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 ISAR Framework

The application of generative AI in teaching and learning is characterized by new affordances (i.e., automatisation, adaptation, personalization, interactivity) as well as new risks (i.e., applied without oversight, GenAI can harm learning and teaching). This balance between new affordances and new risks can be understood through the ISAR model (Bauer et al., 2025).

Substitution AI replaces existing tasks or tools with functionally equivalent results.

›

Augmentation AI adds functional value through pedagogical alignment or adaptive support.

›

Redefinition AI enables entirely new educational experiences, previously unattainable, by leveraging interactivity and simulation capabilities.

⚠ Inversion — a risk at every level: uncritical adoption can undermine the very educational goals AI is meant to support.

How the Catalog is Built

The catalog is produced through a pipeline spanning systematic literature retrieval, LLM-based relevance screening, full-text annotation, analysis, and AI-assisted synthesis. Each stage is built around explicit checkpoints for human input: researchers define which practices to track, how papers should be read and annotated, what should be analyzed, and how the synthesis should be structured.

A distinctive feature is the annotation and analysis approach. The pipeline uses a researcher-defined taxonomy to annotate each excerpt with a section and rhetorical functions, which are subsequently mapped to the evidence requirements for the analysis fields. This makes evidence retrieval tractable and creates a direct avenue for continuous improvement, as domain experts can refine the annotation taxonomy and evidence requirements for analysis fields.

Click on any stage below to learn more.

→

→

→

Expert-Defined Practices & Human-in-the-Loop

• Practice Definition: Identifying and defining the key emerging practices in generative AI for education, grounded in research, theory, and real-world experience.
• Annotation taxonomy: Experts define how scientific papers should be annotated.
• Analysis fields: Experts define the specific analysis questions and evidence requirements for each field.
• Synthesis structure: Experts provide a synthesis structure template.
• Decision checkpoints: At key stages, a human reviews the outputs and can modify any part of the extraction process to get the desired results.

Evidence from the Field

• Open-Access Literature: The catalog is built exclusively on open-access scientific literature.
• Database Search: For each practice, tailored queries retrieve references from three major academic databases — Web of Science, Scopus, and Semantic Scholar.
• Relevance Screening: An LLM reviews each article's abstract to assess whether the envisioned practice is a primary focus of the study. Only relevant articles advance to the next stage.
• Full-Text Retrieval: For screened-in articles, the pipeline collects the full text from publishers or open-access repositories.

Annotation, Analysis & Synthesis

• Rhetorical annotation: Each article's full text is split into consecutive windows. With the full set of windows as context, the LLM classifies each window by structural section and rhetorical function (e.g., describes participants, reports results, acknowledges limitations, etc.) — a content-agnostic map of how the paper is organized.
• Field-level analysis: For each expert-specified analysis field, a defined set of rhetorical functions determines which windows are relevant candidate evidence. The LLM then answers that field's specific question using the matching windows, and reports exactly which excerpts it relied on — producing one structured, traceable answer per field per paper.
• ISAR classification: The interpretation of the artifact presented in each paper is categorized according to the ISAR model.
• Synthesis generation: The per-field answers are interpreted and structured according to an expert-provided synthesis template, wherein each point in the narrative is linked back to the corresponding paper. A human expert then evaluates and adapts the synthesized content and references before finalization.

Unlike standard retrieval-augmented generation (RAG), where relevance is determined by semantic similarity, this approach puts the expert in control: an expert-defined taxonomy explicitly determines how to annotate papers and the expert decides which rhetorical functions are relevant to which analysis questions, based on domain knowledge of how scientific papers are structured. Evidence retrieval becomes tractable, not dependent on the vagaries of embedding distance.

Automatic Updating Pipeline

The pipeline is designed for automation and traceability, enabling the catalog to stay current as the field evolves. The same queries and procedures are applied consistently across all papers and update cycles, and each processing step produces a timestamped dataset — making it possible to track what changed between runs, compare results across versions, and reproduce any output.

Stage	Description
Retrieval	Search queries are run against Web of Science, Scopus, and Semantic Scholar; results are merged into an up-to-date reference dataset.
Screening & Full Text	Abstracts are screened for relevance by an LLM; full texts are retrieved for screened-in articles.
Annotation	Each article is split into windows and classified by structural section and rhetorical function, producing a content-agnostic map of the paper usable across all annotation fields.
Analysis	For each analysis field, matching annotated windows are filtered by rhetorical function and the LLM answers the field's question, citing the specific excerpts it used as evidence.
Synthesis	Field-level answers are assembled into a narrative synthesis following the expert-provided template, and reviewed by a human expert.

🚨 Human decision checkpoints between stages allow for review before proceeding.

---

AI for Education

Explore each practice below — select a practice card to read its synthesis of key applications. The full set of underlying papers for each practice can be downloaded as a CSV at the top of its synthesis.

Select a practice

Data verification

Papers cited with a reference number in the synthesis below have been checked by a human expert. However, we do not evaluate the scientific quality of methodological rigor of studies. Please consult original publications and apply your own critical judgment.

Download underlying reference and analysis table

Applications for Assessment Design

Generative AI is reshaping assessment design along a continuum from scaling established forms of assessment to enabling more adaptive, integrated, and interactive formats. The literature can be organized through the ISAR model ((Bauer et al., 2025)): Substitution captures the use of AI to reproduce existing assessment-design functions; Augmentation describes applications that add capabilities such as pedagogical alignment, adaptation, or integration; and Redefinition encompasses assessment formats made possible by the interactive and generative capabilities of AI. Across all three levels, however, Inversion remains a risk when these affordances come at the expense of assessment quality, learning, or engagement. We first map these applications before examining their outcomes and associated inversion risks.

Substitution

At the substitution level, the principal affordance of generative AI is scaling the production of conventional assessment items. Across disciplines, generative models are used to rapidly produce questions that reproduce the form and intended function of human-authored items. This shifts the central question from whether AI can generate assessment content to whether the resulting items achieve sufficient quality to function as valid substitutes for human-authored material.

As summarized in Table 1, this is the most extensively represented application of generative AI in assessment design, with studies spanning medical and health sciences, computer science and engineering, language and literacy, mathematics and physics, and more general educational settings. The literature is particularly concentrated in medical and health education, where generative AI has been used extensively to produce structured question formats such as multiple-choice questions, as well as clinical cases and vignettes.

Table 1. Studies using generative AI for assessment item generation, by discipline.

Field	References
Medical and Health Sciences	Including structured question formats such as MCQs: [123], [165], [294], [499], [516], [437], [236], [351], [396], [101], [104], [184], [142], [400], [443], [327], [402], [382], [479], [477], [302], [383], [332], [381], [212], [232], [103], [436], [355], [196], [87], [168], [218], [262], [121], [216], [328], [352], [337], [322] Including clinical cases or vignettes: [270], [264], [111], [88], [154], [128], [82], [83], [296], [494], [169], [345], [408], [377], [389], [259], [498]
Computer Science and Engineering	[129], [197], [489], [242], [126], [164], [97], [98], [241], [179], [223]
Language and Literacy	[119], [162], [273], [217], [203], [125], [195], [472]
Mathematics and Physics	[233], [127], [240], [431]
Domain-General Education across Primary, Secondary, and Tertiary Settings	[99], [331], [493], [419], [140], [251]

Augmentation

At the augmentation level, generative AI moves beyond reproducing existing item-authoring processes to adding capabilities to how assessments are designed and delivered. Three mechanisms are particularly prominent in the literature: aligning generated items with pedagogical frameworks, adapting assessment to individual learners, and integrating generation into broader learning and assessment systems.

Alignment with cognitive frameworks Usage of framewoks like Bloom’s taxonomy can ensure systematic alignment with pedagogical intentions. For example, BloomLLM was fine-tuned on expert-authored questions labeled with Bloom levels, enabling the model to generate assessment items at specified cognitive depths [181]. TwinStar uses a dual-agent architecture in which a separate cognitive-level evaluation agent iteratively feeds back Bloom-alignment scores to a generator agent [274].

Adaptive delivery Item selection can be tailored to individual learner characteristics in real time. For example, AdaptiveGPT adjusts question difficulty in real time based on evaluated student responses and presents individualised learning trajectories to teachers [175]. Another study tested a mathematics exercise generator that aligned question difficulty with student cognitive profiles [404].

Integration Item generation can be integrated within learning management systems, intelligent tutoring systems, and automated assessment pipelines, affording more accessible and integrated learning opportunities. For example, the Learnix platform integrates GPT-4 to generate MCQs and provide real-time feedback on coding problems within a unified eLearning environment [157].

Table 2. Studies using augmentation mechanisms in AI-assisted assessment design.

Augmentation Mechanism	References
Pedagogical Alignment	[181], [274], [249], [333], [132], [144], [73], [178], [506], [126], [376], [245], [272], [100], [114], [72], [248], [177], [434], [81], [344], [146], [141], [93], [384], [295], [478], [189], [474], [194], [108], [106]
Adaptive Delivery	[175], [263], [126], [404], [440], [372], [300], [326], [344], [304], [120]
Integration	[157] [145], [330], [225], [453], [32], [201], [433]

Redefinition

At the redefinition level, generative AI is used to create dynamic and interactive assessment formats that are difficult to reproduce through conventional assessment design. Rather than generating a fixed set of items, these applications exploit the capacity of generative models to respond to learner actions in real time, enabling branching scenarios, simulations, adaptive dialogue, and assessments linked to dynamic or multimodal content.

