Abstract:

This paper reports on an exploratory practitioner-research study investigating the use of a voice-enabled Generative Artificial Intelligence (GenAI) agent as a mock interviewer in professional communication training for undergraduate engineering students. While GenAI roleplay tools are increasingly being adopted in education, empirical evidence on their pedagogical value in spoken professional contexts remains limited, particularly where effective performance depends on sustained, context-sensitive interaction. This study examines both the affordances and limitations of such a tool, with a specific focus on positivity bias as a constraint on pedagogical effectiveness.

The study was conducted within an English for Specific Purposes (ESP) module at an Austrian university of applied sciences. Following instruction in job interview techniques, with emphasis on behavioural interviewing, students completed individual mock interviews with a custom-designed voice-enabled GenAI chatbot configured as a domain-informed interviewer with knowledge of relevant course content and the students' engineering context.

Data were collected through a post-activity questionnaire (n=28) and thematic analysis of interview transcripts from consenting participants (n=4). Questionnaire findings indicate broadly positive student perceptions: participants reported high levels of comfort, agreed that the activity supported practice in spoken English and interview skills, and expressed interest in further use of similar tools. However, a prominent theme across both data sources was a pervasive positivity bias in the chatbot's interaction behaviour. The agent frequently produced positive evaluative language that appeared responsive but did not consistently engage with the substance of students' answers, praising weak or repetitive responses without challenge. Transcript analysis corroborated this finding: although the chatbot structured interviews using appropriate question types, it failed to probe underdeveloped answers or recognise when candidates reused examples across multiple questions. The voice interface further introduced turn-taking disruptions, with the agent frequently interjecting before students had completed their responses.

These findings suggest that voice-enabled GenAI agents can offer clear benefits for low-stakes professional skills practice, particularly in terms of student comfort and accessibility. At the same time, current systems may reproduce the surface structure of behavioural interviews without delivering the evaluative challenge that makes such practice pedagogically effective. The paper concludes by proposing prompt- and system-design strategies to mitigate positivity bias and discussing implications for the integration of GenAI tools into professional communication curricula.

Keywords:

GenAI in education, voice-enabled chatbots, positivity bias, English for Specific Purposes (ESP).