AI Role-Play Partner
Students prompt an LLM (ChatGPT, Claude, Gemini) to play a specific role — a hotel receptionist, a job interviewer, a doctor, a hostile debater. They carry out a full interaction in the target language, then the AI gives feedback on their production. Scales 1-to-1 conversational practice indefinitely.
Procedure
- Design a prompt scaffold (teacher-provided, the first few times):
Play the role of a [ROLE]. I will be [ROLE]. Respond naturally, stay in character, and ask follow-up questions. After I say "END", give me feedback on my language: what was accurate, what sounded unnatural, and 2 specific suggestions for improvement.
- Students pick (or are assigned) a scenario: booking a flight / arguing a refund / interviewing for a job / asking for medical advice.
- They copy the scaffold, fill in the roles, and paste to the AI.
- 10–15 minutes of interaction — typed or spoken (voice mode).
- Say END. The AI produces feedback.
- Students pick one piece of feedback to work on. Redo the role-play with that focus.
Why It Works
- Unlimited patience: the AI tolerates slow typing, beginner errors, repetitions without fatigue.
- Pressure-free experimentation: students try risky phrases they wouldn't risk with a human partner.
- Focused feedback: the feedback prompt trains the AI to attend to language, not just content.
- Accessible outside class: students continue practice at home; no scheduling with partners.
Good Scaffolds
Job Interview
You are an interviewer for a marketing assistant role at a Hanoi tech company. Ask me 5 realistic questions. Push me with follow-ups. After "END", evaluate my responses for clarity, grammar, and vocabulary fit for this register.
Customer Complaint
You are a customer service agent. I will complain about a faulty product. Be polite but initially reluctant to refund. Make me justify my complaint in English. Feedback at "END".
Academic Discussion
Play the role of a critical university tutor. I will present an argument about [TOPIC]. Challenge my reasoning with probing questions. At "END", give me 3 specific improvements for academic register.
Variations
- Pair scaffolding: pairs design a scaffold for another pair; swap prompts; each practises the other's scenario.
- Silent observer: one student role-plays; partner watches and notes the AI's and student's language. Debriefs after.
- Pre- and post-recording: same scenario at week 1 and week 6. Students compare to see progress.
- Group-designed character: class collaboratively designs a character, then each student has a solo session with that character.
Tips and Cautions
- LLMs hallucinate. Warn students: the AI may invent incorrect "grammar rules" or strange feedback. Verify with the teacher when unsure.
- Teach prompt refinement: if the AI is too polite, too verbose, too easy, the student should ask it to adjust.
- Protect privacy: no personal information (real names, addresses, payment details) should go into the AI.
- Use as supplement, not replacement. Students need human audiences too.
Source
Warschauer, M., Tseng, W., Yim, S., Webster, T. (2023) The affordances and contradictions of AI-generated text for L2 writers. Journal of Second Language Writing. Kohnke, Moorhouse & Zou (2023) on ChatGPT in ELT.