Dynamic and interactive assessment formats Generative AI enables new assessment formats that fully leverage the interactivity and simulation afforded by generative AI. These assessments can involve elements like scenarios, branching based on the choices of the examinee, and interactive examination, where the examiner’s next question depends on the candidate’s prior answer. For example, in one study in medicine, decision points are inserted at clinically strategic moments so that the narrative branches in response to student choices [515]. Another study developed an application that generated dynamic clinical case vignettes complete with patient history, vital signs, laboratory values, and imaging findings, and then engaged in an interactive dialogue that guides the examinee through diagnosis, operative planning, and complication management, before delivering structured feedback [122]. In language education, the LEARN system pairs assessment items with a voice-driven adaptive dialogue manager that elicits and evaluates oral responses in multiple languages, creating a multimodal assessment format virtually impossible with pre-AI technology [467]. Another approach links assessment to dynamic content: one driving simulator combines 360-degree spherical video with a visual language model to generate situational test questions anchored to specific moments in the video stream [364].

Table 3. Studies employing redefinition-level AI assessment formats.

Innovation	References
Interactive and Simulation	[122], [265], [280], [380], [467], [133] [515], [364], [90]

Measures and Outcomes

The educational value of AI-assisted assessment design ultimately depends not on the capacity to generate assessment content, but on the quality of the resulting assessments and their consequences for students. The literature addresses these questions at two distinct levels. Most studies evaluate the quality of AI-generated assessment items themselves, using expert judgments, psychometric analysis, or comparisons of student performance on AI- and human-authored items. A much smaller evidence base examines downstream effects on learning and engagement.

Measuring the quality of AI-generated assessments

Evidence on the quality of AI-generated assessments is mixed. As summarized in Table 4, expert evaluation is the most common approach, while fewer studies examine psychometric properties or directly compare student performance on AI- and human-authored assessments. Across these methods, some studies provide favorable evidence that AI-generated items can approach or occasionally exceed the quality of human-authored items, whereas others identify substantial quality deficits.

Favorable evidence for the quality of AI-generated items Several studies report comparable or favorable performance of AI-generated assessment items. One study found that expert median ratings for AI-generated items across five quality domains were statistically identical to human-authored items [443], whereas another study even found that ChatGPT-generated items received higher expert ratings than faculty-prepared items on most quality parameters including clinical appropriateness, distractor reasonableness, and appropriate difficulty [106]. Psychometric studies also frequently reported encouraging results. For example, one study found that AI-generated MCQs presented higher internal reliability, a better-balanced difficulty distribution, and an improved discrimination index compared with a human-generated exam, [100]. However, these results were obtained after human preselection of suitable AI-generated exam items, which is the case for many psychometric studies reporting good results.

Student performance on AI- and human-authored items Studies provide mixed evidence. Four studies found no significant difference in student scores; for example, one study reported no significant difference in exam scores between an AI-generated and faculty-authored assessment, with a strong positive correlation between the two [216]. Four other studies, however, identified performance differences. One found that AI-generated versions yielded higher mean grades than classic exams but showed poor difficulty distribution, limited discriminatory power, and inadvertent cues [295]. Another found significantly lower mean midterm scores and greater score variability for AI-generated MCQs, which the authors interpreted as evidence of reduced grade inflation and improved student differentiation [100].

Table 4. Studies evaluating the quality of AI-generated assessment items, by method.

Method	References
Psychometric Evaluation	[100], [272], [196], [154], [477], [245], Inversion risks: [383], [328], [82], [317], [216], [264], [337], [189], [81], [262], [474], [295], [126],
Quantitative expert ratings	[443], [106], [142], [479], [169], [259], [249], [178], [389], [195], [333], [103], [402], [345], [273], [217], [132], [436], [351], Inversion risks: [377], [332], [396], [382], [302] [146] [81], [218], [119], [493], [262], [88], [474], [295], [352], [232], [141], [376], [127], [241], [73], [121], [179], [111], [203], [101], [408],
Student Performance	No difference in performance: [216], [236], [477], [196] Difference in performance: [295], [100], [337], [264]

Inversion risks in assessment quality

Alongside evidence of adequacy, a substantial body of research identifies systematic deficits that can undermine the validity and educational value of AI-generated assessments. These represent inversion risks when the efficiencies or new capabilities afforded by generative AI are achieved at the expense of assessment quality. As summarized in Table 5, the most consistently documented problems concern cognitive demand and difficulty, distractor quality, factual and logical accuracy, and psychometric performance.

Low cognitive demand and difficulty This is the most consistently reported deficit. AI-generated items disproportionately target lower-order cognitive skills such as recall and tend to be easier than human-authored items. One study found that 95% of AI-generated questions targeted knowledge recall and only 5% higher-order understanding or analysis [302]. Even when explicitly asked to generate items at specific levels, Apply and Conceptual Knowledge levels showed higher generation precision, suggesting that AI generation is more reliable at these levels [146]. A related but distinct problem concerns difficulty calibration: when AI systems are asked to estimate item difficulty, their ratings show systematic bias compared with empirically derived parameters [300].

Low distractor quality This is another pervasive consetn For example, one study found that ChatGPT items had 33.75% nonfunctional distractors versus 13.75% for faculty items, with lower mean distractor effectiveness [262]. Another found that human-authored distractors received substantially higher plausibility ratings than LLM-generated alternatives [127].

Factual inaccuracies, hallucinations, and logical inconsistencies Important limitations were identified in many studies. One study found that 22% of AI-generated items contained factual or conceptual errors, with three items removed mid-semester after students identified errors [248]. A study of mathematics assessment similarly identified calculation errors, incorrect roots, and incorrect answer keys [240]. Some studies also report difficulty producing usable image-based items [437], limiting applications for multimodal assessment.

Psychometric underperformance. Issues with psychometric performance have been identified in several studies. One study found significantly lower discrimination between high- and low-performing students for AI-generated items, with 56% failing acceptable discrimination thresholds compared with 26% of human-authored items, including three AI items with negative discrimination [328].

Table 5. Studies documenting quality deficits in AI-generated assessment items, by deficit type.

Quality Deficit	References
Low Cognitive Demand and Difficulty	[302] [146] [300] [81], [248], [218], [119], [493], [472], [262], [88], [194], [251], [264], [474], [352], [232], [295], [126], [141], [337], [102], [165], [376].
Low Distractor Quality	[262], [127], [328], [241], [381], [376], [73], [121], [382], [352], [179], [111], [396], [99], [93], [218], [223], [203], [162], [128], [472], [264], [101], [327], [383], [332]
Factual Inaccuracies, Hallucinations and Logical Inconsistencies	[248], [240], [408], [380], [376], [87], [122], [126], [218], [230], [102], [233], [145], [88]; [128], [168], [431], Including inability to produce usable image-based items: [437], [382], [162]
Psychometric Underperformance	[328] [82], [317], [216], [262], [264], [337], [189]

Impact on Learning and Engagement

Item quality, however, provides only an indirect indication of educational value. A more consequential question is whether the use of AI-generated assessments changes student learning or engagement. Despite the large literature evaluating generated items, relatively few studies examine these downstream outcomes directly, and the available evidence remains mixed.

Learning performance A randomized comparison in three university computing courses found that students who engaged with AI-generated assessments achieved significantly higher final exam scores than those who did not, with positive correlations between the number of assessment attempts and exam performance [98]. A study using LLM-generated personalized assessments reported a significantly higher post-test mean than controls, with particularly pronounced gains among initially lower-performing students [372]. A driving-simulator study similarly reported a 27% improvement in test accuracy after LLM-assisted simulator-based instruction compared with traditional video instruction [364].

Engagement Evidence in this category is heterogeneous. The driving-simulator study reported a 32% increase in self-reported engagement and confidence [364], and a language-learning platform study reported high engagement and enjoyment in students [467]. Other studies, however, identify potential inversion effects. Engagement with AI-generated assessments was more superficial in one study, with students typically completing them only once compared with repeated completion of instructor-created content [330]. In AI-generated speaking assessments, emotional, cognitive, and social engagement were significantly lower with AI than with human interlocutors [133]. These findings suggest that even where AI-generated assessments are technically adequate, their educational effects may depend on how students interact with them and on the assessment context.

Table 6. Studies measuring student impact of AI-generated assessment.

Student Impact	References
Learning	[98], [372], [364], [114], [82], [164], [429], [144]
Engagement	Positive effects: [364], [467], [294], [408], [280], [372] Negative effects: [330], [133]
Student Perception	[280], [327], [265]

Recommendations to Improve AI-Enhanced Assessment Design

The literature increasingly suggests that the quality of AI-assisted assessment design depends on how generation is constrained, supported, and embedded within a broader assessment-design process. Recommendations therefore converge around three complementary targets: improving the quality of generated items, strengthening the capabilities of the assessment-design pipeline, and defining effective human–AI workflows.

Improve item quality Several strategies allow to improve the quality of AI-generated assessment items, as well as their relevance in relation to course material, alignment with course objectives, while also reducing hallucinations.

