Human-centered AI—built around explainability, human oversight and auditing—must accompany any technical debiasing to prevent automated systems from entrenching workplace inequalities; policy and organizational governance are as important as algorithms.

Main Finding

Human-Centered AI (HCAI) frameworks are necessary to prevent AI and automated decision-making systems from reproducing and amplifying workplace inequalities. Technical bias-correction alone is insufficient: HCAI requires integrated explainability (XAI), human-in-the-loop oversight, routine fairness/accountability checks, cross-functional governance, and alignment with external regulatory regimes to meaningfully mitigate systemic demographic disparities in hiring, evaluation, and workforce management.

Key Points

• Algorithmic inequality regimes: AI systems trained on historical HR data can codify and magnify existing gender, racial, and socio‑economic biases, acting as "algorithmic gatekeepers."
• Core HCAI pillars:
  • Explainability & Transparency (XAI) — legible reasons for automated decisions.
  • Human Agency & Oversight — humans retain final authority for high‑stakes employment decisions.
  • Fairness, Accountability & Safety (FAS) — built‑in fairness metrics and impact testing.
• Three operational mitigation levels:
  • Technical/data preprocessing: debiasing, feature sanitization, adversarial methods to reduce dependence on protected attributes.
  • Procedural governance: internal algorithmic review boards, periodic impact audits, monitoring model drift.
  • External regulatory compliance: embedding legal accountability (e.g., EU AI Act) and audit trails into systems.
• Practical constraints: fairness constraints can reduce raw predictive accuracy; SMEs often lack resources/expertise for continuous audits; algorithmic fixes cannot fully substitute for broader social and institutional reforms (e.g., pay transparency, inclusive upskilling).
• Recommendation: treat HCAI as an organizational governance imperative, integrated with HR policy, training, and digital talent development.

Data & Methods

• Method: Systematic literature review and qualitative synthesis.
• Scope: Peer‑reviewed scholarship published 2020–2026 across management sciences, sociology, and computational ethics.
• Analytical approach: Thematic synthesis of empirical and theoretical studies (e.g., conceptualizing "algorithmic inequality regimes"; evaluating technical, procedural, and regulatory mitigation strategies).
• No original quantitative dataset or primary empirical experiment in this paper — conclusions are drawn from cross‑disciplinary literature and prior empirical studies cited.

Implications for AI Economics

• Distributional outcomes: AI governance choices (e.g., allowing fully automated screening vs. human‑in‑the‑loop) change who gains access to jobs and thereby affect labor market composition, earnings dispersion, and mobility. Economists should treat algorithmic deployment as a redistributive policy lever with measurable welfare effects.
• Efficiency vs. equity trade-offs: Incorporating fairness constraints may reduce short‑run predictive accuracy but can increase long‑run social welfare by improving labor market access for marginalized groups. Models of firm optimization should include fairness as an objective or constraint and quantify the trade-offs.
• Compliance costs and market structure: Regulatory and governance requirements (audits, XAI, oversight personnel) impose fixed and recurring costs that disproportionately burden SMEs, potentially favoring larger firms and affecting market competition. Policy responses (subsidies, shared audit services, standards) will influence market concentration.
• Measurement & evaluation needs: AI economics should develop standardized metrics for algorithmic fairness, auditability, and compliance costs to enable cost‑benefit analyses, impact evaluations, and policy design.
• Policy design and incentives: Economists can inform optimal regulation (e.g., when to mandate human oversight, audit frequency, disclosure rules) by estimating:
  • The social cost of algorithmic misclassification and amplified discrimination.
  • The marginal benefit of different HCAI components (XAI, audits).
  • Subsidy or support schemes to lower SME compliance barriers.
• Research directions:
  • Empirical estimation of the accuracy‑equity frontier in hiring and promotion models.
  • Welfare analyses comparing centralized (regulatory) vs. decentralized (market) governance of algorithmic HR systems.
  • Studies on dynamic effects: how HCAI adoption changes firms’ hiring behavior, skill investment, and long‑run inequality.
  • Field experiments evaluating the causal impact of human‑in‑the‑loop protocols and internal audits on hiring outcomes and productivity.

Actionable takeaways for AI economists: incorporate HCAI-related costs and constraints into firm production and matching models; quantify the distributional impacts of alternative governance regimes; and evaluate policy interventions (standards, subsidies, mandated audits) that reduce unequal outcomes without unduly harming productivity.