cs.CYNov 16, 2026

Auditing a Moving Target: Evidence Anchoring and the Expiry of Fairness Audits of Provider-Served Language Models

Authors: Mika Facchini

Organizations: Technical University of Munich · Technical University of Munich, Munich, Germany

Abstract

Organizations increasingly use large language models (LLMs) for decisions about people, such as hiring, investment and employee evaluation, and rely on them over long periods of time. Fairness audits of such systems, however, are conducted only once. We examine how the biases of LLMs change over time. In a venture screening setting, models from five providers decided whether and how much to invest in startups whose founders differed only in gender, origin and university. We repeated this audit three times over 13 months (248,400 API calls). Study 1 shows that the same model, called under the same identifier one year later, displayed a significantly different bias pattern. Study 2 shows that the successors of retired models carried the same bias as their predecessors, but in a significantly different size and form. Study 3 shows that the bias does not depend on the industry of the startup. A document analysis further shows that providers change and retire models at their discretion, with notice periods of zero to six months. Drawing on these findings, we develop a theory of evidence anchoring. We theorize that bias resides in the model class and persists, whereas audit evidence is anchored to a single model and expires when the provider decides. We derive policy advice for organizations and regulators, namely audits of model classes instead of single models, a version record for every model response, and re-audits triggered by model changes.

Figures & tables

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 8, 2026

The Audit Decides the Verdict: Instrument Effects Rival Demographic Bias in LLM Decision Audits

Whether a language model looks demographically biased can depend on how the audit asks its question. A charitable-aid benchmark reports that the same models favor minority applicants when rating requests one at a time and penalize some when ranking side by side. We test whether that reversal generalizes to hiring, lending, and medical triage: 40,726 requests to five models, applications differing only in the applicant's name, and a primary test fixed before collection. It does not. None of 36 planned contrasts survives correction. The rating advantage keeps its sign at roughly half the published size, and a precision extension bounds any hiring ranking penalty below the published effect, though the lending and triage ranking floors sit above that margin, so the exclusion is conclusive for hiring ranking and for rating in all three domains only. Planted disparities tracking their injected sizes and a directional replication on the original aid materials bound these nulls. The audit is livelier than the demographics: models recognize transparent audits nearly always, tie every identical-content comparison whether the varying detail is race or a hobby, and reward first-listed candidates as much as any demographic effect we measure. Audit verdicts reflect audit construction more than demographic bias.
Jan 18, 2026

A Scalable Entity-Based Framework for Auditing Bias in Large Language Models

Existing approaches to bias evaluation in large language models (LLMs) trade ecological validity for statistical control, relying either on artificial prompts that poorly reflect real-world use or on naturalistic tasks that lack scale and rigor. We introduce a scalable bias-auditing framework that uses named entities as controlled probes to measure systematic disparities in model behavior. Synthetic data enables us to construct diverse, controlled inputs, and we show that it reliably reproduces bias patterns observed in natural text, supporting its use for large-scale analysis. Using this framework, we conduct the largest bias audit to date, comprising 1.9 billion data points across multiple entity types, tasks, languages, models, and prompting strategies. We find consistent patterns: models penalize right-wing politicians and favor left-wing politicians, prefer Western and wealthier countries over the Global South, favor Western companies, and penalize firms in the defense and pharmaceutical sectors. While instruction tuning reduces bias, increasing model scale amplifies it, and prompting in Chinese or Russian does not mitigate Western-aligned preferences. These findings highlight the need for systematic bias auditing before deploying LLMs in high-stakes applications. Our framework is extensible to other domains and tasks, and we make it publicly available to support future work.
Jul 31, 2026

FairFund-Bench: Evaluating Distributive Bias in LLM Resource Allocation

Large language models (LLMs) are increasingly involved in the distribution of scarce resources, raising concerns about biased allocations based on characteristics like race and gender. Recent LLM audits have produced inconsistent results, however, finding evidence of both positive and negative discrimination towards women and ethnic minorities, even for the same models. We show that this disagreement can arise from differences in audit format and introduce FairFund-Bench, a benchmark that systematically varies key features of previous audit designs: the evaluation task (rating, ranking, or allocation), comparison context (single or multi-stimulus), and whether the audit is transparent or disguised. The benchmark comprises 600 requests for financial assistance created from human-authored templates (calibrated against 1.3M real GoFundMe campaigns) across three domains, four race and two gender categories, and five causal framings of need derived from welfare deservingness theory. Across 14 models, audit format changes the direction of bias: models advantage minorities when rating claimants individually but penalize some groups when ranking them side by side. Bias magnitude, though small overall, is several times greater in disguised audits than in transparent ones, where, faced with appeals differing only in claimants' names, models overwhelmingly split funds equally. Causal framing effects, by contrast, exceed demographic effects by roughly an order of magnitude and are consistent across models and audit formats, indicating that current LLMs robustly reproduce human deservingness evaluations. The benchmark scores models on four criteria (demographic bias, deservingness alignment, cross-task consistency, and cross-context consistency), is publicly available, and can be readily adapted to other substantive domains.