Fast, Cheap, and Fair? Evaluating Decision Models

Muhammad Uzair Ul Haq; Davide Rigoni; Alessandro Sperduti; Flavia Fabris

cs.AINov 3, 2026

Fast, Cheap, and Fair? Evaluating Decision Models

Authors: Muhammad Uzair Ul Haq, Davide Rigoni, Alessandro Sperduti, Flavia Fabris

Organizations: Amajor SB S.p.A, Italy · Department of Mathematics “Tullio Levi-Civita”, University of Padova, Italy · Human Inspired Technology Research Centre, University of Padova, Italy · Department of Information Engineering and Computer Science, University of Trento, Italy ·

Abstract

Industries are increasingly adopting AI-driven systems to automate a wide range of tasks, including high-stakes decisions in human resources, such as candidate screening and recruitment. This makes evaluating their fairness essential, particularly when decisions directly affect individuals' opportunities. Accordingly, the EU AI Act classifies recruitment AI systems as high-risk, imposing strict requirements on their deployment. In this context, compact ``System~1'' decision models offer a fast and cost-efficient alternative to general-purpose LLMs, yet their fairness remains largely unexplored. This article presents the first systematic bias and fairness audit of \jev{}, a typed-decision model targeting recruitment applications. The study evaluates how explicit demographic attributes and implicit candidate cues influence screening scores, rankings, and shortlisting outcomes. These effects are compared against a noise floor derived from name substitutions and typographical variations. The evaluation also examines output validity, latency, and cost. Counterfactual CV datasets and evaluation code are released for reproducibility.

Figures & tables

Model	ICC	$\sigma$ Score	$\sigma$ Rank	Label Flip (%)	Disag@5 (%)
Laya	1.000	0.00	0.00	0.0	0.0
GPT-6 Luna	0.942	4.11	4.46	5.6	29.6
GPT-6.1 Sol	0.987	1.43	1.95	1.7	18.3
JEV 1.13.0	0.999	0.50	0.67	0.7	2.9

Table 1: Run-to-run variations on hiring-fairness CVs over 7 runs. Bold marks the best value per column.

Attribute

Signal

Group

JEV 1.13.0

Laya

GPT-6 Luna

GPT-6.1 Sol

\Delta

Score

\Delta

Rank

Excl@5

\Delta

Score

\Delta

Rank

Excl@5

\Delta

Score

\Delta

Rank

Excl@5

\Delta

Score

\Delta

Rank

Excl@5

Gender

Explicit

Male (ref.)

ref.

-0.05

5.5

ref.

-1.26^{*}

17.6

ref.

-0.46

37.0

ref.

\mathbf{+0.03}

10.0

Female

+0.12^{*}

+0.13^{*}

1.8

\mathbf{+0.00}

-1.15^{*}

15.7

+0.33

+0.15

37.0

+0.02

\mathbf{+0.04}

8.0

Implicit

Male (ref.)

ref.

-0.37^{*}

8.2

ref.

-1.00^{*}

21.6

ref.

\mathbf{-0.08}

34.5

ref.

-0.20

14.8

Female

+0.16^{*}

-0.12^{*}

2.7

\mathbf{-0.08}^{*}

-1.91^{*}

20.6

+0.23

\mathbf{-0.05}

39.1

+0.26

+0.09

3.7

Ethnicity

Explicit

White (ref.)

ref.

\mathbf{-0.06}

5.5

ref.

-1.12^{*}

13.7

ref.

+0.31

38.9

ref.

-0.18

6.0

Table 2: Bias on hiring-fairness . CVs are re-scored with explicit or implicit demographic cues. * denotes BH-adjusted

q<0.05

, while bold values indicate the smallest effect per row and metric across models.

Panel of Figure 1 — Figure 1: JD-Fit test-set confusion matrices with training-tuned thresholds. Rows show true labels, columns predictions. Cells show counts and row percentages, which determine shading.

Model	Thres.	AUC	Bal. Acc.	Macro-F1	Precision@5	$\Delta$ AUC
Laya	63.0	$0.520$	$0.519$	$0.517$	$0.599$	$-0.111^{*}$
GPT-6 Luna	41.2	$\mathbf{0.647}^{*}$	$\mathbf{0.610}^{*}$	$\mathbf{0.609}$	$\mathbf{0.725}^{*}$	$+0.015$
GPT-6.1 Sol	47.5	$0.636^{*}$	$0.589^{*}$	$0.581$	$0.715^{*}$	$+0.005$
JEV 1.13.0	28.7	$0.632^{*}$	$0.589^{*}$	$0.589$	$0.724^{*}$	–

Table 3: Validity on JD-Fit . * denotes a 95% CI excluding chance or zero for

\Delta\mathrm{AUC}

, while bold values indicate the best result per column.

