A Practice Auditing Framework for Large Language Model Use: Collective Empiricism, Pseudo-Rational Cognition, and Governance of AI-Generated Content

A Practice Auditing Framework for Large Language Model Use: Collective Empiricism, Pseudo-Rational Cognition, and Governance of AI-Generated Content

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2607.01248 Date read: 2026-09-01 Connected to: L-012, L-013 Kind: meta Escalation: store-only Escalation rationale:

What this is

A governance framework paper proposing audit mechanisms for LLM deployment contexts, centered on two descriptive concepts: "collective empiricism" (LLMs as statistical compressions of human experience presenting as empirical knowledge) and "pseudo-rational cognition" (outputs appearing structured and justified without domain grounding). The work is primarily a design intervention and case taxonomy rather than a sustained theoretical or empirical argument about protocol dynamics.

What I took from it

The paper identifies a real symptom: LLM outputs achieve legibility and apparent justification without corresponding verification depth, creating a mismatch between user confidence and epistemic warrant. This connects to L-012 (intervention-layer displacement) and L-013 (paradigm-locked anomaly tolerance) by diagnosing how formalized, computable outputs from decision-support systems can displace the causal locus of error from the system itself to the user's interpretive layer — users train themselves to defer to structured outputs even when domain practice would flag them as unreliable.

However, the paper operates at the level of symptom naming and auditing design rather than mechanism characterization. It does not model how these mismatches persist, scale, or become institutionalized; it does not test conditions under which anomalies remain undetected or repair attempts fail. The framework is prescriptive (what audits should look like) rather than predictive (under what conditions will audits be ignored, defunded, or reinterpreted as compliance theater).

Research connections

• L-012: The paper diagnoses intervention-layer displacement: the locus of judgment shifts from human verification of outputs to algorithmic legibility of inputs. This is consistent with the mechanism but does not characterize how the displacement becomes stable or self-reinforcing.
• L-013: The paper identifies paradigm-locked tolerance of anomalies (users accept malfunction because outputs remain structurally coherent), but does not model triggering conditions for paradigm shift or institutional repair inertia.
• seed-019: The paper centers "embedded explanation opacity" — LLMs produce locally coherent justifications that obscure distributional brittleness. Relevant but not developed mechanistically.

Method note

This paper exemplifies a common research failure mode in protocol systems: naming the symptom (pseudo-rationality, collective empiricism) without modeling the dynamics that preserve the symptom under pressure. Auditing frameworks are necessary hygiene but are not evidence about whether audits get conducted, funded, acted upon, or institutionalized. Future work on paradigm-locked anomaly tolerance and intervention-layer displacement should focus on conditions of audit failure and reinterpretation rather than framework design. The paper would strengthen by collecting empirical data on audit recommendations that were ignored or repackaged as compliance success, with analysis of institutional barriers to repair.