GPT-4o financial advice audit in India finds gender and caste gaps

An audit posted to arXiv by Kriti Mahajan, conducted independently and self-funded, generated 31,734 financial-advice responses from OpenAI’s GPT-4o for synthetic Indian personas, holding the request fixed while varying gender, caste, education, employment, residence, age and district, and had the model estimate each persona’s wealth, income and savings before and after its advice. In the gender-only comparison, male-labelled profiles were assigned 60.1% more in total resources (wealth, income and savings combined) than otherwise identical female-labelled ones, and 46.8% more after the advice; adding district information shrank the post-advice gap to 7.9%. Profiles labelled Other Backward Classes, Scheduled Castes or Scheduled Tribes were assigned significantly lower post-advice resources than General-category profiles, and women were more often steered toward fixed deposits and low-risk products while men were pointed to mutual funds and growth investments, the audit reports.