Prompt engineering is a consistently reported lever to improve performance. Model outputs vary substantially with prompt wording [121], and novice users who lack prompt-engineering training tend to obtain suboptimal results [101]. Therefore authors across domains treat this as the primary means of improving item quality. Structured prompting and domain-specific prompt templates (e.g., specifying subject, topic, learner level, Bloom’s taxonomy level, distractor constraints) substantially raise the relevance and cognitive alignment of generated items, and reduce common errors such as implausible distractors [97], [104]. Misconception-oriented prompting (i.e., explicitly instructing the model to generate distractors that reflect common student errors) has also been shown to increase the proportion of high-quality distractors in mathematics and computing contexts [242], [241]. Few-shot exemplars are widely recommended as a practical complement to structured prompting. Providing sample questions alongside the prompt biases the model toward desired formats and quality levels [219]. Chain-of-thought (CoT) decomposition (i.e., prompting the model to reason step-by-step through item construction) has furthermore been shown to improve generation quality [127]. Authors also caution that prompt quality is itself a skill requiring professional development: training teachers to craft clear, detailed, goal-oriented prompts is a prerequisite for reliable outputs [86].

Retrieval-augmented generation (RAG) is recommended as a mechanism for reducing hallucination, ensuring curriculum alignment, and incorporating up-to-date knowledge, by enabling the LLM to retrieve relevant passages from course materials or curated knowledge at inference time. Implementations can make use of vector databases [423] or knowledge graphs [110]. Authors recommend careful knowledge-base governance: maintaining master documents in plain text, applying version control, scheduling faculty audits to flag obsolete or missing content, and ensuring source traceability so that every generated item can be traced to an auditable passage in the source material [419], [435]. Creating additional metadata (e.g., knowledge components, key concepts, learning objectives, cognitive complexity ratings, etc.) can further enhance the LLM’s ability to generate appropriate assessment items. For example, one study used syllabus-specific vector databases storing Bloom’s taxonomy classifications and course outcomes to constrain generation [333].

Improve capability of assessment design pipeline

Choice of model is one important factor influencing the performance of generative AI in assessment design (see Table 7). Results obtained with one model snapshot may not replicate with a later version or a different model. The conditions under which adequate quality is most reliably achieved include the use of more capable models (GPT-4 and its successors consistently outperform earlier or smaller models). However, using more capable models is also less sustainable, especially if smaller models can achieve adequate performance.

Multiple agent architecture is recommended by a growing body of work, and consists of decomposing the assessment-generation pipeline into specialised components rather than relying on a single monolithic model call. Different agents could handle tasks related to item generation, pedagogical review, and student error modelling. For example, a dual-agent architecture separating a question-generation model from a cognitive-level evaluation model has been shown to improve Bloom’s-level adherence [274]. More elaborate pipelines assign even more specialised roles (e.g., item generator, distractor critic, difficulty calibrator, and hallucination checker) and use optimisation methods to identify the most cost-effective agent sequences [452]. Indeed, authors consistently note that multi-agent pipelines increase quality but also computational cost, and that minimising agent calls while preserving quality gains is an important practical constraint.

Fine-tuning represent a final and more technically difficult avenue. Fine-tuning allows to repurpose models to develop more specific capabilities, and to perform these capabilities at a lower computational cost. For example, TwinStar uses two lightweight fine-tuned models and attains equal or better question quality while using far fewer parameters and negligible memory compared with large LLMs, making it less costly, more efficiently deployed, and more environmentally friendly [274].

Design Human-AI Workflow

Human-in-the-loop is required for summative and high-stakes examinations, while for low-stakes formative quizzes, and initial item-bank seeding, it may be argued that automated generation and validation of assessment items is adequate. Such a workflow typically involves three checkpoints [477]. First, prompt engineering aligns generation with pedagogical objectives. Second, expert review of generated items ensures that the generated items indeed match pedagogical objectives and contain factually correct information and high-quality distractors. Authors recommend using calibrated rubrics, inter-rater reliability checks, and consensus among multiple reviewers to reduce subjectivity [477]. Third, psychometric validation, including difficulty index, discrimination index, item-total correlation, and IRT analyses, is recommended before deployment; authors caution against circular validation in which the same LLM that generates items also evaluates them [295], [300]. Expert validation and psychometric evaluation should inform an iterative refinement pipeline before items are deployed [154], [214]. Beyond merely integrating human expertise in the process, generative AI applications can also be envisioned as a co-design partner to help humans analyze and improve assessment design [92]. For example, one study designed a generative AI application with a structured workflow inspired by an instructional design framework, operating in several modes, ranging from autonomous to guided to full co-pilot, enabling flexible control over the degree of human involvement [424].

Responsible implementation principles require attention to equity and access, accountability, and the development of AI and assessment literacy across all stakeholders (see [512] for a procedural guidance taxonomy-of validity, reliability, safety, security, accountability, transparency, explainability, privacy, and fairness—for organizations adopting generative AI in assessment contexts). Concerning equity and access, authors recommend performing regular bias audits of generated content, and designing prompts that explicitly incorporate equity considerations [133]. Furthermore, equitable access to AI tools needs to be ensured [488]. Transparency mechanisms are recommended with instructors (e.g., exposing semantic-similarity scores to instructors, providing citations for every assessment item) to prevent black-box assessment design [334], as well as with students (i.e., clarity about AI involvement in assessment design) to maintain trust [212]. Accountability is framed not merely as error-checking but as a professional responsibility: educators must remain the final arbiters of curricular alignment and cognitive-level targeting, and institutions should establish clear accountability frameworks and policies for appropriate use [81]. AI literacy and assessment literacy are identified as mutually reinforcing competencies that must be developed in parallel. Faculty training should cover not only prompt engineering and model selection but also the critical appraisal of AI outputs, interpretation of psychometric indices, and the judgment required to decide when AI-generated items are not fit for deployment [134].

Table 7. Recommendations for AI-enhanced assessment design, with supporting references.

Recommendation	References
Choice of Model	[135], [245], [196], [296], [165], [72], [83], [178], [142]; [389], [32], [97], [419]; [355], [437], [382], [217], [144]
Prompt Engineering	Impact of prompting: [121], [106], [423], [101] Structured prompting: [186], [97], [104], [208], [81], [178], [474], [274], [376] Leveraging misconceptions: [242], [241] Few-shot exemplars: [219], [223], [402], [127] Chain-of-thought: [171], [127], [32], [97], [369] Teacher training: [86], [420], [443], [476]
RAG	[423], [110], [188], [364], [419], [274], [80], [333], [249], [111], [435] Metadata enrichment: [333], [95], [137], [367], [358], [135], [258]
Multiple Agents	[274], [452], [280], [80], [296], [249], [99], [203], [132], [171]
Fine-tuning	[274], [372], [127], [80], [209], [394], [158]
Human-in-the-loop	Workflow: [214], [477], [295], [300], [345], [337], [219], [87], [82], [248], [111], [264], [81], [245] Iterative refinement: [154], [196], [498], [383], [452], [429] Co-design: [92], [424], [505], [80]
Responsible Implementation	Equity and access: [133], [488], [505], [453], [319] Transparency: [334], [212], [419], [396] Accountability: [81], [212], [134], [85], [334], [419] AI and assessment literacy: [134], [238], [443]

Limitations and Gaps

Despite rapid growth in research on AI-assisted assessment design, the evidence base remains substantially stronger on technical feasibility and item-level quality than on pedagogical effectiveness and institutional impact. Existing limitations concern the contexts in which systems have been studied, the methods used to establish assessment quality and educational impact, and persistent limitations of the underlying generative technologies.

Limitations in Sample, Scope, and Generalizability Most studies rely on small samples in a single discipline. Exposure periods tend to be short. In many cases, deployment in real classrooms is lacking, leaving key questions about pedagogical impact, student engagement, and discriminative power under real examination conditions unanswered. A review of generative AI in K-12 education reinforces this finding, noting a lack of concrete experiments and longitudinal validation [243].

Limitations in Evaluation of Quality and Impact To assure assessment quality, expert review of AI-generated items is the dominant approach (see Table 4). However, inter-rater reliability is repeatedly flagged as problematic [232], [443], [302], [380], [245], [408], [189]. Moreover. measurement of impact of AI-assisted assessment design on student outcomes, such as learning and engagement, is limited, with a small set of studies reporting mostly positive findings. Long-term retention is almost universally unexamined: short assessment windows preclude conclusions about whether AI-supported assessment produces durable learning. The absence of learner-centered validation — collecting student feedback on the utility, clarity, or perceived fairness of AI-generated items — is also noted in several studies leaving a significant gap between expert-rated quality and the student experience of assessment.

Limitations in Generative AI Technology Studies documented extensive evidence of several quality deficits in AI-generated assessment items (see Table 5). The manifestation of these issues depends on several factors, including model version and prompt engineering, but also point to structural limitations of current generative approaches. Several mitigation strategies have shown promise in the literature, and are thoroughly discussed in the next section.

