Tsinghua’s early AI-taught courses reveal promising advantages and notable challenges.

“What I really want is to harness lift to create my own flying racecar,” says a cheeky engineering student in response to a question about dealing with aerodynamics. “That’s called a jet!” chimes in a boisterous classmate. Their teacher draws the students back on track, saying, “Whether it’s a jet or self-driving cars, we need to learn today’s content to build anything.”

This might seem like an ordinary tangent in a classroom discussion. But here’s the twist: In this case, only the first ambitious engineering student is human. The instructor and the classmate are both created by artificial intelligence, and are known as AI agents. These AI agents form part of a Massive AI-empowered Course, or MAIC, spearheaded by an interdisciplinary team that includes computer scientists from Tsinghua University’s Department of Computer Science and Technology, as well as educational scholars from the Institute of Education.

Many teaching institutions are struggling with the challenge of students surreptitiously using large language models (LLMs) — the kind of deep-learning algorithms behind chatbots — to do their homework for them. But this team is using LLMs to enhance learning by creating personalized lessons that tailor the content and the pace of the lesson to the individual student’s needs, while also giving continued AI guidance to the student along the way.

MAICs, which are created using LLMs, assist teachers in designing courses, automatically generating teacher agents, teaching assistant agents, and peer agents with diverse styles. This enables the rapid creation of a personalized multi-agent course environment for each student, in which they can engage in bespoke online learning activities, guided by multiple AI agents.

Students engage better

Before MAICs, there were massive open online classes (MOOCs), which are web-enabled classes aimed at large-scale participation in terms of numbers of students. In addition to traditional course materials, such as filmed lectures, readings, and problem sets, many MOOCs provide interactions via user forums or social media discussions, as well as immediate feedback to quick quizzes and assignments.

MAICs are also able to be distributed widely, but they can personalize the teaching environment, which has benefits. “We are applying LLMs to radically improve tertiary education,” explains Liu Zhiyuan, a computer scientist who led the MAIC creation project at Tsinghua. As part of a new Tsinghua University AI-empowered education initiative, Liu’s Tsinghua team has tested two MAIC classes on hundreds of university students. They found that the students taught by AI had a class completion rate much higher than those taught traditionally on online learning MOOCs.

The MAICs also do not require advanced technical skills from either teachers or students, Liu adds, so the technology will make education more effective, accessible, and equitable. “It can reach students who might otherwise be deprived of opportunities, helping to bridge the educational gap,” he says.

MOOCs are already hugely popular, attracting more than 100 million learners globally and increasing access to education. But they require teachers to spend a lot of time recording lectures while lacking direct student-teacher engagement. “This model cannot achieve the educational effect of traditional classrooms,” explains Liu.

The Tsinghua team proposed MAICs as an interactive alternative that also encourages equality by improving access for people unable to attend in-person classes due to geographical limitations, work commitments, or health issues. Such MAICs would be cheap and quick to put together, once the platform is in place: A typical MOOC takes 60 hours for the teacher to record, based on an existing course, and all of its constituent parts can cost thousands of dollars per course; by contrast, a MAIC requires less than 30 minutes to upload course materials and the total cost can be as little as US$2.

Each MAIC has an AI instructor, an AI teaching assistant, and multiple AI classmates, with different personalities. While other AI-assisted programs offer limited support to students — a chatbot can answer simple questions, for instance — the developers say that MAICs are unique in that every aspect of the learning process, from teaching to answering students’ questions and facilitating discussions, is carefully conducted by the AI agents. A typical scenario involves the teacher AI dynamically adjusting instructional pace based on the student’s learning progress, deciding in real-time when to transition to the next topic during lectures.

At the same time, the teaching assistant AI and multiple AI classmates actively engage students with questions and foster discussions about the course material, thus creating an interactive and stimulating classroom environment that simulates the vibrancy of a traditional classroom. One advantage of this, says Liu, is that you’re engaging with a range of virtual peers who are deliberately designed to offer varied perspectives.

Massive AI-empowered Courses, or MAICs, use large language models to create agents that act as teachers and students in university level classes. These MAICs have been developed at Tsinghua by a team from the Department of Computer Science and Technology — led by Liu Zhiyuan (pictured) — as well as educational researchers from the university’s Institute of Education.

