VirtualSpeech: the machine talks - Building Socratic agents for reflection, skill building and professional development

As part of a short pilot Victoria Burns, Violet Chan, Giles Barber and Jorge Freire each designed and tested a distinct VirtualSpeech scenario, applying the platform to different educational contexts—from reflective coaching and research mentoring to clinical communication skills and assessment-related practice. Across the pilot, VirtualSpeech was used as a deliberately in-between layer: positioned after instruction but before assessment and framed as a low-stakes space for rehearsal, mentoring-style dialogue, and feedback rather than as a teaching or grading tool.

Victoria articulated this most clearly when contrasting VirtualSpeech with the media formats she typically produces—intros, micro-lectures, screencasts, interviews, and demonstrations. These formats are effective for content delivery and showing students how processes work, but they offer little space for learners to rehearse communication, think aloud, or test how clearly they are expressing ideas. In response, she designed a “Reflective Coaching Studio” linked to a Global Public Health assessment. Students were asked to submit a composite video consisting of a scientific presentation, a lay communication clip, and a recorded critical reflection. Alongside a ten-part DaVinci Resolve – a video and audio editor - tutorial series supporting the technical workflow, the VirtualSpeech scenario was positioned between production and submission, giving students a low-stakes opportunity to practise articulating their reflections, refine structure and clarity, and build confidence before recording the assessed version.

Violet applied VirtualSpeech to clinical communication roleplay on managing difficult patient interactions, building on earlier work with Bodyswaps. While she found the scenario-authoring interface less sophisticated, she judged the feedback to be more specific and more useful for communication skills practice. The browser-based delivery also reduced barriers to access, allowing students to practise anytime and anywhere without headsets or timetabling. At the same time, her testing exposed important constraints: the avatar’s refusal to provide a date of birth highlighted limits in compliance-sensitive, medically realistic interactions. This surfaced a key design boundary—GenAI roleplay can support rehearsal and emotional communication but should be treated cautiously in safety-critical or procedurally rigid contexts.

Giles deliberately used VirtualSpeech as a comparative stress test, recreating a Socratic “Professor Innovator” chatbot originally designed for research proposal mentoring. The avatar successfully followed a questioning brief—probing ideas without supplying answers—but the platform’s built-in emphasis on speech analytics and communication dashboards sometimes conflicted with the goal of conceptual inquiry. This revealed an affordance mismatch: VirtualSpeech excels when the learning objective is articulation and communication but is less well aligned when the primary aim is disciplinary knowledge construction or research design, as the affordances match the core goal of communication skill development and not necessarily subject domain teaching.

Jorge’s own exploratory scenario focused on dialogue with the VR agent to create the possibility of mentoring and leadership action planning within an online course. Here, the strongest learning mechanism mirrored what others observed: a loop of reflect → articulate → receive feedback → re-articulate. Speaking reflections aloud, prompted through open and follow-up questions, surfaced assumptions, gaps, and connections more effectively than writing alone. The experience felt constructive rather than evaluative, though it raised open questions about consistency across learners, analytics, and how far prompt design can be refined.

A short pilot, the affordances and potential of VirtualSpeech certainly merit further exploration.

And taken together, the pilot suggests that GenAI-enabled roleplay tools such as VirtualSpeech are most effective where learners need to practise articulation rather than acquire content: reflective assessment, professional and clinical communication, mentoring, and rehearsal. Their value lies not in replacing teaching or feedback, but in creating a pedagogically bounded space that supports thinking, dialogue, and confidence ahead of high-stakes interaction with staff or assessors.

This finding sits within a much longer tradition of dialogic learning. Sometimes we check maps; sometimes we ask for directions.

For centuries, dialogue has been one of the primary engines of human understanding, from everyday problem-solving to formal education. Socrates’ use of maieutic questioning and irony exemplifies this tradition: knowledge emerges through probing, contradiction, and shared inquiry rather than transmission alone.

In contemporary teaching, this dialogic tradition persists but must be deliberately designed. Different types of questions produce different kinds of thinking—memory, convergent, divergent, and evaluative questions activate distinct cognitive operations and learning outcomes (Ciardiello, 1998). In face-to-face settings, educators move fluidly between these question types. Online, however, higher-order learning does not arise automatically. Research on cognitive presence shows that it depends on structured, scaffolded discussions in authentic scenarios, supporting exploration, integration, and resolution (Darabi et al., 2011), and that the depth of student thinking closely tracks the quality of questions posed rather than instructor visibility alone (Belcher et al., 2015).

This long tradition now faces a rupture. Machines speak, simulate dialogue, and claim pedagogical presence. Generative AI and XR agents can model questioning and feedback, but they also blur boundaries between facilitation, authority, and relational engagement. As artificial voices enter educational dialogue, new possibilities emerge alongside significant epistemic and axiological risks (Freire, 2025).

Recent scholarship positions Socratic GenAI as a promising mechanism for learning by discussion, creative thinking, and mentorship when aligned with dialogic learning principles. Contributions in Teaching and Learning in the Generative Artificial Intelligence Age emphasise that GenAI’s educational value lies not in answer generation, but in its capacity to sustain exploratory dialogue and make learner thinking visible over time (Sampson, Isaías and Ifenthaler, 2026). Empirical evidence from Lee et al. (2025) supports this view, showing that GenAI systems designed around Socratic questioning can enhance problem-solving, critical thinking, and reflective practice.

Read through this lens, the VirtualSpeech experimentation does not point to AI tutors or automated feedback replacing educators. Instead, it suggests a more modest—and more powerful—role: carefully designed GenAI as a conversational companion that helps learners rehearse thinking, articulate ideas, and enter assessment and supervision better prepared for meaningful human dialogue.

 

References (Harvard)

Belcher, A., Hall, B.M., Kelley, K. and Pressey, K.L. (2015). An Analysis of Faculty Promotion of Critical Thinking and Peer Interaction within Threaded Discussions. Online Learning, 19(4). doi: https://doi.org/10.24059/olj.v19i4.544

Ciardiello, A.V. (1998). Did You Ask a Good Question Today? Alternative Cognitive and Metacognitive Strategies. Journal of Adolescent & Adult Literacy, [online] 42(3), pp.210–219. Available at: https://www.jstor.org/stable/40014681

Darabi, A., Arrastia, M.C., Nelson, D.W., Cornille, T. and Liang, X. (2010). Cognitive presence in asynchronous online learning: a comparison of four discussion strategies. Journal of Computer Assisted Learning, 27(3), pp.216–227. doi: https://doi.org/10.1111/j.1365-2729.2010.00392.x

Freire, J. (2025). Digital learning, now with robot voices, chatbots, and AI girlfriends. [online] Blogspot.com. Available at: https://jorgefreireld.blogspot.com/2025/12/robot-voices-chatbots-and-ai-girlfriends.html [Accessed 26 Jan. 2026].

Sampson, D.G., Isaías, P. and Ifenthaler, D. (2026). Teaching and Learning in the Generative Artificial Intelligence Age. Springer Nature.

Lee, J., Hung, J.-T., Yilmaz, S.M., Popescu, D., Cui, C.Z., Grigoryan, G., Joyner, D.A. and Harmon, S.W. (2025). Socratic Mind: Impact of a Novel GenAI-Powered Assessment Tool on Student Learning and Higher-Order Thinking. arXiv (Cornell University). doi: https://doi.org/10.48550/arxiv.2509.16262