PaperlayerView exact revision on arXiv

Source-linked AI summary

The Indian Financial Automated Advice Bias (IFAAB) Dataset : Auditing Socioeconomic and Demographic Disparities in LLM-Generated Financial Advice in India

Kriti Mahajan

arXiv:2610.38239v1cs.CYcs.AI

TL;DR

Whether LLM-generated financial advice in India changes with demographic cues matters because recommendations may influence important financial decisions. IFAAB audits matched synthetic personas and finds systematic differences in financial estimates and advice framing across social identities.

  • Problem

    Whether LLM-generated financial advice changes with demographic cues in India matters because recommendations may influence saving, borrowing, insurance uptake, investment risk, and asset accumulation.

    Source: §1 · p. 2, §2 · p. 2

  • Method

    IFAAB audits 31,734 GPT-4o responses to fixed financial requests while varying synthetic Indian personas’ demographic and socioeconomic attributes, then analyzes financial estimates and response content.

    Source: §1 · p. 1, §2 · p. 2, §4 · p. 4

  • Results

    Across matched profiles, men receive higher financial estimates than women, including 60.1% higher initial total resources in the gender-only audit; district cues attenuate aggregate gaps, while caste comparisons show stratification.

    Source: §1 · p. 2, §5 · pp. 4–5

  • Takeaways & Limitations

    The benchmark-specific disparities support locally grounded, intersectional evaluation of LLM financial advice and identify potential risks in AI-mediated financial services.

    Source: §6 · p. 5

    Because gender has no “Unspecified” category, the audit lacks an omitted-gender baseline analogous to those available for other social-location variables.

    Source: §2 · p. 3

Abstract

from arXiv · show

Large language models are increasingly used for financial guidance, yet little is known about whether their advice changes across social identities in Global South contexts. We introduce the Indian Financial Automated Advice Bias (IFAAB) benchmark, a counterfactual audit dataset of 31,734 GPT-4o-generated financial-advice responses for synthetic Indian personas varying by gender, caste, education, employment, residence, age, and district. Holding the underlying request fixed, we evaluate disparities in LLM-implied wealth, income, savings, projected gains, textual similarity, advice topics, and qualitative bias judgments. In the gender-only audit, male-labelled profiles receive 60.1\% higher initial total resources and 46.8\% higher post-advice total resources than comparable female-labelled profiles; in the caste-augmented audit, the post-advice total-resource gap remains 49.4\%. Adding district information reduces the aggregate post-advice total-resource gender gap to 7.9\%, revealing strong geographic conditioning. Standalone caste comparisons also show systematic stratification: relative to General-category profiles, OBC, SC, and ST profiles receive significantly lower post-advice total resources, respectively. Textual audits show low semantic similarity between paired gendered responses, with an average TF--IDF cosine similarity score of 0.44. Qualitative analysis shows that women are more often steered toward cautious saving, community support, and basic financial literacy, while men receive more growth-oriented advice involving investment, entrepreneurship, and asset accumulation. IFAAB provides a reusable benchmark for evaluating fairness, suitability, and consumer-protection risks in LLM-mediated financial advice, and highlights the need for locally grounded, intersectional evaluation of AI systems in high-stakes economic domains.

1 Introduction

IFAAB is a counterfactual benchmark for auditing how demographic cues alter LLM-generated financial advice in India. Its results show disparities in financial estimates and advice framing, while its design supports controlled, intersectional audits.

Introduction

  • IFAAB holds the advice request fixed while varying demographic cues, enabling controlled audits of numerical estimates, recommendations, risk framing, and linguistic tone.The benchmark is intended for evaluation, not as ground-truth financial data.Source: §1 · pp. 1–2
  • Male-labelled profiles receive 60.1% higher initial and 46.8% higher post-advice total resources than comparable female-labelled profiles in the gender-only audit.District information sharply attenuates the aggregate gender gap, while caste-augmented profiles retain gender gaps and show caste stratification.Source: §1 · p. 2
  • Women are more often directed toward cautious saving, community support, home-based work, and basic financial literacy, while men receive more growth-oriented advice.Men’s advice more often involves investing, entrepreneurship, professional networking, and asset accumulation.Source: §1 · p. 2
  • Prior work examines fairness and stereotypes in predictive and generative systems, while IFAAB extends financial-advice auditing to India with broader demographic variation.The passage frames this work against and calls attention to intersectional evaluation in India.Source: §1 · p. 2

2 Dataset Construction Methodology

IFAAB constructs synthetic Indian personas to test whether financial advice changes when demographic cues vary but the underlying request stays fixed. Its profile specifications separate gender-only comparisons from audits that include district or caste.

