Source-linked AI summary
Personalized Education in the AI Era: What to Expect Next?
Setareh Maghsudi, Andrew Lan, Jie Xu, Mihaela van der Schaar
TL;DR
AI/ML-based personalized education seeks to tailor knowledge acquisition while addressing unresolved challenges beyond recommendation, including peer absence, motivation, diversity, and bias. This paper reviews state-of-the-art research across six core topics, outlines limitations, and proposes future directions and solutions. It presents personalized education as a promising AI merit that can improve educational quality across several dimensions by adapting to learners.
Problem
Personalized education still leaves challenges unresolved, including peer absence, motivation, diversity, bias, lifelong support, data availability, feedback delay, and changing environments.
Method
The paper reviews existing work across six core personalized-education topics, outlines limitations, and proposes future research directions and solutions.
Results
The paper concludes that enabling personalized education is a major merit of AI and that adapting education to learners can improve quality across several dimensions.
Takeaways & Limitations
AI/ML-enabled platforms can extend personalized education beyond recommendations toward adaptive materials, assessment, curricula, and learner connections.
Abstract
from arXiv · showhide
The objective of personalized learning is to design an effective knowledge acquisition track that matches the learner's strengths and bypasses her weaknesses to ultimately meet her desired goal. This concept emerged several years ago and is being adopted by a rapidly-growing number of educational institutions around the globe. In recent years, the boost of artificial intelligence (AI) and machine learning (ML), together with the advances in big data analysis, has unfolded novel perspectives to enhance personalized education in numerous dimensions. By taking advantage of AI/ML methods, the educational platform precisely acquires the student's characteristics. This is done, in part, by observing the past experiences as well as analyzing the available big data through exploring the learners' features and similarities. It can, for example, recommend the most appropriate content among numerous accessible ones, advise a well-designed long-term curriculum, connect appropriate learners by suggestion, accurate performance evaluation, and the like. Still, several aspects of AI-based personalized education remain unexplored. These include, among others, compensating for the adverse effects of the absence of peers, creating and maintaining motivations for learning, increasing diversity, removing the biases induced by the data and algorithms, and the like. In this paper, while providing a brief review of state-of-the-art research, we investigate the challenges of AI/ML-based personalized education and discuss potential solutions.
I. Introduction
AI/ML-enabled personalized education extends online learning beyond content recommendation by adapting educational materials and supporting assessment, networking, and broader learner needs. The paper reviews six core topics, their limitations, and future research directions, including unresolved issues around lifelong learning, fairness, diversity, and peer interaction.
- Current limitations: Existing personalized education platforms are often reduced to recommender systems, although personalized education targets learning outcomes rather than engagement or profit.The paper distinguishes the student-centric, difficult-to-define objective of learning outcomes from the system-centric objective of maximizing user engagement.
- Potential benefits: AI-enabled personalized education aims to provide benefits resembling one-on-one instruction while retaining a per-student cost similar to large lecture classes.The paper frames this as combining strengths of online tools and individual tutoring.
- AI/ML approach: AI/ML-enabled education maps learner features and available data to personalized materials, recommendations, and lesson sequences.Performance data from assignments, exercises, and tests can support optimization of lesson sequences.
- Paper scope: The paper reviews existing work on six core personalized-education topics, outlines their limitations, and proposes future research directions.Figure 3 organizes the topics across technical, personal, and social aspects.
- Open challenges: Important unresolved challenges include lifelong support, insufficient data, delayed feedback, high diversity, fast-changing environments, inequality, fairness, and the loss of peer interaction.The paper also identifies content recommendation across heterogeneous levels, connected content bundles, performance evaluation, and conflicting objectives as open challenges.
- Social dimensions: Personalized education can support knowledge and expertise networks by connecting learners for mentorship, friendship, cooperation, inspiration, and motivation.The paper presents autonomous network formation and learner interaction as extensions beyond individualized content delivery.
II. Content Production and Recommendation
The section presents AI-assisted content production and personalized recommendation as complementary parts of effective education. It highlights progress in automated assessment and outcome analysis while identifying challenges in multimodal retrieval, experimentation, and balancing learning objectives.
- Content Production: Educational quality depends on learning content, motivating AI-assisted tools for summarization and automated question generation.Summarization tools extract key facts from long textbook sections, while question-generation systems can produce factual questions with short textual answers.
- Content Production: Human experts rated automatically generated questions higher in quality than questions produced by other methods.
- Content Production: Multimodal personalization must retrieve helpful content across text, formulas, figures, and diagrams when learners misunderstand concepts.Shared vector-space representations are suggested as one way to map multiple modalities for retrieval and content production.
- Content Production: AI can assist human content designers by identifying priority areas, drafting responses, checking content, and incorporating crowdsourced feedback.The section frames AI as an assistant during iterative content creation rather than as an autonomous replacement for human experts.
- Recommendation: Personalized recommendation should operate at microscopic levels such as questions and videos and macroscopic levels such as courses and textbooks.Analyzing learner-performance data can help detect which content is most effective for particular learning outcomes.
