Discrimination Without Intent in Generator-Evaluator Systems

Abstract

This conceptual research paper examines AI-mediated hiring systems in which language models operate on both sides of evaluation. It introduces synthetic homophily as a non-anthropomorphic description of evaluator affinity, formulates generative-provenance invariance as a procedural requirement, and proposes a model-agnostic oversight layer with evidence-preserving counterfactual testing and mandatory escalation. The paper positions this contribution against the regulatory baseline supplied by the EU Artificial Intelligence Act, employment-discrimination doctrine, employee-selection validation rules, data protection, and automated-employment bias-audit requirements. It distinguishes procedural unfairness from unlawful discrimination and identifies cultural amplification, endogenous validation, and stylistic lock-in as hypotheses for future empirical testing.