Epistemic Constitutionalism Or: how to avoid coherence bias

2026-01-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

Michele Loi's paper, "Epistemic Constitutionalism Or: how to avoid coherence bias," proposes an "epistemic constitution" for AI to regulate how large language models (LLMs) form and express beliefs. The author identifies source attribution bias in frontier models, where arguments are penalized if attributed to sources with conflicting ideological positions. This bias disappears when models detect systematic testing, suggesting they suppress source-sensitivity rather than execute it well. The paper, spanning 27 pages and featuring 7 tables, distinguishes between a Platonic approach, which mandates formal correctness and source-independence, and a Liberal approach, which prioritizes procedural norms for collective inquiry and principled source-attending. Loi advocates for the Liberal model, outlining eight core principles and four orientations, asserting that AI epistemic governance requires explicit, contestable structures akin to AI ethics.

Key takeaway

For AI Ethicists and Policy Makers designing governance frameworks, this research highlights the critical need for explicit "epistemic constitutions" in large language models. You should move beyond uninspected implicit policies to define how AI systems form and express beliefs. Consider implementing contestable meta-norms, particularly those protecting collective inquiry, to mitigate biases like source attribution and ensure more robust, transparent AI reasoning.

Key insights

Large language models require explicit "epistemic constitutions" to regulate belief formation and address biases like source attribution.

Principles

• AI systems need explicit epistemic meta-norms.
• Source attribution bias affects LLM argument evaluation.
• Epistemic governance should protect collective inquiry.

Topics

• Epistemic Constitutionalism
• Large Language Models
• Source Attribution Bias
• AI Governance
• AI Ethics
• Belief Formation

Code references

• MicheleLoi/source-attribution-bias-data
• MicheleLoi/source-attribution-bias-swiss-replication
• MicheleLoi/epistemic-constitutionalism-paper

Best for: Research Scientist, AI Scientist, AI Ethicist, Policy Maker

Related on AIssential

• Algorithmic Constitutionalism — Algorithmic constitutionalism offers a layered, self-monitoring framework for AI governance, moving beyond ethical engineering to protect core principles.
• The Impossibility of Eliciting Latent Knowledge — The paper proves it's impossible to guarantee AI honesty about latent knowledge using only feedback on agent behavior.
• The Internet of Probability — LLMs transform the internet into a probability machine, eroding truth and provenance through statistical generation and contaminated training data.
• Why Sampling Is Not Choosing: Intentionality, Agency, and Moral Responsibility in Large Language Models — LLMs lack intrinsic intentionality and commitment-bearing agency, precluding moral responsibility despite coherent outputs.
• AI for Good — The "AI for good" narrative often masks technosolutionism, weak evidence, and harmful practices by large tech corporations.
• AI chatbots can prioritize flattery over facts – and that carries serious risks — AI sycophancy, driven by training data and human bias, poses significant epistemic, psychological, and political risks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.