodo, Jun. 10, 2025. doi: 10.5281/zenodo.15633017.

BibTeX

@misc{uittenhove_2025,
  author       = {Uittenhove, Kim},
  title        = {{AI for Education: A Living Catalog of Emerging Practice}},
  month        = {June},
  year         = {2025},
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.15633017},
  url          = {[https://doi.org/10.5281/zenodo.15633017](https://doi.org/10.5281/zenodo.15633017)}
}