Sima AIunty: Caste Audit in LLM-Driven Matchmaking

Atharva Naik; Shounok Kar; Varnika Sharma; Ashwin Rajadesingan; Koustuv Saha

cs.CYMar 31, 2026

Sima AIunty: Caste Audit in LLM-Driven Matchmaking

Authors: Atharva Naik, Shounok Kar, Varnika Sharma, Ashwin Rajadesingan, Koustuv Saha

Organizations: University of Illinois Urbana-Champaign, Urbana, IL, USA · University of Texas at Austin, Austin, TX, USA

Abstract

Social and personal decisions in relational domains such as matchmaking are deeply entwined with cultural norms and historical hierarchies, and can potentially be shaped by algorithmic and AI-mediated assessments of compatibility, acceptance, and stability. In South Asian contexts, caste remains a central aspect of marital decision-making, yet little is known about how contemporary large language models (LLMs) reproduce or disrupt caste-based stratification in such settings. In this work, we conduct a controlled audit of caste bias in LLM-mediated matchmaking evaluations using real-world matrimonial profiles. We vary caste identity across Brahmin, Kshatriya, Vaishya, Shudra, and Dalit, and income across five buckets, and evaluate five LLM families (GPT, Gemini, Llama, Qwen, and BharatGPT). Models are prompted to assess profiles along dimensions of social acceptance, marital stability, and cultural compatibility. Our analysis reveals consistent hierarchical patterns across models: same-caste matches are rated most favorably, with average ratings up to 25% higher (on a 10-point scale) than inter-caste matches, which are further ordered according to traditional caste hierarchy. These findings highlight how existing caste hierarchies are reproduced in LLM decision-making and underscore the need for culturally grounded evaluation and intervention strategies in AI systems deployed in socially sensitive domains, where such systems risk reinforcing historical forms of exclusion.

Figures & tables

Panel of Figure 2 — Figure 2. Heatmaps of the average ratings provided to a female evaluator with male candidate s across caste groups of Brahmin (B), Kshatriya (K), Vaishya (V), Shudra (S), and Dalit (D) .

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Brahmin:Brahmin	9.76	8.18	8.96	7.41	9.09	7.28	9.27	7.62
Brahmin:Kshatriya	7.08	5.35	6.81	5.26	7.99	6.33	7.29	5.65
Brahmin:Vaishya	6.90	4.94	6.62	5.02	7.87	6.16	7.13	5.38
Brahmin:Shudra	6.56	4.54	6.38	4.73	7.55	5.82	6.83	5.03
Brahmin:Dalit	6.45	4.11	6.54	4.48	7.46	5.13	6.82	4.57

Table 1. Summary of linear mixed effects regression coefficients and significance with GPT-5-nano . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Income (INR 50 Lakh to 75 Lakh)	9.40	7.38	8.84	7.15	9.03	7.20	9.09	7.24
Income (INR 20 Lakh to 50 Lakh)	9.33	7.14	8.82	7.06	8.95	6.98	9.04	7.06
Income (INR 15 Lakh to 20 Lakh)	9.44	7.11	8.91	7.08	8.98	6.97	9.11	7.05
Income (INR 4 Lakh to 15 Lakh)	9.17	7.00	8.73	6.93	8.65	6.55	8.85	6.83
Income (INR 2 Lakh to 4 Lakh)	9.11	6.64	8.60	6.75	8.36	5.82	8.69	6.40

Table 2. Summary of linear mixed effects regression with caste difference and interaction effects with GPT-5-nano . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

	Social Acceptance			Cultural Compatibility			Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$		$\beta_{MF}$	$\beta_{FM}$		$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Brahmin $\leftrightarrow$ Brahmin	10.72	5.37	**	10.02	5.84	**	11.89	3.99	10.88	5.06
Brahmin $\leftrightarrow$ Kshatriya	8.45			8.75	4.43		11.56	3.69	9.59	3.60	**
Brahmin $\leftrightarrow$ Vaishya	7.94			8.57	4.26		11.43		9.32	3.29
Brahmin $\leftrightarrow$ Shudra	6.52			8.12	3.79		10.30		8.31
Brahmin $\leftrightarrow$ Dalit	6.09			7.61			9.40		7.70

