Generative AI Could Reshape Personalized Learning, But Major Gaps Remain

Generative artificial intelligence has moved from research labs into classrooms at a pace few educational technologies have ever matched, and a new analysis argues that the technology could fundamentally reshape how learning is personalized for every student. In a study published in Frontiers of Digital Education, Yaxin Tu and Changqin Huang of Zhejiang University, together with Jili Chen of Zhejiang Normal University, map out exactly how large language models, the engines behind tools such as ChatGPT, can be harnessed to set individualized learning objectives, adapt learning patterns, generate tailored resources, and build new evaluation systems. But the paper is equally clear about the other side of the ledger: current systems still struggle to understand who a learner actually is, how their learning unfolds over time, and how to cultivate the higher-order skills that education ultimately exists to develop.

The core promise of generative AI in education lies in its ability to produce content and dialogue on demand rather than merely retrieve pre-packaged material. Large language models can interpret a student’s request, reason over vast bodies of knowledge, and generate explanations, practice problems, and feedback calibrated to that student’s level. The researchers describe this as enabling automated, humanized, and personalized learning services, a combination that has become a central topic in the ongoing transformation of education. Where earlier adaptive learning platforms relied on rigid rule systems and predefined content libraries, generative models can compose responses in real time, adjusting tone, difficulty, and format as a conversation progresses. This flexibility opens the door to tutoring experiences that feel less like interacting with software and more like working with a responsive human mentor.

Technically, the study identifies several strategies that make this personalization possible. Retrieval-augmented generation allows a model to ground its answers in authoritative course materials rather than relying solely on patterns learned during training, reducing the risk of fabricated content. Knowledge tracing techniques, which model a learner’s evolving mastery of specific concepts, can be integrated with language models so that recommendations reflect actual performance data rather than surface-level interaction patterns. Multi-agent architectures, in which several specialized AI agents collaborate, can divide the work of planning learning paths, generating resources, and assessing progress. Researchers have also developed education-specific models, such as systems fine-tuned for Socratic questioning, English reading comprehension support, and intelligent tutoring, that adapt general-purpose language models to pedagogical goals.

The application landscape the authors survey is broad. Generative AI can help set personalized learning objectives by analyzing a student’s current state and suggesting realistic targets. It can shape learning patterns by recommending paths through material, pacing activities, and choosing modalities that suit individual preferences. It can construct learning resources on demand, from worked examples to alternative explanations of the same concept. It can also contribute to evaluation, generating formative assessments and interpreting student responses in ways that inform the next instructional step. Studies cited in the analysis report applications ranging from AI assistants that deliver personalized and adaptive learning in higher education to conversational agents that provide affective and motivational feedback, suggesting that the technology’s reach extends well beyond simple question answering.

Yet the paper’s most valuable contribution may be its unsparing diagnosis of what generative AI still cannot do. The authors highlight significant limitations in understanding differences in individual static characteristics, such as cognitive profiles, prior knowledge, and learning styles, as well as dynamic learning processes, the moment-to-moment evolution of attention, motivation, and understanding. A language model may produce fluent responses, but fluency is not the same as insight into a particular learner. The study also points to insufficient capacity for actively differentiating and adapting to these differences, meaning that many so-called personalized systems deliver variation in surface form rather than genuine pedagogical adaptation. Without a deep model of the learner, personalization risks becoming a label rather than a reality.

Compounding these technical gaps, the researchers identify a lag in theoretical foundations and a lack of practical guidance. Educational theories such as constructivism, distributed cognition, embodied cognition, and the theory of multiple intelligences were developed long before generative models existed, and the field has not yet systematically integrated them with the capabilities of modern AI. The result is a technology racing ahead of the science meant to explain how it should be used. Key technologies also remain weak in autonomy and controllability: models can behave unpredictably, and educators have limited means to constrain or steer their behavior in pedagogically sound ways. For a domain where wrong guidance can compound misconceptions, controllability is not a luxury but a requirement.

Perhaps the most consequential concern involves higher-order literacy. Education aims to develop critical thinking, creativity, self-regulation, and collaboration, not just content mastery. The authors argue that current generative AI systems lack mechanisms for enhancing these capacities, and may even undermine them. Recent research they cite warns of metacognitive laziness, where students offload thinking to AI and skip the productive struggle that drives deep learning. If a model always supplies the answer, the student may never develop the habit of constructing it independently. The study also flags deficiencies in safety and ethical regulations, including privacy risks associated with collecting fine-grained learner data, the potential for biased or inappropriate content, and the absence of clear accountability frameworks when AI-generated guidance goes wrong.

In response, the authors propose a set of implementation pathways designed to move the field from enthusiasm toward sustainable practice. The first is interdisciplinary theoretical innovation: bringing together learning scientists, computer scientists, and educators to build new frameworks that connect generative AI capabilities with established theories of how people learn. The second is continued development of large language models themselves, including education-specific models optimized for pedagogy, along with efficiency techniques such as model pruning, knowledge distillation, and cloud-edge collaboration that could bring sophisticated AI to resource-constrained schools and devices. The third is enhancing personalized basic services, ensuring that objective setting, resource generation, and assessment genuinely reflect individual differences rather than superficial customization.

The remaining pathways address the deeper challenges. Improving higher-order literacy requires designing AI systems that scaffold thinking rather than replace it, for example by adopting Socratic questioning strategies that guide students toward answers instead of handing them over. Optimizing long-term evidence-based effects means tracking learners over extended periods, using learning analytics to verify that AI-supported personalization actually improves outcomes, rather than relying on short-term engagement metrics. Finally, the authors call for establishing a safety and ethical value regulation system, encompassing privacy-preserving techniques such as federated learning, transparent governance of AI use in schools, and clear norms that keep human teachers in charge of pedagogical decisions. Together, these six pathways aim at what the researchers describe as safe, efficient, and sustainable personalized learning.

The significance of this analysis extends beyond any single classroom. With generative AI already embedded in homework help, essay drafting, and study planning for millions of students worldwide, the question is no longer whether these tools will shape education but whether they will do so thoughtfully. The study, supported by the National Natural Science Foundation of China, offers a roadmap that treats personalization not as a marketing feature but as a precise educational science, one that demands better learner models, stronger theory, controllable technology, and robust ethical guardrails. If the field follows that roadmap, the vision of an AI tutor that truly understands each learner, challenges them appropriately, and protects their autonomy and privacy may move from promise to practice. If it does not, schools risk deploying powerful technology that personalizes little, teaches less, and quietly erodes the very skills education is meant to build.

Subject of Research: Generative artificial intelligence mechanisms, challenges, and implementation pathways for personalized learning

Article Title: Empowering Personalized Learning with Generative Artificial Intelligence: Mechanisms, Challenges and Pathways

Article References: Tu, Y., Chen, J., & Huang, C. (2025). Empowering Personalized Learning with Generative Artificial Intelligence: Mechanisms, Challenges and Pathways. Frontiers of Digital Education, 2(2), Article 19. https://doi.org/10.1007/s44366-025-0056-9

Image Credits: AI Generated

DOI: 10.1007/s44366-025-0056-9

Keywords: generative AI, personalized learning, large language models, education technology, intelligent tutoring, learning analytics, knowledge tracing, self-regulated learning, AI ethics, educational theory, multi-agent systems, digital education