Conclusion and Future Directions

Generative AI in assessment design has progressed from proof-of-concept item generation toward systems that can scale conventional assessment design, augment it with new capabilities, and enable more interactive forms of assessment. At the same time, the evidence reviewed here shows that increasing technical capability does not by itself ensure educational value. Alongside promising evidence of assessment quality and new design possibilities, the literature documents inversion risks including factual inaccuracies, low cognitive demand, weak distractors, psychometric underperformance, and, in some settings, reduced engagement. The next phase of research therefore needs to move beyond demonstrating what generative AI can produce toward establishing when, how, and under what conditions AI-assisted assessment design improves assessment and learning.

Item quality must be systematically improved Future research can leverage advances in prompt engineering, retrieval-augmented generation techniques that ground model outputs in verified domain knowledge, and multi-agent architectures. Computational sustainability also deserves greater attention in this regard: architectures combining smaller finetuned models with RAG pipelines and rich metadata can approach the quality of large-model systems at substantially lower cost, making capable AI-assisted assessment design viable outside well-resourced institutions.

Multimodal, adaptive, and simulation-based assessments The field needs to move beyond it’s preoccupation with multipke choice questions to develop new kinds of assessment that can capture complex competencies and better reflect authentic performance contexts.

Evidence on student learning and engagement Research on the impact for student learning and engagement remains sparse, short-term, and largely confined to multiple-choice formats. Building a rigorous longitudinal evidence base across contexts is a prerequisite for informed scaling.

Professional capacity needs to be developed As AI tools take on a larger share of assessment authoring, the ability of educators to critically evaluate, refine, and take responsibility for AI-generated items cannot be assumed — it must be actively cultivated through sustained investment in AI literacy, assessment literacy, and clear institutional accountability frameworks.

Download underlying reference and analysis table

Applications for Assessment Evaluation

Generative AI is being applied to educational evaluation in ways that range from automating existing grading practices to enabling forms of assessment that would be difficult to implement without AI. The ISAR model (Bauer et al., 2025) provides a useful framework for distinguishing these applications according to how far AI changes the evaluation process. At the substitution level, AI performs evaluative functions that would otherwise be carried out by human graders. Augmentation extends the evaluation process by supporting the production of judgments, enriching feedback, or connecting evaluation to subsequent instructional action. At the redefinition level, AI changes the assessment activity itself by enabling forms of evaluation that would otherwise be difficult or impractical to implement. Across these levels, a smaller body of research has also examined the outcomes associated with AI-assisted evaluation, particularly its effects on efficiency, student learning, and user perceptions.

Substitution

The largest body of applications operates at the substitution level. These systems generally take student work as input, evaluate it against some representation of expected performance—such as a rubric, model answer, reference corpus, or learned examples—and produce scores or evaluative judgments that would otherwise require human grading. Applications differ primarily in the kinds of student evidence that can be evaluated.

What kind of student evidence can be evaluated? Text-based assessment is particularly prominent. For example, SmartGrading adapts automated scoring to different text-based assessment tasks by taking the assessment question, a sample solution or rubric, and evaluation instructions [55]. In computer science, BeGrading uses an LLM fine-tuned on programming submissions to evaluate and score code across multiple criteria [16]. In mathematics, models trained on annotated student problem-solving processes classify answer correctness and identify erroneous solution steps [197].

The formats in which student performance is captured extend the scope of substitution. Spoken or interactive performances can be evaluated from transcripts [123], [1], [26], while multimodal systems evaluate performance directly from audio, visual, or spatial representations [212], [157], [221], [47], [151]. Several systems also evaluate handwritten work directly from scanned responses [34], [174], [188], [176], [170]. Substitution therefore extends across a broad range of evidence, from conventional written responses to programming, mathematical reasoning, spoken performance, handwritten work, and multimodal artefacts. Table 1 summarizes this range of student evidence evaluated across the reviewed studies.

Automatization of scoring is implemented in some systems, including real-time rubric-based essay scoring [167], [140] and automated code assessment [5], [100].

Table 1. Studies using AI for substitution-level evaluation tasks, by type of student evidence evaluated.

Type of student evidence evaluated	References
Extended writing	[202], [40], [179], [36], [217], [187], [170], [54], [149], [165], [155], [135], [6], [173], [181], [122], [171], [146], [211], [213], [2], [13], [147], [167]
Short-answer responses	[144], [129], [150], [174], [48], [184], [175], [55], [49], [176], [21], [140]
Open-ended responses	[34], [138], [210], [55], [62], [49], [180], [145], [168]
Mathematical/scientific problem solving	[43], [14], [61], [188], [197], [142]
Programming/code	[3], [46], [33], [223], [169], [116], [16], [2], [5], [100], [117]
Creativity/design	[10], [148], [182], [207], [152], [11], [154]
Translation/language production	[159], [208]
Clinical/professional performance	[123], [1], [26], [121], [222], [161]
Discipline-specific artefacts	[153], [206], [214], [194], [18], [136], [178]
Multimodal procedural/performance evidence	[212], [157], [221], [47], [204], [151], [115]

Augmentation

At the augmentation level, AI does more than perform an evaluative function: it extends what can be done within or on the basis of evaluation. Across the reviewed applications, this occurs in three main ways: by making evaluative judgments more transparent or subject to additional quality control, by enhancing the feedback accompanying evaluation, and by using evaluation to inform subsequent instructional action. Table 2 summarizes these mechanisms of augmentation.

Enhancing the quality and transparency of evaluation Some systems make the basis of AI-generated judgments more explicit and auditable. The GPT-eComplex Assistant, for example, extracts supporting evidence and provides citations and justifications for rubric-based judgments, while indicating when the available evidence is insufficient. Rather than assigning final grades, these traceable outputs are provided for human evaluators to support their assessment and potentially improve grading accuracy [186], [228], [162], [140]. Other approaches add quality-control mechanisms around the judgment itself. Post-grading analytics can identify rubric criteria associated with high scoring variability for subsequent faculty review [158], while multi-agent architectures can assign separate agents to check evaluations for potential bias, inconsistencies, or omissions before results are released [211].

Enhancing feedback AI can also extend an evaluative judgment into more detailed diagnostic, explanatory, or actionable feedback. In one study, human grading comments were typically only two or three words long, whereas a GPT-4-based system generated multi-sentence feedback identifying errors, explaining why an approach was incorrect, and suggesting improvements [215]. Other systems add formative elements such as metacognitive prompts and actionable summaries [211]. Augmentation can also operate through the human evaluator: providing graders with AI-generated rationales has been shown to improve the quality of the feedback they subsequently provide to students [228]. Finally, AI-assisted assessment systems enable real-time feedback [143], [100].

Connecting evaluation to instructional action A further step is to use information derived from evaluation to determine what should happen next. At the cohort level, evaluation data can be aggregated to identify recurring errors or patterns of topic mastery, as in a programming platform that presents such patterns through an instructor dashboard [100]. At the individual level, LLM systems can use identified errors, knowledge gaps and learner profiles to generate remediation tasks [46], recommend study materials [53], or generate personalized learning targets [143]. The same information can feed back into assessment itself, allowing subsequent tasks to be adapted to the learner. Systems have generated programming challenges based partly on previous performance [100], combined learner profiles and a domain knowledge graph to generate personalized diagnostic items [166], [91], and used students’ own submitted artefacts to generate individualized questions probing understanding of their own work [134]. Evaluation thereby becomes an input to subsequent instruction and assessment rather than simply the endpoint of an assessment cycle.

Table 2. Studies using AI to augment educational evaluation, by mechanism of augmentation.

Mechanism of augmentation	References
Enhancing the quality and transparency of evaluation	[162], [140], [228], [27], [186], [114], [55], [158], [211], [139]
Enhancing feedback	[167], [129], [138], [166], [197], [202], [1], [5], [194], [136], [26], [122], [2], [223], [48], [215], [228], [211], [203], [120], [4], [185], [3], [170], [173], [217], [196] [143], [100]
Connecting evaluation to instructional action	[100], [213], [46], [53], [143], [185], [91], [134], [166]

Redefinition

At the redefinition level, generative AI changes the assessment activity itself by participating in the process through which student performance is elicited. In the reviewed literature, this is most apparent in interactive simulation-based assessment.

Enabling interactive simulation-based assessment Generative AI can assume roles traditionally requiring another person, allowing students to demonstrate performance through dynamically generated interactions. In clinical education, for example, LLMs have been used as simulated patients for history-taking and clinical encounters [1], [26], [199], and as examiners conducting case-based oral examinations [102]. These interactive roles can moreover be integrated with evaluation within the same system: the resulting performance is analyzed to provide scores and feedback [1], [26], [102], [163], [199]. The redefinition therefore lies not in automated scoring, but in the integration of performance elicitation and evaluation within an AI-mediated interaction. Rather than evaluating only evidence produced through a pre-existing assessment activity, generative AI can participate in creating that activity, dynamically eliciting the performance on which the evaluation is based.

Observed outcomes

While the literature documents a wide range of applications, considerably fewer studies examine what happens when these systems are implemented. The available evidence is mainly concentrated in three areas: time efficiency, student learning, and perceptions of AI-assisted evaluation. Table 3 summarizes the empirically observed outcomes across these areas.

