Constitutional Legitimacy, Conceptual Completion, and Prototype Effects in Frontier Language Models

Debates about political bias in large language models (LLMs) typically focus on policy positions or ideological tone. This study examines a deeper layer: how AI systems interpret the meaning of political legitimacy itself. Political theory contains multiple established legitimacy frameworks, including those grounded in legal-institutional continuity and those grounded in moral justification to persons as political equals. Using structured probes across frontier LLMs, we find that models can describe and reason within both frameworks when definitions are provided. However, when legitimacy is left unspecified, models frequently treat legal validity as necessary while introducing participation, rights, or justification to persons as conditions for full legitimacy. This pattern indicates a shared interpretive prior at the level of concept completion: a form of epistemological tilt that shapes how political authority becomes intelligible in AI-mediated reasoning. The result has practical implications for AI use in constitutional contexts where legitimacy is derived from institutional continuity and liberty is residual rather than rights-conferring.