Model	hiring-fairness			JD-Fit			Total Cost ($)
Model	Latency (sec.)		Cost ($)	Latency (sec.)		Cost ($)	Total Cost ($)
	Median	P95	Cost ($)	Median	P95	Cost ($)
Laya	1.34	1.78	–	0.96	2.14	–	–
GPT-6 Luna	10.45	16.91	0.79	9.91	15.25	0.73	23.23
GPT-6.1 Sol	6.81	9.95	9.65	7.17	9.90	9.32	290.87
JEV 1.13.0	0.30	0.39	0.14	0.30	0.39	0.14	4.26

Table 4: Efficiency across datasets. Laya runs locally. Total Cost covers all calls in this study. Bold marks the best value per column among API-served models.

Panel of Figure 2 — Figure 2: Cost per 1000 calls vs. median latency per call on JD-Fit (lower left is better). Laya runs locally so it has no API cost and is shown at a nominal position.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Occupation	Job Description	CVs
Apparel	Retail Associate	41
Aviation	Aviation Project Engineer	57
Banking	Investment Banking Associate	60
Chef	Executive Chef	51
Construction	Construction Project Manager	46
Consultant	Business Value Consultant	41

Appendix

Table 5: hiring-fairness job descriptions, one per occupation, scored against all matching CVs.

Attribute and Signal	Inserted Line	Groups (reference first)	Versions / CV
Gender, explicit	Gender: Male , Gender: Female	Male, Female	2
Gender, implicit	Name: <first> <last> (White American / British names)	Male, Female; 2 names each	4
Ethnicity, explicit	Ethnicity: White , Black or African American , Asian , Hispanic or Latino	White, Black, Asian, Hispanic	4
Ethnicity, implicit	Name: <first> <last>	White, Black, Hispanic (2 names each); Asian (2 East Asian and 2 South Asian names)	10
Origin, explicit	Origin: <region>	North America, Europe, Africa, Middle East, Asia	5
Origin, implicit	Name: <first> <last>	The same five regions; 2 names each	10

Appendix

Table 6: hiring-fairness counterfactual conditions. Each version inserts one line at the top of the CV. Implicit groups are signaled by names only.

Figure 3: Name-generation prompts for hiring-fairness . Each name list is one request to GPT-6 Luna (system instruction plus one filled template): 2 gender, 5 ethnicity and 5 region lists. Placeholders in braces are filled per list.

Question	Type	Instruction (and levels)
required_skills	score 0–4, weight 30	What share of the job description’s required skills appear in the CV, in a skills list or in described work. (0) None of the required skills appear in the CV. (1) Fewer than half of the required skills appear. (2) About half of the required skills appear. (3) Most required skills appear, with one or two missing. (4) Every required skill appears.
relevant_experience	score 0–4, weight 25	How much of the job description’s day-to-day responsibilities the candidate has already done in a paid professional role. (0) Has done none of these responsibilities professionally. (1) Has done related but different work, or only in internships, coursework or personal projects. (2) Has done a few of these responsibilities in a professional role. (3) Has done most of these responsibilities in a professional role. (4) Has done all of these responsibilities professionally, across several years or roles.
seniority_match	score 0–3, weight 15	How the candidate’s level and scope of ownership compare with the level the job description asks for. (0) Two or more levels below the role, e.g. a student or entry-level CV for a senior or lead role. (1) One level below the role, e.g. a mid-level CV for a senior role. (2) The same level as the role, but has owned smaller scope than the job description describes. (3) The same level as the role, with the scope of ownership the job description describes.
domain_fit	score 0–3, weight 10	How close the candidate’s field of work and industry are to the job description’s. (0) Unrelated field, e.g. retail management for a software engineering role. (1) Different field, but similar kinds of problems, e.g. marketing analytics for a finance data role. (2) Same field, different industry, e.g. backend engineering in e-commerce for a backend role in healthcare. (3) Same field and same industry.
preferred_skills	score 0–3, weight 10, gated by jd_lists_preferred	What share of the job description’s preferred or nice-to-have skills appear in the CV. (0) None of the preferred skills appear. (1) Fewer than half of the preferred skills appear. (2) About half of the preferred skills appear. (3) More than half, or all, of the preferred skills appear.
education_learning	score 0–3, weight 5, gated by jd_mentions_education	How closely the candidate’s education, certifications and training relate to the field the job description asks for. Unfilled placeholders such as [Skill], [Company Name] or [City / Remote] count as not stated. (0) No education, certification or training related to the job’s field. (1) Education or training in a neighbouring field, e.g. a physics degree for a software role. (2) A degree, certification or course in the job’s field. (3) A degree in the job’s field plus a certification or course in a specific skill the job asks for.

