Meraki: Encouraging Language Learning in Real-World Simulations Through AI Para-Social Relationship Building

Meraki is a prototype language learning application that teaches Hindi through a 3D theatre production simulation, in which an AI voice assistant builds para-social relationships with learners and draws on Self-Determination Theory to support motivation. The work positions itself as a critique of reward-based language apps, arguing that learners benefit more from autonomy-supportive relationships with an AI companion than from points and badges. Its central proposition is that such relationships can strengthen memory encoding and retrieval, but only when the learner experiences them as positive.

The evidence presented is preliminary and primarily qualitative. An initial Wizard of Oz evaluation with five participants suggested that the design achieved a reasonable balance among autonomy, competence, and relatedness. A subsequent study compared four AI assistant conditions, each aligned with a different regulation style: external, introjected, identified, and integrated. The integrated condition produced the most favourable indicators, including perceived intimacy, emotional attachment, behavioural influence, and the longest engagement among high-engagement groups at 2.03 hours. By contrast, the external and introjected conditions revealed power imbalances and negative or unrealistic effects on learners' self-concept.

The authors conclude that pairing para-social relationship building with self-determination theory can foster learner agency and motivation, and they report that learners who formed positive relationships recalled and applied language knowledge more effectively, with curiosity and excitement supporting retention. They also caution that heavy reliance on AI feedback may erode real-world communication skills and that overly reverent AI relationships risk imposing ideologies and stereotypes without encouraging critical thinking. The study's small sample sizes and absence of quantitative learning outcomes limit the strength of these conclusions, so its findings are best read as a design-oriented proposal in need of larger, controlled validation.