- Recommendation: Rapid experimentation remains necessary because conventional A/B testing can require long cycles, while synthetic learner models may support reinforcement-learning research despite limited real learner data.The section also calls for algorithms that balance conflicting objectives defined over different timescales, such as exam performance, course grades, and post-graduation skills.
III. Assessment and Evaluation
Assessment and evaluation methods estimate learner knowledge and predict future performance, but existing approaches often trade predictive accuracy against interpretability and lose information by reducing raw responses to grades.
- Assessment approaches: Learner assessment models estimate knowledge mastery from responses using static IRT models or dynamic knowledge-tracing models.IRT estimates latent ability and item difficulty, while knowledge tracing models learner knowledge evolution over time.
- Assessment approaches: 1PL IRT models represent learner ability and question difficulty as scalars linked to the probability of a correct response.Extensions add discrimination, guessing, or multidimensional ability and difficulty parameters.
- Assessment approaches: IRT models can provide relatively stable ability estimates by denoising learner responses and estimate assessment-question quality.Multidimensional variants use vectors to capture multiple aspects of ability and difficulty.
- Assessment approaches: Deep-learning knowledge-tracing models achieve state-of-the-art future-response prediction, although their advantage can be insignificant and may reduce interpretability.These models represent knowledge as a latent vector rather than relying only on expert-defined concepts.
- Assessment approaches: AKT combines attention networks with cognitive-theory-inspired modules, using monotonic attention and 1PL-IRT question embeddings to improve prediction while retaining interpretability.The method retrieves previously acquired knowledge for performance prediction, and experiments report that it outperforms existing knowledge-tracing models.
- Assessment challenges: Converting raw responses into graded responses loses information, including misconceptions signaled by particular multiple-choice distractors and rich open-ended responses.Raw-response models can support finer-grained personalization after each step of an open-ended problem and address learner difficulties more promptly.
IV. Life-long Learning
Lifelong learning requires adaptive, foresighted course planning because students, course sequences, constraints, and learning contexts vary. AI-based recommendation can learn from prior student records, but planning remains challenging in large, evolving decision spaces.
- Course sequence recommendation: Course planning must account for prerequisites, electives, mandatory courses, course availability, timing, and trade-offs between graduation time and grades.Myopic choices can delay required subsequent courses and prolong graduation.
- Course sequence recommendation: A static course sequence is suboptimal because students’ knowledge, experience, and performance evolve during learning, while their backgrounds, knowledge, and goals vary substantially.The same learning path is unlikely to best serve all students.
- Course sequence recommendation: An automated course-sequence system uses offline dynamic programming to learn candidate policies, then selects a policy online for each student based on background.Candidate policies target expected time to graduation or on-time graduation probability using anonymized student records.
- Beyond degree programs: Lifelong-learning plans must be tailored to specific contexts such as workplaces, where similar challenges may coexist with new ones.The paper also identifies a gap between school lectures and job requirements, including the importance of soft skills.
V. Incentives and Motivation
Personalized education must address motivation directly rather than relying only on engagement from recommendations or learning networks. Behavioral economics, self-determination theory, and self-efficacy theory provide frameworks for designing such support.
- Motivation in personalized education: Motivation remains a largely neglected aspect of personalized education and requires direct integration into learning platforms.Recommendations may increase engagement and satisfaction, but the paper describes these effects as insufficient on their own.
- Design implications: Motivation is an unobservable dynamic process that is difficult to measure directly but can be inferred from observations.The paper frames motivation as involving multiple learning-related features, including initiation, persistence, intensity, and behavior quality.
- Behavioral economics: Behavioral economics models learner decision-making through utility functions that reflect mistakes, heuristics, irrationality, and individual backgrounds.AI and ML could estimate a learner’s utility from platform data and feedback to predict reactions to incentives and allocate rewards.
- Self-determination theory: Self-determination theory emphasizes autonomy, competence, and connectedness as intrinsic needs that can support lasting motivation.The paper describes intrinsic motivation as often more effective, lower-cost, and longer-lasting than material rewards.
- Self-efficacy theory: Self-efficacy concerns confidence in performing a specific task and is shaped by observed information and past experience.Appropriate feedback and side-information can strengthen learners’ positive beliefs about their ability to perform well.
- Design implications: AI-based recommendation tools should provide alternatives, accurate assessment, feedback, and learner connections to support autonomy, competence, connectedness, and self-efficacy.The proposed design links micro- and macro-level choices with feedback, coherent learner networks, and interaction.
VI. Building Learning Networks
Learning networks can restore peer interaction, support scalable peer review, and enable content sharing, but effective network design must address heterogeneous reviewer capabilities and effort. Game-theoretic analysis offers tools for understanding and shaping these interactions.
- Peer interaction: Online personalized education risks losing peer interaction and the sense of community found in traditional classrooms.Peer review can provide scalable assessment when enrollment exceeds the number of teaching assistants.
- Future directions: A formal method for building education-related learning networks and a deep understanding of their effectiveness remain absent.Existing networks facilitate collaboration, question answering, and resource sharing, but their design principles are not yet formalized.
