How AI Avatars Are Transforming Language Learning in Higher Education

Théry Béord, French Lecturer at the School of Arts and Humanities, and Head of the Center for Excellence in Teaching and Learning at Sorbonne University Abu Dhabi, answered questions on the university’s innovative AI-powered avatar initiative, developed to support learning during online learning, how it is helping students strengthen speaking skills through interactive conversations and personalised feedback, and what it signals for the future of AI in higher education. 

Challenges of oral production skills in fully remote learning & limits of traditional online platforms

One of the main challenges we faced was how to maintain meaningful oral language practice in a fully online setting. While platforms such as Blackboard Ultra are effective for delivering content and managing coursework, they are not designed to support sustained, interactive speaking tasks. 

In the intensive French Foundation program at Sorbonne University Abu Dhabi, students attend a weekly 90-minute oral production class dedicated to developing this skill. Each week, they complete a structured activity based on a stimulus composed of a visual document and a short text on a contemporary issue linked to what has been studied in their other classes. They are expected to describe and analyse the material, present a critical perspective, express and justify their opinion, and engage in a debate with the instructor. Each interaction typically lasts approximately 15 minutes. 

This is a demanding exercise, even in face-to-face settings. It requires not only linguistic competence, but also the ability to organise ideas, structure discourse, and engage in critical thinking across a range of topics. 

Moving this type of activity online made things more complex. Even in small groups, facilitating such interactions is challenging for instructors. More importantly, students have limited opportunities to practise speaking independently outside of class, and it is difficult to recreate, through traditional tools, the conditions of real-time interaction that are essential to oral skill development. 

This gap highlights the need for solutions that can extend oral practice beyond the virtual classroom while maintaining the interactive nature of spoken communication. 

AI-powered avatars for spoken interaction, debate & spontaneity in online learning

To address this challenge, we worked (and continue to work very closely) with VRAI Learning, an AI learning platform, to develop AI-driven avatars capable of supporting oral practice in a more interactive way. 

Each week, a new avatar is specifically trained on the topic associated with the students’ task. This ensures that learners engage in focused, meaningful discussions rather than generic conversations. 

The avatars simulate the role of an examiner or interlocutor. They listen to the student’s analysis, respond to their arguments, ask follow-up questions, and challenge their position to sustain a genuine debate. They are also visually realistic, taking the form of human-like characters—male or female—which, combined with the coherence and relevance of their linguistic responses, contributes to making the interaction feel natural and engaging for students. 

This creates a form of guided interaction that encourages spontaneity while maintaining a clear pedagogical structure. 

A key aspect of the project is the continuous training and refinement of the models. We collaborate on a daily basis with VRAI Learning to adjust the avatars’ behaviour, ensuring that their responses remain coherent, relevant, and aligned with the expectations of the task. 

Practical use of the tool: student sessions, preparation & feedback

In practical terms, students install an application on the device of their choice (phone, tablet, or computer). Each week, they access a new AI avatar trained on the topic of the week. 

The activity mirrors classroom practice. Students are first given a stimulus consisting of a visual document and a short paragraph introducing a contemporary issue. After a 30-minute preparation phase, they begin the interaction with the avatar. 

The task starts with a short uninterrupted monologue, during which the student presents their analysis, followed by a debate phase. During the discussion, the avatar reacts to the student’s arguments, asks questions, and introduces counterpoints to sustain the exchange. The interaction typically lasts between 5 and 10 minutes. 

At the end of the exchange, the avatar switches to English and provides detailed feedback based on clearly defined criteria: organisation and development of ideas, discourse structure, vocabulary, grammar, and pronunciation. The model is trained to assess each criterion individually and provide concrete recommendations—for example, suggesting connectors or expressions to improve the structuring of arguments. 

For example, in a session on the impact of social media, a student may analyse an image illustrating changing communication habits, supported by a short explanatory text. The avatar may then challenge the student’s position by introducing an alternative perspective, prompting them to react, justify their opinion, and refine their argumentation. 

On the instructor’s side, all interactions are accessible through a dedicated platform. Teachers can review the full transcript, the AI-generated evaluation, and the feedback provided. This report can be shared with students, adjusted if needed, and used as a basis for follow-up and progression. 

Student response: confidence, participation, motivation & speaking development

The pilot phase of this initiative has been conducted by two instructors working with small groups, involving approximately forty students in total. This has allowed us to test the tool in real teaching conditions while maintaining close pedagogical follow-up. 

Early student feedback has been very positive, particularly in relation to confidence, participation, and willingness to speak. Many students highlighted that the tool creates a low-pressure environment, allowing them to practise speaking “without feeling judged or embarrassed,” which is essential for building confidence. 

Ease of use also plays an important role. Students reported that the application is intuitive and easy to navigate, enabling them to focus on the task itself. This supports regular engagement and sustained participation. 

Another strong point in the feedback is the perceived value of the AI-generated feedback. Students described it as clear, constructive, and helpful in identifying areas for improvement. The fact that feedback is immediate and personalised appears to support more focused and effective learning. Some students also appreciated the flexibility of the interaction, noting that the system can accommodate moments of hesitation or misunderstanding. 

Strengthening teaching and learning through data, accessibility, collaboration, AI & innovation

A key strength of the initiative lies in its continuous improvement model. After each session, we systematically collect student feedback and analyse the full transcripts of interactions, along with the AI-generated evaluations available on the instructor platform. This provides us with detailed insight into how students engage with the tasks, the types of difficulties they encounter, and the quality of the feedback produced. In close, ongoing collaboration with VRAI Learning, we use this data on a weekly basis to refine and retrain the avatars. Adjustments are made to improve the relevance of responses, the coherence of interactions, and the precision of the feedback. This iterative, data-driven process ensures that the tool remains closely aligned with our pedagogical expectations.  

What was initially conceived as an emergency solution, developed in response to constraints unrelated to pedagogy, now shows strong potential to become a lasting component of our teaching practices, even after a full return to campus. 

You can download the reference video here.