AI Digital Humans Bring Real-Time Tutoring to Pre-Recorded Courses

Pre-recorded video lectures have become one of the dominant ways students around the world consume educational content, offering the freedom to learn anytime and at any pace. Yet that flexibility comes at a well-documented cost: when a learner hits a conceptual wall halfway through a recorded lesson, there is no instructor present to answer a question, no classmate to ask, and often no parent equipped to help. A research team led by Qi Liu and Yunhao Sha of the University of Science and Technology of China, together with colleagues at the Hefei Comprehensive National Science Center and Hefei Normal University, argues that this interaction gap is quietly eroding motivation for millions of self-paced learners. Their response, published in Frontiers of Digital Education, is an intelligent teaching system built around a digital human — a lifelike virtual instructor powered by large language models that can converse with students in real time while they watch recorded material.

The problem the researchers set out to solve is structural rather than cosmetic. Live-streamed classes, their analysis notes, are frequently derailed by scheduling conflicts between instructors and students, which limits accessibility and pushes learners toward recorded alternatives. Those alternatives, in turn, strip away the real-time interaction and direct guidance that make live teaching effective. Studies cited by the team, including comparisons of student performance and study behaviors across live and pre-recorded formats during the COVID-19 pandemic, suggest that the absence of immediate support leaves students stuck on obstacles far longer than they would be in a classroom. For younger learners, the burden often falls on parents who lack the subject expertise to intervene, and repeated unresolved stumbling blocks cause motivation to diminish considerably. The team’s central claim is that the trade-off between flexibility and interactivity is not inevitable — it can be engineered away.

At the heart of the system is a large language model that serves as the digital human’s cognitive engine. Rather than answering questions from a fixed script, the model generates tailored, context-aware responses to individual student queries, adapting its explanations to the learner’s apparent level of understanding and progress through the course. The authors draw on a rapidly maturing body of work on LLMs in education, from technical reports on GPT-4 to surveys of pedagogical alignment, which show that these models can support lesson planning, Socratic questioning, and personalized response generation. By embedding such a model inside a virtual human presented alongside recorded lectures, the system aims to simulate the presence of a human instructor — one who answers questions, offers tailored guidance, and tracks individual progress without ever needing to be scheduled.

The digital human itself is more than a chat window. The researchers integrate advances in talking-head generation and human reaction modeling, fields that have produced transformer-based methods for generating natural facial expressions, gestures, and conversational responses. Text-to-speech techniques such as FastSpeech provide controllable, robust voice output, allowing the virtual instructor to speak its answers aloud rather than displaying them as text. This multimodal presentation matters for a reason the team traces through the virtual-human literature: embodiment and personalization increase a learner’s sense of self-identification with the agent, and perceived plausibility shapes how much students trust and engage with virtual characters. A disembodied chatbot attached to a video player, the authors suggest, would not produce the same social presence as an animated instructor who appears to be teaching alongside the recorded lesson.

Personalization is the second pillar of the design. The system draws on the research group’s earlier work on exercise-aware knowledge tracing, a machine learning approach that estimates what a student knows by analyzing their responses to exercises and predicting future performance. By combining this kind of learner modeling with the generative capabilities of LLMs, the digital human can adjust not only what it says but how it says it — simplifying an explanation for a struggling student, offering a deeper extension for an advanced one, or revisiting a prerequisite concept the model infers has not been mastered. The team’s related work on an intelligent interaction platform for personalized digital tutors, presented at the ACM Web Conference 2025, emphasizes empathetic and adaptive learning experiences, indicating that affective dimension — recognizing frustration and responding encouragingly — is treated as a design goal rather than an afterthought.

The technical architecture reflects a broader trend in applied artificial intelligence: rather than training a monolithic model from scratch, the system composes existing components. The LLM handles dialogue and reasoning; external domain knowledge can be injected to ground answers in course-specific material, addressing the well-known tendency of language models to hallucinate or drift from the curriculum; and the rendering pipeline converts textual responses into speech and synchronized facial animation. The authors also engage with the growing literature on readability, noting comparisons showing that LLM-generated educational content can approach the readability of human-written material — a prerequisite for a virtual tutor that students will actually want to listen to. The result is a pipeline in which a student’s spoken or typed question triggers comprehension of the query, retrieval or conditioning on relevant course knowledge, generation of a pedagogically appropriate answer, and delivery through an expressive virtual persona.

What makes the approach notable in the crowded field of educational AI is its positioning. Chatbots and LLM tutors already exist, and studies such as one on NewtBot, an LLM-as-tutor chatbot for secondary physics, have examined how students interact with them. Comparisons of perceived cognitive load across AI chatbots, pre-recorded videos, and live lectures suggest that each format carries distinct mental demands. The digital human system attempts to occupy a middle ground: it preserves the on-demand availability of recorded courses while layering on the responsive, conversational engagement of a live session. The authors frame this explicitly as bridging the gap between the flexibility of pre-recorded lessons and the instantaneous engagement typically associated with live teaching — a framing that speaks directly to the accessibility problems that motivated the work.

The researchers are candid about the challenges that accompany such systems. Their bibliography includes systematic reviews of the practical and ethical challenges of LLMs in education, surveys of LLM ethics, and analyses of the risks of deploying large language models across societal contexts. Accuracy, bias, over-reliance, and privacy all remain live concerns for any system that puts a generative model in front of learners, and the team’s own ethics statement notes that participant data was anonymized before statistical analysis. The funding acknowledgments — including grants from the National Natural Science Foundation of China and the Key Technologies R&D Program of Anhui Province — signal institutional backing for continued development, and the article’s early citation record suggests the work is already circulating among education-technology researchers.

The implications reach beyond any single course platform. If virtual instructors can reliably reproduce the most valuable feature of live teaching — immediate, personalized response to confusion — then the economics of one-to-one tutoring could shift dramatically, particularly for students in regions or circumstances where professional educators are scarce. The system also raises questions that the field is only beginning to grapple with: how much social presence a synthetic instructor should project, how to prevent students from forming misplaced attachments to virtual agents, and how to certify that an AI tutor’s explanations are pedagogically sound. For now, the study stands as a concrete demonstration that the components — language models, knowledge tracing, expressive avatars, and speech synthesis — have matured to the point where they can be assembled into a coherent teaching system. The recorded lecture, long the loneliest format in education, may soon come with someone to talk to.

Subject of Research: LLM-driven digital human tutoring systems for interactive pre-recorded online courses

Article Title: Advancements in Digital Humans for Recorded Courses: Enhancing Learning Experiences via Personalized Interaction

Article References: Advancements in Digital Humans for Recorded Courses: Enhancing Learning Experiences via Personalized Interaction. (n.d.). https://doi.org/10.1007/s44366-025-0072-9

Image Credits: AI Generated

DOI: 10.1007/s44366-025-0072-9

Keywords: digital humans, large language models, pre-recorded courses, online education, personalized learning, virtual instructors, knowledge tracing, talking-head generation, text-to-speech, educational technology, student motivation, human-computer interaction