Meeting students where they are

The AI instructors do not just answer students’ queries but also analyze why a student might have posed that question and then adapt the course to their level and needs accordingly. The design of the instructor was modified based on student feedback, explains Yu Jifan, a computer scientist and education researcher at the Institute of Education at Tsinghua University, who co-led the study. For example, if the system identifies an opportunity to enhance student's higher-order thinking, it might subtly increase the participation of certain inquiry-driven peer agents to encourage deeper analysis and reflection.

As mentioned previously, the Tsinghua team analyzed 100,000 learning records from more than 700 students taking two university-level classes, one on the topic of AI and one teaching study skills, over three months. They also interviewed 20 students along the way. They found the AI-taught group slightly outperformed a human-taught control group in course tests.

“Students who were more engaged showed improvements in knowledge, higher-order thinking, and self-efficacy,” says Yu. The team speculates this could be because the AI agents stuck rigidly to course content, while human lecturers tend to digress. The AI agents are also encouraging, while human teachers are not always supportive.

But while test scores may have been good, there are ethical concerns about the adoption of LLMs in education, related to inaccuracy, discrimination, and security. One issue is that chatbots often ‘hallucinate’, confidently presenting incorrect statements as facts. To counter this, the MAIC’s AI instructors’ teaching output was regularly monitored by subject experts, and both students and teachers were able to flag obvious errors. LLMs also amplify biases inherent in their training data. To address this, the team attempted to weed out discriminatory data during the design stage, but admits these are ongoing challenges.

The team also encrypted student data for privacy. These ethical concerns are a priority, says Yu. “If there is a large amount of unsafe content, it is difficult to see how the system can provide equal educational opportunities for everyone,” he says.

Another worry is that students taught by AI might struggle with developing certain social skills. The team introduced the AI classmates to help, with some success, explains Liu. “AI classmates encourage students to participate in discussions as the students do not fear judgment from real peers,” he says. This confidence can then be carried over into real-world classrooms.

Yu Jifan, who has a Ph.D. in Computer Science from Tsinghua, is now an educational researcher at the Institute of Education at Tsinghua. He leads the implementation of Massive AI-empowered Courses at Tsinghua, which incorporate insights from an educational perspective.

Tsinghua’s pilot AI teachers

In 2023, Tsinghua University began piloting eight courses in science, engineering, and the social sciences using intelligent teaching assistants, which have been met with great enthusiasm from lecturers. Since then, computer scientists at Tsinghua University have rapidly explored two complementary ideas: developing more advanced AI capabilities for education, a project led by Wang Hongning and his team, and creating richer AI-empowered educational environments, a project led by Liu and Yu’s team. These efforts have provided both the confidence and capacity for future scaling and dissemination, says Liu.

The team plans to further scale up their MAIC at Tsinghua by expanding the number of participants to a few thousand people so they can conduct more comprehensive evaluations and upgrades and are preparing a series of new features, such as automatic knowledge profiling for students. “We are also continually recruiting educators who are interested in this educational model to participate,” says Yu.

The hope is to eventually expand to other universities across the country and worldwide. Scaling up should be easy, says Liu, since by design, LLMs can handle, and adapt to, many students simultaneously. Not every class can be converted to an MAIC, however. The system is best suited to introductory lessons and those that do not involve teaching hands-on skills.

That said, AI’s teaching capabilities are not without limitations. They may not be particularly adept at facilitating deeply subjective, abstract, and value laden questions, for example, they cannot really ‘understand’ the topics they teach. “AI tutors operate based on algorithms and data, lacking personal experiences, emotions and the intuitive insights that humans bring to the table,” says Liu. These include granular and multi-disciplinary knowledge, and more complex human analyses, such as the ability to quickly understand a students’ cognitive blind spots. Most of all: “AI cannot replicate the ability to inspire and motivate,” says Liu.

For these reasons, he says, AI-tutors will never replace human teachers. However, MAIC has the potential to enhance both the student and human teachers’ experiences.

Ideally, AI classes will remove the burden of teaching repetitive tasks, enabling human lecturers to focus on the creative and interpersonal aspects of teaching. “Human teachers can instead concentrate on cultivating students’ higher-order cognitive abilities,” says Yu. “I think the role of the human lecturer will become even more significant, not less.”

References:1. Yu, J. et al. From MOOC to MAIC: Reshaping Online Teaching and Learning through LLM-driven Agents, preprint: arXiv.2409.03512 (2024).2. ‘Tsinghua AI teaching assistant is here! Opening a new era of teaching’, Tsinghua University, 5 March 2024.

Editors: Ma Mingwei, Li Han