This working paper documents the construction of MIRO (Multi-model Interreligious Roleplay Outputs), a dataset designed to be used in exploring the capacities of Large Language Models (LLMs) to produce theology — statements committing to a position within a religious tradition — rather than generic, balanced knowledge about religions. Four LLMs (OpenAI's GPT-4o and GPT-5, Anthropic's Claude Sonnet 4, and Google's Gemini 2.5 Flash) were queried both through their APIs and through their proprietary web interfaces. To elicit committed theological output, the models were prompted to roleplay twelve personas — six Christian and six Muslim — spanning a baseline lay believer, historically grounded individual figures, and tradition-specific scholarly or movement positions. All personas were asked the same question: to outline their view on other religions. Persona descriptions were kept deliberately minimal, so that the theological substance of each response can be attributed to the models' own elaboration from training data rather than to detailed prompting. The paper sets out the rationale behind the choice of models, channels, prompt design and personas, and reflects on the methodological trade-offs involved. MIRO is intended as a shared resource for interdisciplinary analysis of how generative AI mimics contemporary theological discourse, and of the patterns and biases such mimicry reveals.
Kamal et al. (Mon,) studied this question.