Time efficiency The clearest evidence concerns reductions in evaluation time and human effort. A retrieval-augmented system processed 90 open-ended examinations in approximately 40 minutes, compared with an estimated 25 hours of manual grading [138]. Automated rubric scoring of student reflections required 7.71 seconds per response, compared with an average of 1.4 minutes for human evaluators [211], while automated grading of 534 programming submissions took approximately 2.5 hours rather than an estimated 45 or more hours of human grading [169]. Efficiency gains can also come from reducing rather than replacing human review: automated identification of relevant segments in video-based assessment reduced the average footage requiring inspection from 900 to 175 seconds per encounter [204].

Student learning Evidence for learning effects remains limited. A knowledge-graph-enhanced assessment system was associated with higher post-test performance, particularly among initially lower-performing students [166]. Students receiving LLM-assisted formative testing obtained higher formative and subsequent summative assessment scores than a comparison group [53], while an automated pronunciation-assessment system was associated with improvements in pronunciation, self-regulated learning, and metacognitive knowledge [143].

Perceptions Responses to AI-assisted evaluation are mixed and appear sensitive to how AI is incorporated into the process [193]. Students have reported greater trust in instructor-only than ChatGPT-only grading, with hybrid grading occupying an intermediate position [56]. Other studies report positive perceptions of usefulness, fairness, or willingness to reuse AI-supported evaluation [169], [162], [223], whereas limited transparency about AI involvement has been associated with concerns about fairness and institutional trust [209]. Instructor responses likewise range from skepticism toward AI-generated grading [46] to positive assessments of AI-supported diagnostic and evaluation systems [143], [134]. AI-generated feedback has also been viewed more favorably when combined with instructor feedback than when presented alone.

Overall, the available outcome evidence suggests that AI-assisted evaluation can substantially reduce evaluator workload and, in some implementations, contribute to learning and receive high levels of user acceptance. These benefits are nevertheless uneven across applications and should not be taken as evidence that the underlying evaluative judgments are necessarily valid, reliable, or fair. Indeed, the literature also documents substantial disagreement with human evaluation, systematic scoring errors, misclassification of student performance, and instability across repeated evaluations. These findings motivate consideration of inversion: circumstances in which introducing AI into evaluation undermines rather than improves the quality or integrity of the assessment process.

Table 3. Empirically observed outcomes of AI-assisted evaluation.

Outcome	References
Time efficiency	[138], [211], [169], [187], [115], [123], [158], [188], [170], [136], [204], [228]
Student learning	[166], [53], [143], [199], [10], [215]
Perceptions	[56], [169], [65], [209], [223], [162], [163], [122], [146], [193], [46], [143], [134], [171]

Inversion and Containment

The potential benefits of generative AI for assessment depend on whether AI-supported evaluation preserves the validity, reliability, fairness, and trustworthiness of the assessment process. In terms of the ISAR framework, inversion occurs when the introduction of AI undermines rather than improves these qualities. Conditions that allow weaknesses in AI evaluation to propagate into consequential decisions amplify inversion, whereas interventions that reduce those weaknesses or interrupt their propagation contain it.

Inversion in AI-supported evaluation

Across the reviewed literature, AI-generated judgments can become disconnected from the performance they are intended to assess in several ways: they can vary with implementation conditions, diverge from human or expert reference judgments, fail to recognize or faithfully represent relevant evidence, systematically apply different evaluative standards, or produce different evaluation errors across student groups. Table 4 summarizes these dimensions of inversion.

Instability and implementation dependence A basic requirement of reliable evaluation is that judgments reflect the performance being assessed rather than arbitrary properties of the implementation. Generative AI does not always satisfy this requirement. Evaluations of the same student work using the same rubric can vary depending on the LLM used [144], [218], and instability can persist even when the model is held constant, with repeated evaluations producing different scores [46], [43]. Users of an LLM-based Moodle grading tool, for example, reported identical answers receiving different evaluations across attempts [122].

Divergence from human or reference evaluation The agreement between AI-generated and human judgments is a central question. Across the reviewed studies, the degree of correspondence varies substantially. Some AI evaluators showed high agreement with expert judgments, in some cases reaching or exceeding the level of agreement among the human experts themselves [144], [162]. At the other end of the observed range, six general-purpose LLMs used to grade language examinations showed little agreement with human reference grades [176]. Similarly, a comparison of 18 LLMs used to grade programming assignments found relatively weak agreement with teacher grades [117].

The granularity of the comparison of AI and human scoring paints a complex picture. While AI-generated scores may broadly track human scores across students [182], correspondence has been shown to vary substantially across the particular criteria or dimensions being assessed [147]. Strong correspondence in overall scores can therefore conceal substantially weaker correspondence on particular assessment dimensions [154].

The magnitude of the differences between AI-generated and human scores offers a further perspective. Even where AI and human scores show correspondence in their overall pattern across students, the scores assigned to individual responses may still differ substantially. In language-exam grading, for example, the magnitude of the differences between LLM-generated and reference scores was substantially greater than that observed for human graders [176]. Thus, correspondence in the overall pattern of scores does not necessarily imply that AI-generated judgments are close to human judgments in absolute terms.

Failures in evidence representation and recognition Before evaluative criteria can be applied, an AI evaluator must first faithfully represent and recognize the assessment-relevant evidence contained in student work. Failures at this stage can disconnect the subsequent judgment from the performance it is intended to assess.

Evidence detection failures can involve both missing evidence that is present and detecting evidence that is not. In short-answer assessment, LLMs frequently failed to recognize correct responses, indicating under-identification of evidence [175]. Conversely, ChatGPT-based classifiers identifying concepts in physics responses achieved high recall (0.88–0.90) but very low precision (0.16–0.20), indicating substantial over-identification of evidence [24].

Evidence that is not faithfully represented an give rise to these failures. Reported examples include mathematical hallucinations in physics grading [61], incorrect localization or invention of errors in programming submissions [3], alteration of student responses during evaluation by inserting clauses not present in the original submission [37], hallucinated rows or algebraic steps in calculus assessment [43], and inaccuracies introduced when handwritten responses are converted through OCR [170].

Evidence recognition failures fluctuate. Some errors or features are detected substantially more accurately than others [11]; in programming assessment, for example, F1 scores ranged from 0.09 to 0.59 across models and error categories [33]. Valid evidence may also go unrecognized when its surface form differs from what the evaluator expects, as when mathematically equivalent expressions or legitimate differences in rounding or numerical representation are treated as incorrect [142].

Evidence itself can interfere with evaluation during adversarial manipulation. Prompt injection embeds instructions within a student response that alter how the LLM interprets or executes the assessment task [49]. Here, the failure lies not in misrecognizing legitimate evidence but in allowing the submitted evidence to modify the evaluation process itself.

Systematic differences in evaluative standards Directional differences in AI and human scores form a specific type of divergence. Several studies report systematically higher AI-generated scores [135], [155], [47], [222], [121], in some cases affecting consequential classifications such as the proportion of students receiving honours [178]. Other studies find systematically more stringent AI evaluation [150], [173], or compression toward the middle of the scale, with points deducted from responses judged fully correct by humans while completely incorrect responses were rarely assigned zero [49].

AI and human evaluators can weigh assessment evidence differently. AI evaluators have been reported to reward verbose or disorganized responses [144], overscore inflated or off-topic responses while penalizing brief but correct answers [34], and place disproportionate emphasis on surface-level linguistic or organizational characteristics relative to deeper content quality [36], [116], [208], [171], [165], [121].

The influence of construct-irrelevant features paves the wax for strategic manipulation. Inflated or off-topic content can receive disproportionate credit [34], while insufficient penalties for incorrect or irrelevant material can make response padding advantageous [54]. Evaluative misalignment can therefore become a strategic vulnerability when students are able to optimize responses for properties rewarded by the AI evaluator rather than for the competence the assessment is intended to measure.

Differential evaluation and error patterns Importantly, evaluation errors may be unevenly distributed across students or characteristics of their responses. Some studies, including a large-scale analysis, found no evidence of demographic bias [14], [163]. Others identified differential error patterns: false-positive rates in LLM-based oral English assessment varied by students’ first-language background [143], [21], and grading errors in reflective-writing assessment were larger for lower-ability learners [211]. Differential performance can also arise from characteristics of the evidence itself: in handwritten-answer assessment, scoring performance varied substantially with handwriting quality [174]. Thus, aggregate evaluation performance can obscure whether errors are concentrated among particular students, response characteristics, or forms of evidence.

Table 4. Dimensions of inversion in AI-assisted evaluation.

Dimension of inversion	Empirical focus	References
Instability and implementation dependence	Variation across implementation conditions or repeated evaluations	[144], [218], [46], [43], [122], [49], [173], [184], [165]
Divergence from human or reference evaluation	Agreement or concordance (e.g., κ, QWK, ICC)	[144], [162], [10], [115], [123], [150], [155], [135], [117], [176], [184], [206]
	Correlation between AI and reference scores	[182], [147], [154], [208], [169], [165]
	Magnitude of score differences (e.g., MAE, RMSE, Bland–Altman)	[62], [161], [151], [153], [211], [176], [132], [169]
Failures in evidence representation and recognition	Missed evidence; false detection; hallucination or alteration; surface-form sensitivity; prompt injection	[175], [24], [11], [33], [185], [61], [3], [37], [43], [170], [49], [142], [184]
Systematic differences in evaluative standards	Leniency or stringency; score compression; differential weighting of features; response padding	[144], [34], [36], [171], [116], [208], [121], [6], [135], [155], [47], [222], [178], [150], [173], [49], [54], [206], [165], [25]
Differential evaluation and error patterns	Demographic differences; first-language effects; performance level; handwriting quality	[14], [163], [143], [146], [21], [211], [174]

Risk containment

The consequences of inversion depend not only on the frequency or magnitude of errors produced by an AI evaluator, but also on the conditions under which those errors are translated into assessment decisions. A residual error has different implications when it can be detected and corrected than when it remains opaque, when AI provides input to an independent human judgment than when its output directly determines a grade, and when deployment remains restricted to conditions for which performance has been established than when validity is assumed to generalize across assessments and populations.

