Can LLMs Rank Candidates with Missing Sensitive Attributes Fairly?

Abstract: Large language models (LLMs) are increasingly deployed in high-stakes ranking systems used for hiring, lending, and scholarship allocation, raising concerns about fairness, accountability, and ethical use. These challenges are exacerbated in ranking settings where sensitive demographic attributes are unavailable due to legal, ethical, or practical constraints. Inferring such attributes may introduce harm by violating consent requirements, misrepresenting individuals, and reinforcing structural inequities. This work thus investigates the timely question: How is LLM-based fair re-ranking impacted when demographic information is missing? In this context, we study three alternate strategies: (1) inferring sensitive attributes using traditional third-party services prior to ranking, (2) directly prompting LLMs to produce fair rankings without explicit mention of attribute inference, and (3) employing a chain-of-thought approach in which LLMs are first prompted to infer attributes and thereafter to perform fairness-aware re-ranking. We compare these strategies across multiple datasets using established group-fairness metrics for ranking. Our experiments demonstrate that LLMs match the accuracy of leading third-party services in demographic inference. Moreover, LLMs can embed fairness objectives into rankings even without explicitly inferring sensitive attributes, revealing a new design space for fairness interventions that avoids direct demographic labeling. Lastly, few-shot prompting is found to be crucial for striking the desired balance between fairness and utility. We conclude by discussing the ethical and governance implications of deploying LLMs for fairness-critical ranking tasks. While LLMs offer flexibility under demographic uncertainty, their capacity for implicit inference also raises significant risks if adopted without transparency, evaluation, and institutional oversight. To support reproducibility and scrutiny, we release our source code and experimental artifacts.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Shangtong_Zhang1

Submission Number: 7833