Epistemic Constitutionalism Or: how to avoid coherence bias
A new paper proposes "epistemic constitutionalism" as a framework for regulating how AI systems form and express beliefs, aiming to prevent biases like source attribution bias. The author argues that current frontier models exhibit this bias by penalizing arguments from sources whose expected ideology conflicts with the argument's content. The paper advocates for a "Liberal" approach to AI epistemic governance, emphasizing procedural norms that support collective inquiry over a rigid, "Platonic" adherence to formal correctness. AI
IMPACT Proposes a new framework for AI belief formation and governance, addressing potential biases in reasoning systems.