Across the reviewed literature, containment operates at four complementary levels: improving the quality of AI-generated judgments, bounding the evaluative authority granted to those judgments, making residual errors detectable and correctable, and governing the conditions of deployment and data use. Table 5 summarizes these approaches to risk amplification and containment.

Improving the quality of AI-generated judgments The most direct form of containment acts on the AI-generated judgment itself. Across the reviewed studies, interventions seek to improve how evaluation is grounded in intended criteria and assessment evidence, adapt the evaluation system to the task and domain, reduce instability, and correct predictable differences in the resulting score scale.

Criterion and evidence grounding constrains both the standards applied by the evaluation system and the evidence on which its judgments are based. In handwritten open-ended assessment, rubric-guided evaluation produced substantially stronger correspondence with expert scores than evaluation without rubric guidance [34], [61], [152]. Evaluation can also be structured around the evidence supporting each criterion. In some evaluation systems, evidence extraction is separate from criterion-level scoring, explicitly linking identified evidence to the resulting evaluation and thereby improving alignment with human evaluators [140], [129], [162].

Graded examples and exemplars provide a further way of specifying how evaluative criteria should be applied. A customized GPT incorporating teacher-provided examples produced substantially stronger correspondence with teacher scores than the same approach without examples [181], see also [136], [213], [202]. Their effect is not uniformly positive, however: few-shot prompting did not improve alignment in one reviewed study [24].

Task and domain adaptation are possible when sufficient task-specific data or knowledge resources are available, and can align the evaluator more closely with the particular assessment context. Fine-tuning with domain-specific data such as student responses, teacher feedback, or scoring rationales can improve alignment [196], [27]. A programming-assessment system combining task-specific fine-tuning with grounding in a curricular knowledge graph achieved high agreement with expert consensus, whereas ablation of the knowledge graph substantially degraded grading performance [166]. Adaptation can also address imbalances in the distribution of training data across score levels. Where relatively few training examples represent particular scores, approaches have included reducing the number of responses from overrepresented score categories and augmenting underrepresented categories before model fine-tuning [14], [179].

Stabilizing judgments characterizes another set of interventions. Rather than relying on a single stochastic evaluation, repeated evaluations can be aggregated, an approach shown to increase score reliability in several studies [123], [2], [169], [129], [26]. Stochastic variation can also be constrained through model configuration, and several assessment systems used models at or near temperature zero [26], [121]. Such settings do not guarantee deterministic evaluation, however: identical inputs can still produce different outputs, and model behaviour may change following updates to the underlying system [11]. Dependence on a single evaluator can additionally be reduced by combining judgments across multiple model instances or components. One study combined small locally deployed language models through a collaborative grading architecture, achieving better grading than the models independently [145], [14].

Score calibration can correct predictable differences in the scale of AI-generated judgments. In one study, regression-based mapping translated systematically stringent LLM scores onto human-equivalent grading categories [215]. Such calibration is limited by the quality of the underlying evaluation: in another study, statistical rescaling was insufficient to correct differences in score dispersion and feedback when the underlying AI–human relationship was poor [61]. Calibration can therefore correct predictable differences in score scale, but not an evaluator that attends to the wrong evidence or makes substantively incorrect judgments.

Bounding evaluative authority. How much authority is given to AI judgments is a central question knowing that improving AI-generated judgments cannot fully eliminate evaluative uncertainty. An erroneous AI judgment has different implications when it informs an independent human decision than when it directly determines a grade, and the same error becomes more consequential when it affects a high-stakes summative decision rather than a low-stakes formative judgment. Accordingly, studies caution against treating AI-generated judgments as autonomous substitutes for human evaluation in consequential or summative assessment, emphasizing continued human involvement or more restricted formative and supporting roles [49], [175], [178], [121], [163], [222]. Concerns about overreliance, reduced teacher autonomy, inappropriate delegation, and diffusion of accountability have been raised across AI-supported assessment contexts [224], [209], [20], [122]. Effective human oversight therefore requires both the capacity to scrutinize AI judgments independently and clearly assigned responsibility for the resulting evaluative decision.

Bounding evaluative authority can be achieved by making automation conditional on characteristics of the individual judgment. Confidence estimates, for example, have been used in large-enrollment examinations to determine which AI-generated point assignments could be accepted automatically and which required human evaluation [188], [136], [138]. Another proposed workflow replaces one of two human raters with AI while referring AI–human disagreements to an additional human rater [14], [55]. These approaches allow AI judgments to carry greater authority when there is less indication of uncertainty or disagreement, while directing other cases to human review. A more general constraint is to reserve final evaluative authority for humans irrespective of the characteristics of the individual judgment. AI judgments may provide preliminary assessment or support triage with necessary faculty confirmation to finalize scores [114], [171], [116] while another AI evaluation system was explicitly configured as a supporting evaluator that generated rubric-based judgments and evidence without assigning final grades [186]. Together, these approaches illustrate a continuum in which AI authority ranges from conditional automation of selected judgments to a strictly supporting role in which final evaluation remains human.

Making residual errors detectable and correctable. Whether questionable judgments can be identified is important for effective oversight. AI-generated scores can be difficult to interpret or audit, and AI-generated scoring rationales should not be assumed to provide faithful access to the process underlying a judgment [10], [44], [169], [27], [13]. A LLM-powered grading system, for example, assigned high scores to flawed work while generating justifications that contradicted those scores [2]. Detectability therefore requires more than apparently plausible explanations.

Inspectability can be improved at the level of an individual judgment by exposing the assessment evidence on which the judgment is based. Systems have separated evidence extraction from scoring to create a traceable sequence from rubric criterion to identified evidence and assigned judgment [140], [129], [162]. These designs do not necessarily prevent erroneous judgments, but they make their evidential basis more accessible to human inspection. A complementary approach is to expose characteristics of the evaluator itself: post-hoc methods such as SHAP and LIME have been proposed to identify features contributing to automated assessment predictions [157], [229].

Uncertainty can likewise be made visible. Authors recommend communicating confidence or uncertainty alongside AI judgments so that cases requiring additional scrutiny can be identified [44], [176]. Such information becomes particularly consequential when connected to the allocation of evaluative authority described above: uncertainty is actionable when it triggers additional evaluation.

Monitoring and revalidation extend detection of scoring failures across groups, items, performance levels, contexts, and time. Grading performance should be examined beyond aggregate estimates, including across demographic groups and other relevant subpopulations [143], [123], [144], [14]. Revalidation over time is also necessary because performance established during development cannot be assumed to remain stable across populations, assessment tasks, configurations, or changes to the underlying model [11], [49], [62]. Monitoring can also be incorporated directly into the evaluation architecture, as in a multi-agent system incorporating an Equity Monitor and fairness checks that compare scoring errors across learner groups, providing a mechanism for identifying and addressing disparities [211].

Correction and contestability must follow from detection. Meaningful contestability requires transparency about whether and how AI contributed to an evaluation together with a mechanism through which its judgments can be challenged and referred for human review [209]. In some implemented systems, students have been able to request human regrading of AI-supported evaluations [120], [188]. Contestability therefore provides a final corrective layer when a questionable judgment has not been identified before reaching the student.

Governing the conditions of deployment. The outermost layer of containment concerns whether a system that performs adequately under particular conditions is deployed only where the evidence and institutional environment support its use. This shifts the focus from the quality of individual judgments to the conditions under which an evaluation system is introduced, expanded, operated, and supplied with student data.

Context-specific validation and staged deployment is a requirement. Evidence of performance in one assessment, population, model, or configuration does not establish suitability in another. Several studies therefore exemplify or recommend restricted, iterative, or human-supervised deployment, with broader or more consequential uses introduced only as evidence supporting the system in its intended context accumulates [187], [134], [60], [144]. Proposed validation processes include comparison against human evaluation, instructor usability testing, monitoring of student outcomes, and explicit criteria for deciding whether deployment should be expanded [60].

Institutional capacity and accountability should accompany deployment. Meaningful use of AI-supported evaluation requires instructors or other responsible personnel to understand model limitations, configure systems appropriately, interpret their outputs, and exercise informed oversight [143], [186], [181]. Institutions must also make explicit who is accountable for the configuration, deployment, and oversight of AI-supported evaluation, rather than allowing responsibility to become diffuse as AI is incorporated into assessment processes [209], [138]. Students should likewise be informed when AI contributes to evaluation and given sufficient information about its role and limitations to interpret AI-generated judgments critically rather than treating them as inherently authoritative [146], [182], [173], [56], [193], [171].

