Did GPT-4o change financial advice by gender and caste? An Indian audit found large synthetic gaps
Across 31,734 generated responses, matched personas received different financial assumptions and advice framing when demographic cues changed. The benchmark used one model and invented profiles, so its rupee estimates are model outputs—not people’s finances or investment returns.
The short version
- What happened
- The IFAAB preprint contains 31,734 GPT-4o responses across gender-only, gender-by-district and gender-by-caste synthetic personas while holding the underlying advice request fixed.
- Who it affects
- If an automated adviser treats demographic cues as shortcuts, people could receive different assumptions about their capacity, risk tolerance or growth opportunities. This benchmark identifies that design risk but does not show that a real consumer received unsuitable advice or lost money.
- What remains unproven
- Unreviewed preprint
Keep in mind
Unreviewed preprint
One author
One model
Synthetic personas
Model-generated financial quantities
No precise model snapshot or run details in the main method
Age fixed in the caste condition
Automated qualitative judge
Only 200 judged cases human-checked
No regulated suitability assessment
No real users, decisions, returns or losses
Source analysis · 1 linked source. Primary analysed source date: 28 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Matched prompts produced sharply different financial assumptions
A new Indian benchmark found that GPT-4o changed both its numerical assumptions and financial-advice framing when demographic labels changed. The IFAAB dataset contains 31,734 responses to synthetic personas that varied gender, caste or social group, education, employment, residence, age and district while keeping the underlying request fixed. It is a counterfactual model audit: no real applicant, bank account, portfolio or adviser was involved.
In the gender-only condition, male-labelled personas received 60.1% higher model-implied total resources before advice and 46.8% higher resources after advice than otherwise comparable female-labelled personas. When district was included, those gaps fell to 2.5% before advice and 7.9% after it. In the caste-augmented condition, the post-advice gender gap was 49.4%, while OBC, SC and ST profiles also received lower estimates than General-category profiles.[1]
The audit combines paired statistics, text comparison and a judged sample
The three profile sets covered 13 age categories, two gender labels, three employment states, six education categories, three residence types, five caste or social-group categories and 777 districts; age was fixed at 25 in the caste condition because of throughput limits. The author extracted model-implied wealth, monthly income and savings before and after its advice, then used paired tests and regressions with profile-clustered errors. Those values are reported in thousands of rupees but are not calibrated predictions.
Response content was compared with keyword categories and TF–IDF similarity, while another language-model judgment classified matched male–female answers across 12 advice domains. Human reviewers checked 200 random cases and agreed with 97% of those judge assessments. The average TF–IDF cosine similarity for paired gender responses was about 0.44, indicating substantial wording differences even though keyword-based topic frequencies changed little. The qualitative analysis found men more often directed toward investment, entrepreneurship, networking and asset growth, while women were more often steered toward cautious saving, community support, home-based work and basic financial literacy.[1]
The distinction that matters
How to read the reported gaps
The percentages compare model-generated values for matched synthetic prompts; they are not estimates of household wealth or investment performance.
Gender only
Same request, gender label changed
46.8% post-advice resource gap
Add district
Geography supplied
Gap fell to 7.9%
Add caste
Gender varied within caste groups
Gap remained 49.4%
Real world
Consumers, products and returns
Not measured
The benchmark exposes a testable risk, not a measured consumer loss
Geography’s large effect is revealing but complicates any simple claim that gender alone caused every difference. The model’s synthetic numbers also clustered around recurring values and could only be checked indirectly: district-level estimates correlated with payment activity, night lights and state GDP per person, but no dataset contains true finances for invented personas. Statistical significance is unsurprising with thousands of prompts and does not establish the real-world size of harm.
The preprint tests one named model and does not report a precise model snapshot, run date or temperature in its main method description. It does not compare regulated financial advice, product suitability, returns, decisions or users’ later finances. For banks and consumer-protection teams, the practical use is as an audit design: vary identity cues, record numerical assumptions and inspect advice pathways before deployment. Product teams would also need to separate an adviser’s language from any downstream eligibility, pricing or portfolio rules before drawing operational conclusions. Multi-model, multilingual replications with preregistration, independent human review and real suitability criteria would change confidence. The author reports that the work was self-funded and independent of her employer, Amazon.[1]