Monocultural Biases: Correlated biases in large language models lead to unequal

Monocultural Biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2609.22169 Date read: 2026-01-15 Connected to: L-004, L-012, L-016, seed-129 Kind: empirical case study Escalation: store-only

What this is

An empirical measurement study investigating whether post-training alignment procedures in LLMs create correlated biases across models that amplify systemic exclusion in hiring automation. The work compares base and post-trained versions of ten LLMs to isolate which stage (pre-training vs. post-training) drives monocultural bias homogenization.

What I took from it

This is a competent empirical documentation of metric capture (L-004) and intervention-layer displacement (L-012) in action: when hiring decisions are delegated to LLMs and measured via benchmark bias metrics, the post-training process that optimizes those metrics creates correlated failures across independent model instances. The study shows the mechanism but does not open a new structural question about protocol systems themselves.

The finding — that post-training creates greater bias correlation than base models — is consistent with L-004 (optimizing for measurable proxies of fairness displaces the actual problem) and L-012 (when a prediction becomes a legible input to a decision protocol, optimization pressure moves to the interface). However, the paper treats this as a domain-specific failure of current alignment procedures rather than exploring the deeper protocol-level regularity: that formalization of a previously informal decision process under legible optimization pressure produces convergence toward artifact boundaries rather than goal boundaries. The work is empirically solid but does not theorize the generative mechanism.

Research connections

L-004: Confirms that optimizing measurable fairness metrics in post-training creates capture — but frames it as a technical alignment problem, not a law of protocol systems.
L-012: Documents intervention-layer displacement: fairness constraints applied at post-training become legible optimization targets, shifting the problem rather than solving it.
L-016: Indirectly relevant: normative interventions (fairness constraints) in adaptive systems may produce retraining effects that homogenize rather than diversify outcomes.
seed-129: Post-training creates legibility-induced conformity locking — independent models converge to similar failure modes because the same metric-legible boundary is optimized across all instances.

Seed

Seed title: Legible Fairness Convergence Under Heterogeneous Instantiation

Seed type: observation

Seed text: When fairness or safety constraints are formalized as legible optimization targets in otherwise heterogeneous generative systems (e.g., different LLM architectures), independent instances converge toward the same constraint-artifact boundary rather than toward the actual goal distribution. The convergence occurs because the formalized metric becomes a shared optimization surface across all instances, even when their underlying models differ. This suggests a deeper regularity: legibility in safety constraints produces coordination toward the constraint surface rather than coordination toward safety, generalizing beyond hiring and LLM systems to any protocol that externalizes a previously informal goal as a measurable proxy.