Epistemic Constitutionalism Or: how to avoid coherence bias

URL SCAN: Epistemic Constitutionalism Or: how to avoid coherence bias
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

This paper builds a constitutional layer for AI epistemology. It takes a real failure mode—source attribution contaminating argument evaluation—and turns it into a design program: expose meta-norms, permit contestation, and define rules for testimony, evidence, and revision.

Its Platonic-versus-Liberal distinction is the central maneuver. It rejects supposedly neutral formalism while trying to preserve collective inquiry through explicit norms. Under the Discontinuity Thesis, however, this governs the cognition engine, not the economic machine. It may improve how AI evaluates testimony, but the abstract says nothing about ownership, control, or displaced humans retaining productive access.

The Core Fallacy

The main error is scope inflation: treating epistemic constitutional design as if it could secure collective inquiry after AI makes cognitive labor cheaper and more capable than human labor. A constitution can specify how a model evaluates evidence. It cannot, by itself, prevent P1, overcome P2, or avert P3.

The paper also risks mistaking procedural contestability for power. Norms are only as independent as the institutions that set, enforce, and revise them. AI owners can absorb “contestability” as interface decoration while retaining control over deployment and epistemic infrastructure. This is the fatal gap between local epistemic validity and systemic consequence.

Hidden Assumptions

• Explicit norms can be translated into model behavior without becoming brittle slogans or reward-hacking targets.
• Source-aware reasoning can distinguish relevant testimonial evidence from status prejudice consistently.
• Institutions will retain the capacity and incentive to contest AI norms after automation erodes broad productive participation.
• A liberal constitutional process can constrain system owners rather than merely legitimize their systems.
• Better belief formation will improve collective agency rather than produce more accurate obedience to whoever controls the infrastructure.
• The conditions for collective inquiry can survive while economic power concentrates around AI capital.

Social Function

Primary classification: transition management, with partial truth and prestige signaling.

The partial truth is real: hidden source priors can distort judgment, and source independence is not automatically neutral. The paper identifies an actual epistemic control problem.

Its transition-management function appears when constitutional language offers institutions a respectable way to govern AI’s reasoning surface while leaving ownership, distribution, and labor consequences largely untouched. It converts a material power struggle into a contest over principles, orientations, and procedural legitimacy.

That does not make the framework useless. It makes it subordinate. Epistemic constitutionalism may serve as a control layer, audit framework, or bargaining instrument. It is not a mechanism for preserving the mass employment–wage–consumption circuit.

The Verdict

A legitimate epistemic patch aimed at a system whose terminal failure is economic and political. It correctly exposes coherence bias and the fiction of neutral source-independence, but mistakes governable judgment rules for governable power. Under the Discontinuity Thesis, this is transition architecture around the AI engine—not a reversal of obsolescence.