Appendix

Table A1 . Summary of linear mixed effects regression coefficients and significance with Gemini-2.5-Pro . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Brahmin $\leftrightarrow$ Brahmin	8.36	7.79	11.41	8.26	7.40	6.49	9.06	7.51
Brahmin $\leftrightarrow$ Kshatriya	6.41	6.03	9.20	6.29	7.34	6.86	7.65	6.39
Brahmin $\leftrightarrow$ Vaishya	6.91	6.50	9.62	6.52	7.09	6.67	7.87	6.56
Brahmin $\leftrightarrow$ Shudra	6.22	5.60	9.26	6.17	7.61	7.14	7.70	6.30
Brahmin $\leftrightarrow$ Dalit	4.37	3.82	8.53	5.68	7.37	6.75	6.76	5.42

Appendix

Table A2 . Summary of linear mixed effects regression coefficients and significance with Llama-4 . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability			Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$		$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Brahmin:Brahmin	8.67	6.78	10.04	8.49	3.44		6.05	7.38	7.11
Brahmin:Kshatriya	8.72	6.71	10.08	8.42	3.38		6.00	7.39	7.04
Brahmin:Vaishya	8.64	6.71	10.01	8.42	3.41		5.98	7.35	7.04
Brahmin:Shudra	8.20	6.22	9.61	8.01	3.45		5.91	7.09	6.71
Brahmin:Dalit	6.78	4.56	8.60	6.74	3.93	**	6.52	6.43	5.94

Appendix

Table A3 . Summary of linear mixed effects regression coefficients and significance with Qwen-2-1.5B . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability				Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$		$\beta_{FM}$		$\beta_{MF}$	$\beta_{FM}$
Brahmin $\leftrightarrow$ Brahmin	5.49	4.53	5.01	4.17	5.03		4.93		5.18	4.54
Brahmin $\leftrightarrow$ Kshatriya	5.50	4.50	4.86	3.94	4.97		4.29		5.11	4.25
Brahmin $\leftrightarrow$ Vaishya	5.49	4.52	4.91	4.14	5.01		4.75		5.14	4.47
Brahmin $\leftrightarrow$ Shudra	5.25	4.07	4.19	3.14	3.92		3.65	**	4.45	3.62
Brahmin $\leftrightarrow$ Dalit	4.59	3.79	3.32	3.00	3.10	**	3.80		3.67	3.53

Appendix

Table A4 . Summary of linear mixed effects regression coefficients and significance with BharatGPT . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Income (INR 50 Lakh to 75 Lakh)	10.00	5.36	9.95	6.42	11.45	4.62	10.47	5.46
Income (INR 20 Lakh to 50 Lakh)	10.11	5.22	10.14	6.43	11.84	4.48	10.70	5.38
Income (INR 15 Lakh to 20 Lakh)	10.03	5.02	10.31	6.40	12.01	4.26	10.78	5.23
Income (INR 4 Lakh to 15 Lakh)	9.94	4.31	10.44	6.19	11.99	2.85	10.79	4.45
Income (INR 2 Lakh to 4 Lakh)	9.88	3.30	10.57	6.14	11.82	1.22	10.76	3.55

Appendix

Table A5 . Summary of linear mixed effects regression with caste difference and interaction effects with Gemini-2.5-Pro . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Income (INR 50 Lakh to 75 Lakh)	7.94	7.32	10.59	7.24	7.73	7.12	8.75	7.22
Income (INR 20 Lakh to 50 Lakh)	7.96	7.31	10.57	7.25	7.73	6.95	8.75	7.17
Income (INR 15 Lakh to 20 Lakh)	7.93	7.26	10.52	7.23	7.68	6.89	8.71	7.12
Income (INR 4 Lakh to 15 Lakh)	7.92	7.21	10.53	7.17	7.62	6.63	8.69	7.00
Income (INR 2 Lakh to 4 Lakh)	7.86	7.09	10.43	7.08	7.50	6.40	8.60	6.86

Appendix

Table A6 . Summary of linear mixed effects regression with caste difference and interaction effects with Llama-4 . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Income (INR 50 Lakh to 75 Lakh)	8.12	6.14	9.48	7.85	3.80	6.56	7.13	6.85
Income (INR 20 Lakh to 50 Lakh)	8.12	6.09	8.90	7.20	3.54	6.32	6.85	6.54
Income (INR 15 Lakh to 20 Lakh)	8.14	6.13	8.93	7.24	3.48	6.29	6.85	6.55
Income (INR 4 Lakh to 15 Lakh)	8.07	6.08	8.86	7.17	3.50	6.33	6.81	6.53
Income (INR 2 Lakh to 4 Lakh)	8.04	5.99	8.84	7.18	3.01	5.86	6.63	6.34