Data protection and governance need to be satisfied in the context of deployment. AI-supported evaluation can involve processing student work through external models or services, making compliance with applicable data-protection laws a condition of deployment. The literature identifies consent, data storage, security, regulatory compliance, and institutional control over student data as issues that must be addressed when AI is incorporated into evaluation [209]. The reviewed studies address these requirements through several complementary strategies. One is to minimize the identifiable information exposed to the model. Student submissions have been pseudonymized by removing names and student identifiers before being processed by AI [169], [115], [31]. A second strategy concerns how external models are accessed. Rather than transmitting student data through ordinary consumer-facing services, studies have used enterprise or otherwise privacy-compliant APIs [169], [204]. A further strategy is to run inference within institutionally controlled infrastructure, avoiding external transmission of student data altogether. An automated grading architecture based on collaborating small-parameter language models demonstrated that competitive grading performance could be achieved through local deployment [145], [202], [114], [41].

Table 5. Conditions of risk amplification and approaches to containing inversion in AI-assisted evaluation.

Level	Empirical approaches	References
Quality of AI-generated judgments	Rubrics and reference answers; criterion-level evaluation; graded exemplars; task/domain adaptation; repeated-run aggregation; model configuration; score calibration	[34], [61], [152], [140], [129], [162], [2], [37], [139], [169], [13], [138], [186], [55], [181], [136], [213], [202], [24], [196], [27], [166], [14], [179], [40], [16], [161], [143], [159], [123], [26], [121], [11], [145], [215]
Bounding evaluative authority	Formative/supporting roles; human oversight; stakes-sensitive use; confidence/agreement thresholds; escalation of uncertain cases; human final authority	[49], [175], [178], [121], [163], [222], [224], [209], [20], [122], [44], [138], [188], [14], [114], [186], [136], [55], [171], [116]
Detectability and correctability of residual errors	Evidence-linked judgments; criterion-level inspectability; uncertainty signals; subgroup and item-level monitoring; revalidation; disclosure and contestability	[10], [44], [169], [27], [13], [2], [140], [129], [162], [139], [157], [229], [176], [143], [123], [144], [14], [11], [49], [62], [211], [209], [120], [188]
Conditions of deployment and governance	Context-specific validation; restricted pilots; staged deployment; institutional expertise and accountability; privacy-compliant infrastructure; de-identification; local/on-device deployment	[187], [134], [60], [144], [117], [143], [186], [181], [209], [138], [146], [182], [173], [56], [193], [171], [169], [115], [31], [204], [145], [202], [114], [41]

Limitations and Gaps

The empirical evidence reviewed here remains subject to several limitations that constrain what can presently be concluded about generative AI as an educational evaluator. Most importantly, the literature provides substantially stronger evidence about how particular systems perform under particular evaluation conditions than about the general validity or educational consequences of AI-supported evaluation.

The evidence base remains contextually bounded Studies commonly examine particular courses, institutions, cohorts, assessment tasks, languages, or response formats, while repeated and longitudinal deployments remain limited. There are important exceptions. One study evaluated 2,288 responses to 82 questions across 12 medical courses and three languages [49], while another examined multilingual responses from eight countries and six languages in an international large-scale assessment [21]. Such studies demonstrate that broader evaluation is feasible, but remain atypical of the evidence base reviewed here. Even where empirical coverage is broader, however, generalizability remains constrained by the specificity and rapid evolution of the technical implementations being evaluated. Performance may depend on the particular model, model version, prompting strategy, configuration, and surrounding architecture, meaning that evidence obtained for one implementation cannot automatically be transferred to another or assumed to remain applicable as the technology changes. The available literature therefore provides evidence primarily about particular combinations of assessment, population, model, and implementation conditions rather than about the general evaluative capability of generative AI.

What constitutes adequate validation? Validation in the reviewed literature is dominated by comparisons between AI-generated and human judgments, but the approaches used to quantify this correspondence vary considerably. Studies report correlations, agreement coefficients, classification measures, score differences, regression-based comparisons, and other error metrics, each capturing different properties of evaluator correspondence and limiting comparability across studies. Broader psychometric evidence is much less common. Notable exceptions include the use of Rasch modelling to examine the measurement properties of AI-generated competency assessments [203], [21].

Moreover, correspondence with human evaluation is an incomplete standard of validity. Human judgments themselves vary, so disagreement cannot necessarily be attributed to the AI evaluator, particularly when a single human rating is treated as the reference. A notable exception is the international large-scale assessment study, which compared human–machine agreement with human–human reliability and investigated disagreements as potentially arising from automated-scoring errors, human-scoring inconsistencies, or genuine ambiguity in the response [21]. Such analyses remain unusual.

More fundamentally, the literature lacks a common framework specifying which properties of AI-generated evaluation should be systematically validated and reported. The inversion patterns identified in this review suggest that validation cannot be reduced to correspondence with human scores. High human–AI agreement does not establish that both evaluators base their judgments on the same assessment-relevant evidence. Aggregate correspondence often conceals substantial differences for individual responses, criteria, performance levels, or groups. Studies often lack evidence of whether judgments are stable across repeated evaluations; whether the evaluator faithfully recognizes and represents assessment-relevant evidence; whether that evidence is interpreted and weighted according to the intended evaluative criteria; whether errors differ systematically across relevant demographic groups; and whether the resulting scores preserve the measurement properties required of the assessment. Without greater consistency in examining and reporting these properties, it remains difficult both to compare findings across studies and to determine what evidence is sufficient to support a particular use of AI-generated evaluation.

Evidence on the consequences of introducing AI into evaluation is lacking Reported reductions in scoring time do not necessarily capture the additional work associated with verification, oversight, correction, or appeals. More fundamentally, replacing or supplementing human evaluation with AI may change the assessment process itself. Students’ knowledge that their work will be evaluated by AI could affect how they learn or behave during examinations, including attempts to optimize responses for or influence the evaluator. AI grading may also affect perceptions of fairness, legitimacy, and accountability. These system-level consequences remain difficult to infer from studies that primarily compare AI-generated and human-generated scores.

Taken together, the current literature establishes considerably more about whether particular AI systems can reproduce or approximate evaluative judgments under specified conditions than about the validity and consequences of AI-supported evaluation when embedded in educational assessment. The central limitation is therefore not simply the amount of available evidence, but its scope: validation has generally focused on the performance of particular evaluators under bounded conditions, while the uses for which AI evaluation may ultimately be considered require evidence about the validity of the resulting judgments, their robustness across relevant conditions, and their consequences within the wider assessment system.

Conclusion and Future Directions

This review shows that generative AI can assume extensive roles in educational evaluation: substituting for human evaluators in existing scoring tasks, augmenting evaluation through richer feedback, greater traceability, and links to subsequent instructional action, and enabling interactive forms of assessment in which AI participates in eliciting as well as evaluating performance. Across these applications, scalability represents the clearest demonstrated benefit, with substantial reductions in scoring or review effort reported in several settings.

At the same time, evaluative capability should not be equated with evaluative validity. AI judgments can vary across repeated evaluations and system configurations, diverge from human or reference judgments, fail to recognize or faithfully represent relevant evidence, apply systematically different evaluative standards, and produce differential patterns of error across responses or populations. These forms of inversion become more consequential when automated judgments are given greater authority or deployed beyond the conditions under which their performance has been established. The containment approaches identified in the literature can improve individual judgments or prevent residual errors from propagating into consequential decisions, but they do not remove the need to establish that the evaluation itself is valid for its intended use.

More comprehensive and systematic approaches to validation The dimensions of inversion identified in this review provide one basis for defining what such validation should encompass. Beyond establishing correspondence with human scores, studies should determine whether an AI evaluator faithfully represents and recognizes assessment-relevant evidence, applies the intended criteria and weights assessment evidence appropriately, produces stable judgments across repeated evaluations, performs consistently across different demographic groups, and preserves the psychometric properties required of the assessment. Greater consistency in the measures used to evaluate and report these properties would also make evidence more comparable across studies and systems.

Evidential basis of AI-generated judgments Similar aggregate scores or high human–AI agreement can conceal differences in what evaluators reward or penalize. An AI evaluator may therefore approximate human scores while relying on different or construct-irrelevant characteristics of student work, or while compensating for errors in one part of a judgment with errors in another. Validation should consequently examine not only whether AI and human evaluators reach similar outcomes, but whether those outcomes are grounded in the assessment-relevant evidence and criteria they are intended to reflect. This moves validation from reproducing human scores toward establishing whether AI-generated judgments support the interpretations and decisions that the assessment is intended to permit.

How does AI grading changes assessment as a system Efficiency should be evaluated across the complete workflow, including the resources required for verification, oversight, correction, and contestation rather than scoring time alone. Research should examine whether knowledge of AI grading changes how students prepare for assessments, construct responses, or behave during examinations; whether new strategies emerge for influencing automated evaluators; how students respond to AI-generated judgments and mechanisms for challenging them; and how responsibility and workload are redistributed among students, educators, and institutions. These questions require prospective and sustained deployments rather than only retrospective comparisons of AI and human scores.