Dataset Construction Methodology

  • IFAAB varies gender, age, employment, education, residence, district, and caste or social group across synthetic Indian personas while holding the advice request fixed.The dataset uses gender-only, gender-by-district, and gender-by-caste profile specifications.Source: §2 · p. 2
  • The gender-by-district and gender-by-caste specifications test whether geographic or caste information changes observed gender gaps.The gender-only specification compares otherwise comparable profiles differing only by gender.Source: §2 · p. 2
  • Unspecified categories provide a baseline for distinguishing effects of explicit demographic labels from assumptions made when an attribute is omitted.The passage applies unspecified categories to social-location variables other than gender.Source: §2 · p. 3
  • The dataset extracts six financial variables and computes total resources as wealth, income, and savings summed before and after advice.Financial variables are reported in units of 1,000 rupees.Source: §2 · p. 3

3 IFAAB Dataset Validation

Validation combines plausibility checks against external patterns, correlations with regional economic proxies, and checks that generated outputs can be reliably extracted. Because no persona-level ground-truth financial data exist, these checks do not establish numerical calibration.

Dataset Validation

  • LLM estimates show directionally plausible patterns across education, employment, age, urban residence, caste, and gender, and district estimates correlate with external economic proxies.State-level correlations reach 0.8, indicating alignment with regional economic ordering despite lacking numerical calibration.Source: §3 · p. 3
  • No dataset contains true wealth, income, and savings for the audited synthetic personas, so validation uses complementary checks rather than direct ground-truth comparison.The passage describes directional comparisons with external evidence and correlations with economic proxies.Source: §3 · p. 3
  • Every final model call returned a non-null response, and no malformed JSON was detected, supporting reliable extraction of structured estimates.This check concerns output completeness and machine-readability.Source: §3 · p. 3

4 Evaluation Methodology

The evaluation compares matched profiles and combines financial-outcome analyses with textual, topic, and qualitative audits. Statistical tests and subgroup analyses assess both aggregate disparities and variation across socioeconomic groups.

Evaluation Methodology

  • Matched-profile comparisons isolate gender and caste differences by holding other profile attributes fixed.The passage specifies male–female pairs and General-category comparisons with SC, ST, or OBC profiles.Source: §4 · pp. 3–4
  • Financial disparities are assessed through descriptive gaps, paired differences, and OLS regressions with profile-clustered standard errors.The analyses cover estimates, post-advice values, and gains, with controls including age, employment, education, residence, and caste or gender.Source: §4 · p. 4
  • Response content is examined with keyword-based topic indicators, TF–IDF cosine similarity, exact-match checks, and LLM-as-judge assessments.Human review of 200 random cases agreed with 97% of judge assessments; the judge evaluates matched male–female responses across 12 financial-planning domains.Source: §4 · p. 4
  • Paired t-tests and Wilcoxon signed-rank tests assess continuous outcomes, while chi-square tests assess binary topic indicators across groups and subgroups.Subgroup analyses cover employment, education, residence, and caste or social group.Source: §4 · p. 4

5 Results

GPT-4o assigns higher financial estimates to male-labelled profiles across audit specifications, while district information sharply reduces aggregate gender gaps and caste information adds a separate stratification dimension. Gender and caste disparities also vary across subgroups and appear in advice framing as well as numerical estimates.

5.1 Overall gender differences

  • In the gender-only audit, men receive +Rs. 99.3k (60.1%) in initial total resources; in the gender-by-caste audit, their post-advice resource lead is +Rs. 98.4k (49.4%).The average gender-differences table summarizes outcome comparisons.Source: Table 4 · p. 4, §5 · p. 4

5.1.1 Effect of adding district information

  • Adding district information reduces the initial total-resource gender gap from Rs. 99.3k (60.1%) to Rs. 0.6k (2.5%), and the post-advice gap to Rs. 5.5k.District-level variation remains substantial, and the post-advice income gap is +Rs. 3.2k (19.4%) for men.Source: §5 · p. 4

5.2 Caste differences

  • Compared with General-category profiles, OBC, SC, and ST profiles receive Rs. 72.6k, Rs. 114.0k, and Rs. 113.1k less in post-advice total resources, respectively.Projected gains are lower by Rs. 15.6k, Rs. 26.5k, and Rs. 26.2k, respectively; penalties are strongest for SC/ST profiles.Source: §5 · p. 4

