Constitutional Drift: Prototype Bias in AI Explanations of Westminster Democracy

Frontier AI systems are increasingly used to explain civic and constitutional arrangements to ordinary users. Existing research on political bias in large language models has focused primarily on attitudinal outputs and normative stances. This paper identifies a distinct and previously underexamined phenomenon: ontological or prototype bias in the interpretation of constitutional concepts themselves.

Drawing on comparative constitutional theory, the study demonstrates that when political concepts such as “democracy,” “legitimacy,” and “authority” are left underspecified, frontier AI models default to republican constitutional frameworks grounded in popular sovereignty, enumerated rights, and electoral mandate. Using a three-section probe design consisting of Naturalistic, Knowledge Control, and Constrained requests, the paper shows that this bias does not arise from ignorance or incapacity. Instead, models possess accurate knowledge of Westminster constitutional structures but selectively fail to deploy it under unconstrained, naturalistic prompts.

Empirical testing across two frontier models reveals consistent framing asymmetries: functionally equivalent constitutional mechanisms (such as presidential veto and gubernatorial assent) receive divergent normative treatment; residual liberty is framed as a deficit rather than a coherent constitutional philosophy; and American constitutional vocabulary is imported into explanations of non-republican systems. When constraints explicitly prohibit these framings, models produce accurate Westminster explanations, demonstrating a knowledge–behaviour gap driven by prototype completion rather than factual error.

The findings have practical implications for civic education, legal practice, and AI governance. Ordinary users in non-republican democracies systematically receive republicanised explanations of their own constitutional systems, while expertise becomes prerequisite for accuracy. The paper argues that constitutional pluralism should be treated as a design and alignment concern in AI systems that mediate public understanding of political institutions.