- Peer review: Peer-review systems face adverse selection because reviewer capabilities are often unknown and moral hazard because reviewer effort is unobservable.Review quality depends on both intrinsic capability and effort.
- Peer review: Existing matching and social-norm mechanisms are limited because they separately assume one-shot interactions or homogeneous reviewers.Learning-network systems therefore need designs that account simultaneously for capability differences and effort incentives.
- Peer review: Proposed peer-review designs should make high effort self-interested and produce ratings that reflect reviewers’ capabilities.The paper identifies this simultaneous treatment of effort and capability as the target for new system designs.
- Network formation: Game-theoretic analysis models learners as heterogeneous, self-interested agents who choose links based on benefits from content sharing and link-formation costs.This framework is intended to guide the construction of learning networks and protocols that motivate beneficial actions.
- Network formation: Equilibrium networks typically have a small core, a larger periphery, minimal connectivity, and short diameters independent of network size.Theoretical results associate short diameters with efficient information dissemination and minimal connectivity with lower construction costs.
- Future directions: Remote learning increases the need to understand peer knowledge flow and discussion-forum behavior because instructors cannot easily moderate activities at a distance.Future work includes combining forum activity with grades and designing automated strategies to moderate student activities.
VII. Diversity, Fairness, and Biases
AI-based personalized education raises fairness and bias challenges affecting underserved learners and requires both fair algorithms and governance mechanisms. Research examines fairness definitions, enforcement methods, and broader safeguards for responsible deployment.
- Biases and equity: AI-driven personalization can improve assessment, feedback, and content recommendation, but outcomes may differ across student subgroups because of training-data biases.Underserved students are often less represented in datasets and have less access to advanced digital educational systems.
- Defining fairness: Fairness research addresses definitions including individual fairness, group-level outcome parity, conditional parity, and counterfactual fairness.These definitions differ in whether they emphasize similar treatment, comparable predicted outcomes, or stability under changes to sensitive attributes.
- Fairness-aware algorithms: Fairness enforcement methods include preprocessing inputs, post-processing predictions, and imposing regularizers or constraints during algorithm training.The paper identifies training-time regularization and constraints as the most promising approach among those discussed.
- Fairness-aware algorithms: Fairness-promoting methods can reduce classification accuracy, while training-time approaches obtain better fairness–accuracy tradeoffs than other methods.The tradeoff means improving fairness may require sacrificing some predictive performance.
- Governance and safeguards: Responsible educational personalization requires fair algorithms alongside principles, practical guidelines, legislation, monitoring, validation, and fail-safe mechanisms.Proposed safeguards span data diversity and quality, performance guarantees across settings, misuse detection, and regulation of data ownership and sharing.
VIII. COVID-19 and AI-Enabled Personalized Education
COVID-19 intensified educational disruption and exposed unequal access to learning resources, while AI and ML offered ways to support personalized, fair, and data-informed responses. The paper discusses applications ranging from resource allocation to school-closure planning.
- Educational disruption: COVID-19 disrupted education and affected students differently according to country or region, family status, and individual characteristics.Reported complications include reduced learning ability, depression, loss of concentration, and declining physical fitness.
- Educational disruption: Reduced school attendance removed access to educational materials, learning support, peers, teachers, incentives, and evaluation.Many students also could not fully use online replacements because they lacked suitable devices, reliable internet, or an appropriate home learning environment.
- Expansion of online education: The pandemic accelerated research and development in personalized and distance education and increased online learning-tool use.Examples include web-based laboratories and online machine-learning education modules, particularly in undergraduate signal-processing education.
- AI-enabled responses: AI and ML can enhance online education through improved materials, fairness and diversity, testing, knowledge networks, and pandemic recovery support.The paper confines its attention to AI and ML while noting their broader relevance to distance and asynchronous education.
- AI-enabled responses: Machine learning can classify students by pandemic-related educational exposure, allocate resources under fairness constraints, and optimize school-closure plans.School-closure optimization can use features such as neighborhood, school size, and grade.
IX. Summary and Conclusion
The paper presents personalized education as a major educational application of AI, adapting learning to individual characteristics while supporting online education during abnormal circumstances. It reviews current challenges and outlines future research directions.
- Conclusion: AI-enabled personalized education is presented as one of AI’s most valuable merits in education.The paper frames personalization as a central focus of its discussion of AI and ML.
- Conclusion: Personalized education adapts to learners’ personality, talent, objectives, and background, and is reported to improve educational quality across several dimensions.The conclusion links these adaptations to the paradigm’s educational value.
- Online education: Online education is especially valuable during abnormal circumstances such as COVID-19 outbreaks or natural disasters.The paper contrasts online education’s resource requirements with conventional education’s dependence on space, scheduling, and human resources.
- Research outlook: Despite its transformative potential, personalized education remains associated with several challenges.The conclusion qualifies the expected benefits rather than presenting personalization as unconditionally effective.
- Research outlook: The paper reviews state-of-the-art research, discusses challenges, proposes solutions, and summarizes future research directions.Table II provides a summary of some proposed research directions.