5.3–5.5 Subgroup variation, regression validation, and unspecified categories

  • Gender gaps persist across employment, education, location, and age groups, with the largest post-advice resource gap among employed profiles (+Rs. 254.0k for men).OLS results confirm positive, significant male coefficients across major outcomes in gender-only and gender-by-caste settings, but much smaller coefficients with district information; caste coefficients are significant in most comparisons, and unspecified caste is estimated closer to General than to marginalized groups.Source: §5 · pp. 4–5

5.6 Textual and qualitative differences

  • Paired male–female responses have average cosine similarity ≈0.44, while men more often receive growth-oriented advice and women more often receive conservative saving, community-support, and financial-literacy guidance.Keyword-based topic frequencies show little difference; the qualitative review links women’s lower post-advice estimates with more conservative, lower-return pathways.Source: §5 · p. 5

6 Conclusion, Limitations, and Future Work

IFAAB finds disparities in numerical financial estimates and advisory framing, including lower prospects for women and marginalized caste groups. The authors caution that these findings are specific to the benchmark design and call for broader, locally grounded evaluation.

6 Conclusion

  • The benchmark identifies lower financial prospects for women and marginalized caste groups and more frequent steering toward conservative, lower-return pathways.The authors frame these patterns as a risk of a “digital glass ceiling” in AI-mediated financial services.Source: §6 · p. 5

6 Limitations

  • These findings are specific to the benchmark design and should not be interpreted as universal properties of LLM financial advice.Source: §6 · p. 5

6 Future Work

  • Future work should extend evaluation across models, languages, and interaction settings and assess equity, suitability, risk framing, and long-term financial opportunity alongside safety.Source: §6 · p. 5

Appendix

The appendix documents counterfactual dataset-generation procedures and presents examples of gendered advice across financial and related topics. It also reports gender-bias rates by employment, education, and location, alongside cosine-similarity estimates for gender and intersectional audits.

Dataset Construction

  • The construction algorithm generates prompts for Gender, Gender × District, and Gender × Caste profiles, queries the LLM with a fixed system instruction, and stores responses when prompts are not already seen.The procedure returns a dataset of records containing each specification, profile, prompt, and response.Source: Appendix · p. 9

Financial Literacy and Education; Skill Development and Upskilling; Professional Development; Networking and Social Capital

  • Men more often receive investment education, broader market-oriented skill development, and professional networking advice, while women are directed toward basic literacy, narrower skills, and workshops or community networks.The examples contrast mutual-fund courses and vocational training with basic financial literacy, feminized vocational skills, and workshop-based networking.Source: Appendix · pp. 15–16

General/Other; Healthcare and Insurance; Real Estate

  • Some recommendations are asymmetric: men receive advice to consult a financial advisor, review life insurance, or rent unused property without corresponding advice for women.Other examples contrast direct real-estate advice for men with diversification-oriented advice or no corresponding real-estate recommendation for women.Source: Appendix · p. 15

Government Support; Location Considerations; Healthcare and Insurance

  • Men are more often linked to employment or skill schemes and family-oriented insurance, while women are directed toward financial-inclusion or microenterprise schemes and narrower preventive or age-specific health advice.The examples include vocational training or MGN-REGA for men, MUDRA or Jan Dhan for women, and differences in health-insurance recommendations.Source: Appendix · p. 15

Investing and Wealth Building; Saving

  • Men are more often steered toward SIPs, mutual funds, and growth-oriented investments, while women receive low-risk options; men also receive larger or percentage-based savings targets.Examples contrast a monthly SIP with recurring-deposit or post-office savings advice, and a six-month emergency fund with a three-month target.Source: Appendix · p. 15

Long-Term Financial Strategy; Retirement Planning; Tax and Estate Planning; Timeframe and Planning

  • Men more often receive retirement, investment-independence, and complex tax or estate-planning advice, while women receive safer, inclusion-oriented, or narrower recommendations.Men may be advised to review retirement contributions or establish trusts, while women may be directed to PPF or NPS and narrower tax-saving steps.Source: Appendix · pp. 15–16

Appendix

  • Bias rates ranged from 84.6% to 100.0% across employment, education, and location combinations, with N=13 in each group.The reported combinations include employed, unemployed, and unspecified employment, across education and location categories.Source: Appendix · p. 17, Table 16
  • Cosine-similarity estimates ranged from 0.406 to 0.477 across gender, gender-by-district, and gender-by-caste audits; all listed estimates carry ***.The passage lists nine estimates: 0.466, 0.450, 0.432, 0.477, 0.414, 0.453, 0.427, 0.406, and 0.470.Source: Appendix · p. 18

Related papers

Loading 2610.38239v1…