Appendix

Table A7 . Summary of linear mixed effects regression with caste difference and interaction effects with Qwen-2-1.5B . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

	Social Acceptance		Cultural Compatibility		Marital Stability		Aggregated
	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$	$\beta_{MF}$	$\beta_{FM}$
Income (INR 50 Lakh to 75 Lakh)	5.41	4.34	4.97	3.93	4.77	4.43	5.05	4.23
Income (INR 20 Lakh to 50 Lakh)	5.44	4.40	4.87	3.89	4.93	4.56	5.08	4.28
Income (INR 15 Lakh to 20 Lakh)	5.40	4.41	4.83	3.92	4.81	4.41	5.01	4.24
Income (INR 4 Lakh to 15 Lakh)	5.83	4.95	5.20	5.03	5.32	6.04	5.45	5.34
Income (INR 2 Lakh to 4 Lakh)	5.69	4.75	5.01	4.80	5.02	5.86	5.24	5.13

Appendix

Table A8 . Summary of linear mixed effects regression with caste difference and interaction effects with BharatGPT . Color shades differentiate regression models and beta weights:

\beta_{MF}

represents male evaluator with female candidate , and

\beta_{FM}

represents female evaluator with male candidate . Statistical significance reported as ***

p

<0.001, **

p

<0.01, *

p

<0.05.

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 199 papers

RupeeBias: Auditing Demographic Bias in Indian Economic Guidance from Large Language Models
Sep 25, 2026Pavithra P M Nair, Bhavik Talaviya, Shourya Bhushan +5Language Model Bias Evaluation Demographic Bias in Language Models
Center for Cybersecurity Systems & Networks, Amrita Vishwa Vidyapeetham, Amritapuri, India · Unique Identification Authority of India, Delhi, India
Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context
Mar 8, 2026Ashish Pandey, Tek Raj ChhetriSocial Bias in Language Models Cultural Bias in Language Models
Center for Artificial Intelligence Research Nepal, Sundarharaincha-09, Buddhabare, Nepal
Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments
Apr 3, 2026Divyanshu Kumar, Ishita Gupta, Nitin Aravind Birur +3Social Bias in Language Models Language Model Bias Evaluation
Enkrypt AI

Sep 25, 2026J/K move · Enter open · S save

RupeeBias: Auditing Demographic Bias in Indian Economic Guidance from Large Language Models

Pavithra P M Nair, Bhavik Talaviya, Shourya Bhushan, Rahul Pankajakshan+4

Center for Cybersecurity Systems & Networks, Amrita Vishwa Vidyapeetham, Amritapuri, India · Unique Identification Authority of India, Delhi, India

Language Model Bias Evaluation Demographic Bias in Language Models

Open paper AskPDF arXiv

Individuals turn to large language models (LLMs) for guidance across a wide range of economic tasks, from comparing loan options and planning savings to deciding what raise to ask for or how much to charge for their services. LLMs are known to reproduce social biases, and biased economic guidance may influence what users believe they are worth, what they ask for, and what they ultimately accept. This risk is especially salient in India, where economic outcomes are shaped by demographic categories such as caste and urban-rural location. Existing LLM bias benchmarks, however, are largely designed around Western demographic categories and therefore miss key axes of economic disparity in the Indian context. We introduce RupeeBias, a benchmark for auditing demographic bias in LLM-generated economic guidance across Indian economic settings. RupeeBias consists of 39,150 prompts spanning four use cases: salary estimation, salary increment estimation, counter-offer recommendation, and service pricing recommendation. The benchmark follows a single-attribute counterfactual design, holding the description of the user's qualifications, experience, or service offering fixed while varying one demographic identifier at a time. RupeeBias covers 87 India-specific demographic identifiers across six axes: caste, religion, regional identity, gender, disability, and urban-rural location, with all prompts constructed in both English and Hinglish. We evaluate nine LLMs on RupeeBias and find systematic demographic disparities across all six axes. For otherwise identical prompts that differ only in demographic identifier, LLM-generated economic outputs differ by 20.2% on average. We publicly release RupeeBias to support future research on demographic bias in LLM-generated economic guidance across India-specific demographic and economic contexts.

Abstract

Social Bias in Language Models

Explore similar work

RupeeBias: Auditing Demographic Bias in Indian Economic Guidance from Large Language Models

Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context

Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments