For this literature review research project, I chose to work on the topic of Artificial Intelligence (AI), Adaptive and Personalized Learning, which is one of the main areas our Advanced Technology for Learning class focuses on. Indeed, AI, and more especially Gen AI, have quickly become both a fascinating and concerning tool for education. I believe it is one of our mission as instructors to learn what are the drawbacks and the advantages of this kind of technology in order to be able to use them in the best possible way and stay relevant in front of our students - and the general public, who is currently discovering new AI technologies everyday.

     As a Ph.D. student in French SLATE (Second Language Acquisition and Teaching Education) and a teaching assistant of French at the university level, my goal is to find a way to make language learning as easy, efficient and entertaining as possible for my students. I am particularly interested in new technologies, including AI and video games, and how instructors can use them as an effective tool to teach French as a foreign language. Indeed, I have previously coded a short narrative video game and conducted a pilot study using Microsoft Copilot with some of my students as part of a pilot study to help guide my choice of dissertation topic. I have also collaborated with colleagues on an innovative assessment project aimed at modernizing oral exams in the French department of our institution.While evolving Gen AI chatbots may pose a threat to language education by enabling instant translation, I believe they can also serve as valuable tools to support learning. 

This research topic explores the influence of voice-based generative AI on the speaking engagement level of second-language learners of French when participating in a roleplay-style dialogue. It also aims to assess the extent to which Microsoft Copilot’s AI is technically well-suited for facilitating roleplay activities for learners of French as a foreign language. This research article also focuses on the benefits of using new generative AI technologies for the oral skills of language learners.

    In order to explore this topic effectively, I will explore scholarly sources, such as research literature articles found on online databases (ResearchGate, ERIC, Google Scholar, etc.). Certain books written by recognized scholars working on the topic of language learning will also be used to address the theoretical framework of this work project. Finally, both the process and results of one of my own research work in progress about Microsoft Copilot’s AI (Petre, 2025) will be used in order to bring some information about concrete applications and venues for future research.

     This topic is relevant to advancing research in SLATE given the recent transformation of language teaching methods with the widespread integration of generative AI. Indeed, tools such as ChatGPT or Deepseek now allow users to engage in oral communication with a virtual interlocutor in any language, at any time. While some studies have begun to examine the benefits of such tools in language classrooms (Huang, 2024; Muthmainnah, 2024; Qiu, 2024), the topic remains underexplored in the context of teaching French as a second language through AI-based video games and roleplay. Research in SLATE has demonstrated that the use of roleplay has a positive impact on students’ oral skills and engagement (Vanisree, 2024). However, no study has yet examined the potential impact of combining Gen voice-based AI and roleplay to enhance students' speaking and listening skills in French. Language learning is undergoing a true technological revolution, and it is crucial for L2 instructors to adapt their pedagogy to the new tools made available to the general public. 

     It is still very interesting and useful to analyze the voice-based Gen AI topic through the eye of language theories, even if they were developed even before the democratization of the World Wide Web. Most language learning theories are still valuable to this day, despite the technological revolution and the incredible progress made since the time scholars theorized their ideas.

     Three main educational theories are relevant to this topic about Gen AI, roleplay and oral skills in SLA: Vygotsky’s (1978) sociocultural theory, Van Lier’s (2010) ecological theory, and Csikszentmihalyi’s (1990) Flow Theory.

     This work was first framed using Vygotsky’s (1978) sociocultural theory, which posits that language development primarily depends on social interactions. Vygotsky proposed that social interaction and cultural context are fundamental to cognitive development. He emphasized oral dialogue and social exchanges as key to knowledge construction. Learning occurs first on the interpersonal level (between people), then on the intrapersonal level (within the individual). Through conversation and collaborative discourse, learners acquire new concepts and refine their understanding. According to this theory, tools (linguistic, cultural, or technological) mediate thinking and learning. Language is considered the most powerful tool for shaping thought. Learning is facilitated by someone more skilled—e.g., a teacher, peer, or even a tool like AI—that helps the learner move forward. Research within the sociocultural paradigm is particularly concerned with language practice in practical circumstances, allowing students to apply their knowledge in concrete situations by engaging in simulated conversations (Huang, 2024). Vygotsky’s sociocultural theory also experienced come criticisms, such as the lack of specificity (especially concerning the concept of Zone of Proximal Development, the “sweet spot” between what a learner can do alone and what they can do with help, which is hard to measure). Some criticisms say Vygotsky downplays individual cognition or overemphasizes culture and interaction (Daniels, 2001). Critics from Piagetian or cognitive traditions also argue he underrepresents biological development (Piaget, 1962). Despite those criticisms, Vygotsky’s educational theory still connects with my topic through several aspects. First, as mentioned above, the Gen AI could be considered as a “More Knowledgeable Other” (KMO) as describes by Vygotsky. Microsoft Copilot’s AI plays the role of an KMO by providing scaffolding during language roleplays. It offers corrective feedback, models correct speech, and can adapt to the learner’s level—mimicking the role of a tutor or peer. Secondly, the AI could be used to target the Zone of Proximal Development. Indeed, an AI can be programmed to recognize a learner’s current level and tailor prompts, vocabulary, and scenarios that push the learner just beyond their comfort zone, but not so far that they’re lost. This is precisely where Vygotsky said learning happens. Plus, using spoken interaction with AI mirrors this theory’s claim that language mediates thought. The voice-based format is especially powerful, as learners internalize language structures through oral use, which reinforces both comprehension and production. Even though the interaction is with a machine, it simulates authentic social exchanges. This aligns with Vygotsky’s emphasis on dialogue and social interaction as necessary for language development. Finally, AI such as Copilot can integrate cultural elements into roleplay (e.g., greetings, politeness norms, real-life scenarios), which supports Vygotsky’s idea that cultural tools shape learning and language use. For example, Copilot's AI manages to reproduce the natural flow, voice, and tone of a native speaker, and also adapts the vocabulary he uses according to the context and to the person using it. Considering the information it uses comes from an infinite Open database, it is able to simulate a conversation about very niche hobbies or the latest news or trends on the internet, and can use a very sophisticated or a very relaxed/slang vocabulary. It is even able to use different dialects or language variations and accents (Metropolitan French, but also Quebec French, Belgian French...). Those are examples of interesting cultural elements students have the chance to explore through this tool.

     The second main educational theory I used to frame this work about Gen AI and oral skills is Van Lier’s (2010) ecological approach. Van Lier’s approach draws inspiration from ecological psychology, especially Gibson, and sociocultural theory, like Vygotsky. He argued that language learning is embedded in a rich, dynamic, and interactive environment—a language ecology. One of the central concepts of this theory is the concept of affordances. An affordance is a possibility for action provided by the environment. In language learning, affordances are opportunities to notice, interpret, and use language in meaningful ways (e.g., a conversation, a sign…). According to this approach, language is learned through meaningful engagement instead of passive exposure. Learners are not passive recipients of input. They are active agents who explore, choose, and construct meaning. In other words, participation in authentic communicative events (conversations, storytelling, role-play) enhances learning. Context and interaction are crucial in Van Lier’s ecological theory. Indeed, language learning is situated and context-sensitive. The social, cultural, and physical environment shapes the learning process. Interaction is key—both with people and with tools or media (e.g., AI, books, real-life tasks). Identity, motivation and emotions play a role in how learners engage with their linguistic environment. Learning is tied to how learners perceive and experience the world around them. Van Lier emphasizes the embodied nature of language use—speaking, listening, and interacting involve the full person in context. His theory was criticized by other scholars for being to conceptually ambiguous (terms like affordance, ecology, and emergence can be hard to define clearly, making empirical research difficult) and too broad or idealistic, especially in institutional settings where flexibility and rich affordances may be limited. Certain critics from more cognitive or SLA-oriented backgrounds also argue it downplays grammatical instruction and the role of deliberate practice. Despite those critics, Van Lier’s theory is well-adapted to research and studies on AI-based conversation tasks. Indeed, the AI creates affordances: it gives learners opportunities to speak, react, negotiate meaning, reformulate, and engage in realistic oral exchanges.
Learners choose to engage, and through that engagement, learning happens. The AI can also offer multiple paths or dialogue options, letting learners direct the flow of conversation. This supports agency, a key tenet of Van Lier’s framework. Moreover, AI-based roleplays can simulate real-life contexts (e.g., ordering food, going to a job interview, chatting with a friend). This situated use of language promotes deeper, contextualized learning—core to the ecological approach. Even though the interaction is with AI, learners are immersed in a dynamic linguistic environment where they receive feedback, adapt, and respond. The AI can then be seen as a semiotic and interactive environment in which language practices emerge.

     Finally, in recent years, both qualitative and quantitative research has focused on the potential of Generative AI chatbots in supporting oral skills practice for engaging in foreign language conversations (Lorentzen & Bonner, 2023; Phuong, 2024; Qiu, 2024). Such studies are based on Csikszentmihalyi’s (1990) Flow Theory, which is the third main theory used to frame my work. According to this theory, flow is a psychological state of deep focus, enjoyment, and immersion in an activity. Csikszentmihalyi developed this theory to explain how people experience optimal engagement—when they are “in the zone.” Although originally developed in the context of creativity, sports, and leisure, it has become highly influential in education and language learning. Flow occurs when the challenge of the task matches the learner’s skills—not too easy (boredom), not too hard (anxiety), but just right. In other words, it requires a challenge-skill balance. Moreover, learners need to see the results of their actions right away to adjust and stay engaged. They must feel in control of their actions, even if the task is challenging. Flow is rewarding in itself—the activity is done for its own sake, not just for external rewards. Some critics argued that flow is hard to define and measure, since it is subjective, and different people experience it differently. It is also difficult to guarantee or reproduce flow in educational settings. Moreover, effective learning often involves frustration, confusion, or discomfort, which are not associated with flow. However, despite those criticisms, concrete implementations of AI in foreign-language education have been reviewed and illustrated the role of AI in helping learners to master pronunciation (Almelhes, 2023). Following the gist of this theory, if students perceive AI-mediated roleplay as an immersive and intrinsically motivating experience, they may enter a ‘flow’ state, characterized by deep focus and enjoyment, leading to improved oral production and sustained motivation in language learning. The AI can adapt its prompts and difficulty to each learner’s oral proficiency level. As a reminder, in the context of AI, especially with generative models like ChatGPT or Microsoft Copilot, a prompt is the name given to any input (text, image, audio, etc.) given to the model to elicit a specific output. It serves as a starting point or instruction that guides the AI's response. In text-based interactions, a prompt can be a question, a command, or a scenario.

This ‘flow’ state helps keep learners within their “flow zone”: not bored, not overwhelmed. Using AI-based roleplay conversations also allow scenarios to be designed with clear communicative goals (e.g., “Order a coffee,” “Convince a friend to go out”). This clarity helps learners focus and stay engaged as far as they can choose the type of scenario, level, or even tone of the conversation. Finally, roleplay can be fun and rewarding in itself—especially when learners experience a sense of achievement in communicating effectively in a foreign language. AI’s storytelling can enhance motivation and joy.

     Generative AI chatbots equipped with Automatic Speech Recognition (ASR), the most well-known of which is ChatGPT, have become a prolific research topic in Second Language Acquisition (SLA) over the past few years. While they can be used to practice reading and writing skills, recent research has also focused on oral skills. Studies have demonstrated that engaging in a conversation with a generative AI can lead to significant improvements in proficiency (Phuong, 2024), including enhanced pronunciation, fluency, and rhythm accuracy (Lopez-Minotta et al. 2025), as well as better language retention through personalized feedback (Suhail, 2024). Similarly, speaking with a virtual rather than a real interlocutor has been shown to reduce anxiety levels (Lopez-Minotta et al. 2025; Muthmainnah, 2024) and increase motivation (Anjum et al. 2025; Mavropoulou & Arvanitis, 2024). Qiu (2024) also demonstrated the use of ChatGPT can enhance willingness to communicate with their study conducted on nine Chinese students. By providing shy learners with a creative, encouraging and non-judgmental interlocutor who pushes the student to engage in a conversation as much as they can, AI then helps foreign language students to master their oral skills.

     Interacting with AI has also been linked to improvements in speaking engagement. In their mixed methods study on EFL learners, Huang (2024) established that there was a significant correlation between speaking performance and behavioral, cognitive, and emotional engagement as well as a significant role of AI affordances on speaking engagement as determined via structural equation modelling (Huang et al. 2025). Some qualitative studies based on students’ and teachers’ reflection journals even advocate for an AI-driven classroom rather than a direct instruction-focused learning environment (Asrifan & Dewi, 2024).

     However, several gaps in research remain. First, while SLA studies on ChatGPT are numerous, research on other voice-based AI platforms remains scarce. Additionally, ChatGPT’s dominance is accompanied by a focus on English as a Second Language (ESL), with AI-assisted language learning being particularly studied in Asia. The teaching of certain languages, including French, has yet received little to no research attention. Finally, although most studies highlight the benefits of generative AI, there is a noticeable lack of research on its potential limitations. Issues such as translation inaccuracies, transcription errors in students’ responses, and technical problems (Lorentzen & Bonner, 2023) are rarely mentioned. This study thus aims to fill these research gaps by examining how oral conversations with Microsoft Copilot, a relatively understudied AI, can influence the speaking engagement of students learning French as a second language, whether positively or negatively, while also highlighting specific constraints inherent to conversing with a virtual interlocutor.

     What about Gen AI chatbots as a tool for Roleplay in SLA? Among the popular teaching strategies to increase engagement, gamification is increasingly associated with AI (Saptiany et al. 2024) and roleplay. Research showed that AI may very well fit this type of activity by offering an infinite number of personalized scenarios to those who know how to use the appropriate prompts thanks to various open-source Large Language Model (LLM) platforms (Surve & Ghatule, 2024), thus maximizing student engagement by creating an “immersive environment” (Mohammed et al. 2024, p. 57) on which the student has control. In order for a prompt to be 'appropriate' for roleplay-based second language acquisition purposes, it should meet several criteria such as matching the learner proficiency level, mentionning not using grammar or vocabulary too easy or too difficult, or providing clear and specific instructions in order to avoid ambiguity. It should also be skill-focused, meaningful, engaging and culturally relevant.

     However, research on the use of generative AI for roleplay in SLA remains scarce, particularly concerning the analysis of oral skills. Most studies on AI and its possible gamification dimension in SLA focus primarily on how to gamify AI-driven lessons, such as Duolingo’s reward systems featuring points or badges (Surve & Ghatule, 2024). The exercise of building a fictional story through oral interaction with generative AI, along with its potential benefits, remains largely unexplored. Similarly, little research has been conducted on the possible challenges associated with this type of pedagogical activity. One exception is Saeki et al.’s (2024) study findings which mentions AI’s difficulty in following a reasonably long and structured prompt and maintaining logical coherence in its answers. This study therefore aims to provide insights into both the advantages and challenges of integrating AI-mediated roleplay into oral language practice.

In light of the rapid advances in generative AI and its growing use in Second Language Acquisition, some of the studies mentioned earlier in this work would merit replication on a larger scale in order to lend greater credibility to the results obtained.

Speaking of my experience, I conducted some mixed-methods research on a short sample of French intermediate students (7 participants) at a university level. After giving the same pre-pared prompt to the AI, students were left alone and had to converse for 4 minutes with Microsoft Copilot’AI doing a roleplay. They then had to share the recordings of the conversation with me and completed a 16 questions post-conversation survey about their experience with the AI. The results I obtained for this work in progress pilot study aligned with emerging research on voice-based gen AI previously mentioned. However, as I stated above, research and practical experiments on this specific topic is still very scarce, if not completely inexistent. Reproducing this type of specifically adapted study design with longer conversations or different prompts would be a possible solution to enrich the research on AI’s adaptability for conducting roleplays in a foreign language. It would also be valuable to repeat this type of task multiple times within a longitudinal study to examine its long-term benefits for oral skills. Future researchers might also find it relevant to replicate this type of study with learners at different levels of French proficiency. This would make it possible to assess whether the AI can successfully carry out a roleplay conversation suited to novice learners, or whether it is capable of sufficiently challenging students with more advanced proficiency in French. Other romance languages, such as Spanish or Italian, could also be tested.

Finally, as stated before, one of the biggest challenge of this topic is that Microsoft Copilot’s AI is a tool that has been too rarely investigated so far as compared to ChatGPT or Deepseek. Although this study focuses solely on Microsoft Copilot, new AIs are being developed every week, with continual improvements in their oral communication abilities. Reproducing this type of study with other voice-based AIs—whether mainstream tools such as ChatGPT, Deepseek, or Gemini, or AIs specialized in language learning—would offer valuable comparisons for both academic research and industry, especially regarding usability concerns. It is likely that in the coming years—or even months—this type of conversation-based task will become increasingly common in roleplay-style games designed for foreign language practice and powered by AI. This phenomenon can already be observed to some extent in games, although these are not focused on language learning.

To date, the only existing research on the use of AI-driven oral roleplay to improve students’ speaking skills is, in fact, my own. In this section, I will therefore connect the results of my pilot study (Petre, 2025) to the three theoretical frameworks discussed above. Overall, the results of my study aligned with emerging research on voice-based generative AI in foreign language learning and provided a meaningful illustration of the hypotheses proposed by Vygotsky’s sociocultural theory, Van Lier’s ecological perspective, and Csikszentmihalyi’s Flow theory. While none of these theories explicitly addresses AI, my interpretation suggests that AI-based roleplay conversations appear to foster the kinds of learning benefits described by these frameworks. However, it is important to note that this area of research remains largely unexplored, and it is still too recent for theory-driven limitations to have been fully considered by major scholars in language education. Based on my observations while conducting this study, the primary limitations seem to arise in the way AI supports Vygotsky’s Zone of Proximal Development (ZPD). Although the AI is designed to adapt to the learner's level, some participants reported a lack of sufficient challenge. Similarly, while most students described the experience as engaging and enjoyable, some technical issues related to usability—such as problems with voice detection or the AI not following the prompt accurately—represent new limitations that existing theories have yet to address.