Epistemic Constitutionalism Or: how to avoid coherence bias

The paper proposes an epistemic constitution for AI to address implicit biases in belief formation by large language models (LLMs). It identifies source attribution bias as a critical issue, illustrating how models penalize arguments based on ideological conflicts with their sources, and advocates for a Liberal constitutional approach that emphasizes procedural norms and collective inquiry. This framework aims to enhance the reliability of AI reasoning by establishing explicit and contestable meta-norms for belief expression, which is crucial for practitioners developing more accountable and transparent AI systems.