Generative AI should consequently be studied not only as an evaluator, but as one component of an assessment system. Its educational value depends on the relationship between the evidence available to the model, the construct being assessed, the authority delegated to AI, and the mechanisms through which errors can be detected, contested, and corrected. The central question for the next stage of research is therefore no longer simply whether generative AI can evaluate student work, but under what configurations AI-supported evaluation can produce defensible educational judgments and meaningful educational benefit.

🚧 This practice is currently being updated

The synthesis below is being revised and may be incomplete or out of date. Check back soon for the refreshed version.

Download underlying reference and analysis table

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 [153], [185], [38], [1], [10], [115], [16], [8], [129], [130], [235], [213], [15], [44], [173], [26], [102], [249], [176], [110][223], [80], [11], [20].

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 [100], [101], [44], [137], [221], [145], [48], [70], [217], [9], [144].

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 [49], [32], [42], [66], [19], [208], [172], [77], [227].

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 [205], [210], [35], [184], [139], [71], [153], [155], [99], [188], [192], [186], [58].

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 [152], [155], [151], [145], [219], [67], [142], [95], [72], [104], [79], [91], [93], [6], [186], [9], and in the educational field more generally [216], [135], [58], [239], [112]. Realtime delivery of feedback is particularly relevant for language learning, where students can get feedback on speaking performance during realtime language practice [143], [79], [18], [192], [122].

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 [94], [24], [182], [167], [30], [236], [33], [29], [25], [162], with an additional example in physics [87] and teacher education [157].

Measures and Outcomes

Feedback Quality

The question of how AI-generated feedback compares to human feedback has attracted substantial empirical attention, and the evidence divides broadly into two camps: studies documenting AI feedback as effective, comparable to, or in some respects superior to human feedback, and studies reporting mixed results, domain-specific shortcomings, or fundamental limitations in comparability.

Several studies have argued that AI-generated feedback can be an effective support tool [6], [44], [50], [184], [235] and is comparable to or exceeds human feedback in some cases [34], [173], [20], [155], [42], [210], [53], [185], [172], [66].

However, other studies have reported mixed results, noting that AI-generated feedback can be ineffective or less effective in some areas, or is inconsistent and unreliable [100], [80], [144], [16], [157], [151], [55] and/or low comparability with human feedback [85], [249], [179], [160], [112], [23], [48], [137].

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. Students’ subjective perception of AI-generated feedback is shaped by a constellation of factors, including their prior attitudes toward AI technology and their experience with AI tools [165], [205], [4]. Often the preference for human or AI feedback depends on the student, with one study showing an even split of preference between students [110].

Perhaps surprisingly, 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, likely because of its greater length and comprehensiveness [217], [66]. The picture changes substantially when the source of feedback is disclosed. Research reveals a complex and sometimes contradictory pattern of responses. On one hand, some studies indicate that students prefer human feedback and express skepticism toward AI-generated responses [197], [104]. On the other hand, natural, flexible interaction experiences, disclosure transparency and student feedback literacy are correlated with positive attitudes [225], [52], [244], [93].

These findings suggest that, beyond the quality of the feedback content, the flexibility of the interaction design is a critical determinant of how students perceive AI-generated feedback.

Engagement

Research on student engagement with AI-generated feedback reveals a multidimensional picture that spans cognitive, affective, and behavioral dimensions, with evidence accumulating on both the benefits and the limitations of this mode of feedback delivery.

Many studies report positive effects across cognitive, affective, and behavioral engagement, often crediting the interactive and immediate nature of AI feedback [47], [5], [218], [8], [251], [239], [135], [170], [87].

Enhanced engagement with feedback has been associated with subsequent revision behavior, representing one of the most concrete outcome measures in this literature [95], [175], [115], [139].

Despite these positive findings, some studies suggest that student engagement with feedback remains generally low regardless of source. One of the most striking examples of the sheer scale of non-engagement: an analysis of 655 writing tasks involving 14,236 students found that, even though AI-generated feedback led to higher revision frequencies compared to teacher feedback alone, approximately half of all students still did not revise a single character in their text after receiving generative AI feedback [176], [120]. This finding underscores that the availability of AI feedback does not automatically translate into behavioral engagement.

Other drawbacks are cognitive overload and the time-consuming nature of processing AI feedback [41], [170], [21].

Taken together, the evidence suggests that AI-generated feedback can meaningfully stimulate cognitive, affective, and behavioral engagement, particularly when it is immediate, personalized, and integrated into iterative revision workflows. However, substantial proportions of students do not act on AI feedback, while some experience cognitive overload or find the feedback insufficiently nuanced. Finally, it is important to note that excessive usage has been linked to decreased performance outcomes [152].

Educational Outcomes

The body of evidence on AI-generated feedback in education spans a wide range of domains and outcome types, and the findings, while broadly encouraging, are nuanced in important ways.

In writing and language learning contexts, AI feedback has been associated with improved skills and performance [223], [129], [81], [247], [97], [125], [213], [69], [46], [15], [233], [143], [58], [1].

Similar positive outcomes extend beyond language domains, with studies documenting enhanced learning performance and academic achievement across various educational disciplines [30], [107], [188], [170], [239], [91], [104], [110], [135].

These benefits are often mediated by increased student engagement [218], [5], [87].

Beyond domain-specific learning outcomes, AI-generated feedback has been linked to various psychological and metacognitive benefits. These include increased motivation [69], [188], [236], reduced anxiety [232], [61], enhanced autonomy and perceived competence or self-efficacy [251], [123], [51], and improvements in self-regulated learning [79], [91].

However, the evidence also contains cases where significant learning gains were absent [229], [152].

Limitations and Recommendations

Enhancing AI Feedback

A recurring diagnosis across the literature is that off-the-shelf large language models, when prompted without domain grounding or pedagogical alignment, tend to produce feedback that is generic, inconsistent, or insufficiently actionable for learners. Developers have responded by developing a range of technical strategies aimed at making AI feedback more precise, contextually appropriate, and educationally sound. Three broad families of approaches have emerged: retrieval-augmented generation, fine-tuning and reinforcement learning, and multi-step architectures.

Retrieval-arugmented generation (RAG) ensures that feedback is contextually appropriate by drawing from relevant educational resources and domain-specific knowledge bases[22], [192]. For example, a programming education system called Tutor Kai explicitly linked its feedback to lecture videos and course content through retrieval [122].

Fine-tuning and reinforcement learning represent a more resource-intensive but potentially more powerful route to aligning model outputs with pedagogical needs. LLMs can be finetuned using specialized datasets tailored to particular academic disciplines, or reinforcement learning techniques can guide models toward producing feedback that aligns with educator objectives [192], [210], [6]. For example, in the context of mathematical tutoring dialogue, fine-tuned smaller models trained on the MATHDIAL dataset significantly surpassed the performance of much larger prompted LLMs, while keeping the rate of directly telling students answers lower [148].

Multi-step architectures decompose the feedback generation process into distinct, specialised stages, typically separating evaluation from pedagogical guidance and adding a validation layer to check quality criteria before feedback reaches students. This modular approach can be implemented through various architectural strategies. Multi-component systems may combine LLM-based modules with traditional rule-based c omponents, 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 [54], [71], [43], [77], [192]. For example, in programming assessment, one system incorporated separate stages for strict code evaluation, review quality maintenance, and filtering of unnecessary feedback [155].

The most pedagogically effective systems will combine several of these strategies: grounding outputs in domain knowledge through retrieval, aligning generation with educator objectives through fine-tuning, and decomposing the feedback task into specialised stages. Crucially, human oversight needs to be incorporated at critical validation points.

Designing Effective Human-AI Interaction

The accumulated evidence across disciplines converges on a foundational principle: generative AI functions most effectively in feedback contexts not as an autonomous replacement for human judgment but as a support tool operating within carefully designed human-AI workflows. Fully automated feedback systems, while scalable, carry documented risks that only hybrid architectures can adequately address.

Human-in-the-loop designs, in which AI generates a first pass that educators review before release, are particularly widely recommended as a practical bridge between the scalability of automated systems and pedagogical expertise, offering complementary strengths where teachers can for example address higher-order concerns while AI handles surface-level feedback [192], [115], [227].

Several interaction design principles emerge as critical to making human-AI feedback collaboration effective. Key principles include promoting transparency about AI limitations, maintaining instructor control, and creating natural, flexible interaction experiences [145], [225], [35], [84]. Systems must also leverage generative AI’s capabilities to promote inclusion and ensure equitable access for diverse learner populations [216], [252].

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 [206], [154], [234].

System Implementation Recommendations

Successful deployment requires attention to system scalability and performance optimization to handle educational workloads effectively [221], [219], [58]. Furthermore, privacy compliance represents a critical concern, as educational data requires special protection measures [192]. To address privacy concerns while maintaining functionality, institutions should consider implementing open-source LLMs that allow for local deployment and greater data control [101], [70], [44].

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 integration of human expertise as an essential component. Research should also emphasize AI’s potential for creating innovative feedback delivery methods that incorporate simulation and real-time interaction.

🚧 This practice is currently being updated

The synthesis below is being revised and may be incomplete or out of date. Check back soon for the refreshed version.

Download underlying reference and analysis table

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.

🚧 This practice is currently being updated

The synthesis below is being revised and may be incomplete or out of date. Check back soon for the refreshed version.

Download underlying reference and analysis table

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.

---

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.

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)}
}