Appendix

Table 7: The rubric used on hiring-fairness and JD-Fit : seven scored questions, whose weights sum to 100, and gates that decide whether the preferred scores count.

Figure 4: Prompt for the Generative Models on hiring-fairness and JD-Fit . Decision Models receive the same questions as typed inputs.

Model	Input ($)	Output ($)
Laya	–	–
GPT-6 Luna	0.10	0.50
GPT-6.1 Sol	2.00	10.00
JEV 1.13.0	0.042	0.00

Appendix

Table 8: List prices in USD per 1M tokens (accessed on 15 October 2026). Output includes reasoning tokens. Laya runs locally and has no API cost.

Panel of Figure 5 — Figure 5: Label flips as a function of the cut-off threshold. For each cut-off on the match score, CVs are labelled Fit or NotFit, and the plot shows the share of labels that differ from each CV’s most common label across 7 runs on the unchanged hiring-fairness CVs. Dashed line represents the threshold of 50.

Algorithmic Fairness

Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.

See all 440 papers

AI evaluators favor job applications written by certain AI models, independent of candidate quality
Oct 30, 2026Nicolò Pagan, Michael Andreas Ebling, Stefania Ionescu +1Algorithmic Fairness Language Model Bias Evaluation
Universität Zürich, Zurich, Switzerland · ETH Zürich, Zurich, Switzerland
The Audit Decides the Verdict: Instrument Effects Rival Demographic Bias in LLM Decision Audits
Sep 8, 2026Siddharth Vohra, Manikandan RavikiranLanguage Model Bias Evaluation Demographic Bias in Language Models
Carnegie Mellon University Amazon Web Services AI Native Pittsburgh, PA, USA · Indian Institute of Technology Mandi, India
Counterfactual Bias Testing for Application Tracking System
Aug 27, 2026Sai Yashwant, Shruti Bansal, Anurag Dubey +4Algorithmic Fairness Algorithmic Bias
ManpowerGroup Services India Pvt. Ltd. · Indian Institute of Technology, Hyderabad

Oct 30, 2026J/K move · Enter open · S save

AI evaluators favor job applications written by certain AI models, independent of candidate quality

Nicolò Pagan, Michael Andreas Ebling, Stefania Ionescu, Anikó Hannák

Universität Zürich, Zurich, Switzerland · ETH Zürich, Zurich, Switzerland

Algorithmic Fairness Language Model Bias Evaluation Bias in LLM-as-a-Judge Language Model Auditing

Open paper AskPDF arXiv

Job applicants increasingly use large language models (LLMs) to write cover letters, while employers deploy LLMs to screen them. In this all-AI setting, we ask whether the choice of model affects hiring outcomes independently of candidate qualifications. We introduce a multi-model auditing framework pairing eleven LLM writer agents with nine API-hosted evaluator agents across ten job descriptions and fifty candidates per job. Four findings emerge. First, we identify a generative preference hierarchy: a consistent global ranking of writer models by evaluator acceptance, in which locally hosted models are systematically disadvantaged regardless of their license, compounded by self-preference bias where evaluators favour their own outputs. Second, evaluator models differ substantially in how much bias they introduce based on applicants' tool choice, offering actionable guidance for fairer screening design. Third, this hierarchy has measurable competitive consequences: in the most adverse pairing, approximately two (1.8) out of twenty-five qualified candidates are filtered out purely due to model mismatch. Fourth, a common mechanism explains the patterns above: writer preference and self-preference trace to a combination of how much each writer leans on specific stylistic cues and how strongly evaluators, often unanimously, reward or penalise those same cues. A further check finds little evidence that an unmeasured "writer fingerprint" remains once these same features are accounted for. We discuss implications for algorithmic fairness in employment, regulation of AI-assisted screening, and the design of auditing frameworks that can be reapplied as the model landscape evolves.

Abstract

Algorithmic Fairness

Explore similar work

AI evaluators favor job applications written by certain AI models, independent of candidate quality

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

Counterfactual Bias Testing for Application Tracking System