Epistemic Subordination: Generative AI and the Infrastructure of Knowledge

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.18758 Date read: 2026-09-02 Connected to: L-004, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

A critical theory paper arguing that generative AI systems encode dominant cultural epistemologies as default infrastructure through training data compression. The work frames this not as output bias but as structural subordination of minority knowledge frameworks via probabilistic baseline effects.

What I took from it

The paper maps onto L-004 (Goodhart Generalization) and L-012 (Intervention-Layer Displacement) but does not advance either substantially. It restates the mechanism of proxy capture—training data distribution becomes optimization target, dominant patterns dominate output—as a knowledge-infrastructure problem. This is sociologically interesting but mechanically familiar: the paper describes where optimization pressure concentrates (the statistical baseline of the model) rather than discovering a new law governing how or why infrastructure becomes the locus of pressure displacement.

The work does not isolate a generalizable mechanism absent from the current inventory. It applies existing proxy-capture logic to a new domain (epistemology / cultural representation) without revealing novel dynamics about how protocols, systems, or artificial infrastructures behave under adoption or scaling pressure. The insight is correct but derivative.

Research connections

L-004: Confirms proxy capture in training objectives (majority representation as measurable proxy for balanced knowledge), but offers no new mechanism or boundary condition.
L-012: Describes the phenomenon (optimization pressure shifts to model baseline) but treats it as domain-specific rather than a general protocol property.
seed-073 (Correlated Failure Under Proxy Consensus): Tangentially relevant—the paper suggests that consensus on statistical baselines may create correlated failures across downstream systems, but does not develop this.

Seed

Seed title: none