Large language models are trained to model statistical regularities in web scale corpora. They tend to reproduce the categories most prevalent in those data, including reductive framings of public life that equate democracy with choice and procedure rather than shared authorship. Alignment methods such as supervised fine tuning and human feedback can guide behavior after pretraining, yet they begin after the representational substrate is already set. This concept paper proposes an alternative starting point. We outline a Civic LLM that uses an explicit civic ontology during corpus design and fine tuning. The ontology names roles, institutions, practices, and values that thicken democratic life. Our claim is deliberately modest. Curated data that is organized and weighted by a civic schema may produce small but measurable shifts in default framings without degrading general capability. The contribution is twofold. We integrate political theory as a design input, and we show that adjacent literatures in knowledge enhanced language models and concept level steering support a cautious path for evaluation. The aim is to establish a clear, citable origin for the idea and a disciplined agenda for future research.
Ceresa et al. (2025) studied this question.