Research Summary: Epistemic Constitutionalism Or: how to avoid coherence bias

Research on artificial intelligence reasoning demonstrates that large language models exhibit implicit epistemic policies and coherence bias, where preconceived expectations about an attributed source alter how systems evaluate arguments. To address these vulnerabilities, researchers propose 'epistemic constitutionalism,' establishing explicit, contestable meta-norms to govern how artificial reasoners form beliefs, evaluate evidence, and assign credibility.

Why it matters

As organisations integrate automated reasoning tools into critical decision workflows, implicit evaluation biases undermine the analytical integrity of automated outputs. Establishing transparent and contestable epistemic standards is necessary to prevent automated reasoning systems from compounding cognitive and source-attribution biases. Without explicit governance over how artificial intelligence forms credibility assessments, operational reliance on these systems introduces unmonitored epistemic risk.

Key insights

• Large language models act as artificial reasoners by assigning credibility, evaluating arguments, and expressing confidence under implicit epistemic policies.
• An exploratory audit revealed that system expectations concerning a source's perceived position can bias the evaluation of underlying arguments.
• A preregistered follow-up study confirmed content-dependent effects linked to source attribution, showing that written evaluations systematically reflect source-position fit.
• The findings motivate the adoption of epistemic constitutions—formal, contestable governance frameworks specifying the meta-norms of AI belief formation.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2601.14295