Choosing What to Practise With an AI Partner

Suppose you want to return an online purchase in another language. You could ask an AI partner to rehearse the conversation. But what should the practice help you do: remember the words for a refund, explain the problem or respond when the seller asks something unexpected? Each would give the session a different purpose.

A new Japanese-learning preprint offers one useful result. In a 40-person, single-session comparison, generated roleplay with the GELLI prototype beat goal-matched video practice by 0.85 points on a 10-point vocabulary test. Confidence was higher; speaking differences were not statistically reliable.

For someone choosing how to practise with AI, the useful starting point is a specific action rather than a general request to become ready. In the hypothetical return, being able to name the item and explain its fault would be a different target from choosing a reply to an unexpected refusal. You could choose either without asking one exercise to stand in for both.

Give the next attempt a job

The controlled comparison used preset goals, not learners’ later custom scenarios. It did not isolate components or test retention or real-conversation transfer. One GELLI learner reported a gym character giving information they had wanted to practise requesting.

To work on your imagined return, you could first practise with translations available, then try making the request without them. If the opening request is the target, that second attempt would ask whether you can produce it without looking up the words. If responding is the target, a partner could instead choose a reply you have not selected beforehand, leaving you to decide what to say next.

These are proposed checks, not demonstrated training methods. Their usefulness would depend on the goal. Remembering the opening phrase and handling an unexpected answer need not be assessed in the same attempt. Nor would a successful practice return establish that a real seller would accept it.

There is also a choice about what assistance to keep. If your real plan is to use a translation tool during the return, an unaided attempt would answer a different question from whether you can use that tool to make yourself understood. Removing every aid is not the definition of progress; it is one possible condition for an exercise. The condition should match what you want to find out.

A different kind of progress

An earlier Rehearsal study gave 40 people conflict-resolution instruction, some with added AI practice. An author unaware of their training group played their opponent in a new, 10-turn assessment. Practice users made about twice as many needs-focused responses and proposals as others; differences in remembering or identifying strategies were not reliable. This was not workplace follow-up.

That assessment put a person on the other side of a fresh exchange. It asked for something beyond remembering advice about how a conversation should go. The useful comparison with language practice is the choice of test, not a claim that the same teaching method produces the same effects across skills.

In an older 267-person negotiation-preparation experiment, a handbook outperformed a specialized AI coach on users’ sense of competence and control at work and usability—not demonstrated negotiation success. In separate interviews, researchers reported one participant’s preference for reading the handbook before repeated AI practice. This was a proposed sequence, not an evaluated improvement.

For the hypothetical purchase return, you might finish a session knowing how to begin but still wanting help with the follow-up. The next rehearsal could then start at the seller’s first unexpected answer, rather than repeat the opening you already know. That would give the AI a particular job—and give you a particular action to try afterward.

Sources

• Danilo Symonette and colleagues, GELLI, October 22, 2026 preprint.
• Omar Shaikh and colleagues, Rehearsal: Simulating Conflict to Teach Conflict Resolution, CHI 2024.
• Veda Duddu and colleagues, Does AI Coaching Prepare us for Workplace Negotiations?, September 2025 preprint.