Counterfactual, Per-Decision Bias Auditing for Automated Hiring: Localizing and Explaining Disparate Impact in Applicant Tracking Systems

Shallow read  ·  2026  ·  source  ·  all reading

L-004 L-012 L-014

Counterfactual, Per-Decision Bias Auditing for Automated Hiring: Localizing and Explaining Disparate Impact in Applicant Tracking Systems

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.21537 Date read: 2026-09-02 Connected to: L-004, L-012, L-014 Kind: content Escalation: store-only Escalation rationale:

What this is

A tool paper presenting the AI Bias Firewall (AIBF), a method that bridges group-level fairness auditing and per-decision explainability in hiring systems by connecting individual prediction attributions to disparate impact legal standards. The work sits at the intersection of legal compliance and interpretability for automated decision-making.

What I took from it

The paper demonstrates a specific instantiation of L-012 (Intervention-Layer Displacement) and L-014 (Strategic Boundary Concentration Under Computable Legality): once hiring bias becomes legally defined in computable terms (disparate impact ratios operationalized as per-decision counterfactual thresholds), the optimization pressure migrates. The AIBF method itself exemplifies this pattern—by making the legal standard legible to both auditors and potentially the hiring system's designers, it creates a new target surface. The paper does not theorize this displacement; it engineers around it. The work confirms that legal formalization of bias (a form of protocol ossification via L-001) creates a demand for localized legibility that preserves compliance while potentially obscuring upstream proxy capture in feature engineering or training data labeling.

However, the paper does not investigate whether finer-grained per-decision auditability changes hiring system behavior itself, nor does it examine whether auditing pressure redistributes bias rather than eliminating it—both phenomena that would deepen L-012 and L-014. It is a competent compliance instrument, not a theory-advancing empirical or conceptual contribution.

Research connections

  • L-004 (Goodhart Generalization): Hiring systems optimized against disparate impact ratios may capture the metric (group-level fairness) while individual decisions remain biased along unmeasured dimensions.
  • L-012 (Intervention-Layer Displacement): Formalizing disparate impact as a legible audit signal may displace optimization pressure from the group-level metric to individual decision boundaries or upstream feature selection.
  • L-014 (Strategic Boundary Concentration Under Computable Legality): Once hiring bias is rendered computable and auditable, optimizing agents may concentrate behavior at the boundary of legal compliance rather than moving away from bias altogether.

Seed

Seed title: Audit Legibility as Proxy Boundary Optimization Seed type: observation Seed text: When a protocol violation (hiring bias) is formalized as a computable audit signal that is then made legible to the system's operators or designers, optimization pressure may concentrate at the boundary of detectability rather than at the underlying violation. Per-decision auditing, by localizing bias detection, enables finer-grained compliance gaming: systems may learn to distribute bias below the individual-decision threshold while maintaining aggregate disparity. The mechanism generalizes to any protocol where formalization enables legible audit signals—compliance becomes a legibility problem, not a structural one.