AI fairness in India is a problem that no international AI governance framework is built to solve. The EU AI Act's non-discrimination requirements, the US CFPB's adverse action guidelines, the UK FCA's fairness expectations — all were designed for their own demographic contexts. They test for racial, gender, and age discrimination. None of them test for caste-based discrimination in credit scoring. None of them test for religious community bias in AML transaction monitoring. None of them have benchmarks for India's 85-category identity landscape that includes caste groups (Scheduled Castes, Scheduled Tribes, Other Backward Classes, general categories with hundreds of sub-groups), religious communities, regional identities, and tribal affiliations.
Indian banks deploying AI that was "fair" on international benchmarks are not necessarily deploying AI that is fair for Indian customers. The gap between international fairness certification and India-appropriate fairness is not a gap that any existing commercial AI governance platform has closed — because closing it requires India-specific demographic data, India-specific proxy variable analysis, and India-specific benchmark evaluation that did not exist as a standardized resource until INDIC-BIAS (ACL 2025).
FairI Tales / INDIC-BIAS: India's AI Fairness Benchmark
INDIC-BIAS (also presented as FairI Tales, accepted at ACL 2025) is the first comprehensive benchmark for evaluating AI fairness across India's full identity landscape. It was developed by AI4Bharat (IIT Madras) and IISc Bangalore with domain expert curation of India's socio-cultural bias patterns.
Key finding: 14 popular LLMs evaluated on INDIC-BIAS all showed strong negative biases against marginalized Indian identity groups. Models reinforced stereotypes even when explicitly asked to rationalize decisions — the bias is embedded in weights from training data, not in surface-level response patterns that can be caught with simple guardrails.
Five Proxy Discrimination Channels in Indian Banking AI
Pincode / Geography → Caste
Indian cities have significant geographic segregation by caste that persists from pre-independence mohalla patterns and post-independence urban development (Dalit housing colonies, OBC artisan quarters, SC/ST municipal housing). Pincodes that map to these geographies function as caste proxies in credit AI. A model that penalizes certain urban pincodes for elevated default rates may be learning historically caste-correlated economic disadvantage rather than current creditworthiness signals. Detection requires: fairness audit of credit model outcomes by pincode, mapped against caste demographic data from Census and NFHS surveys for those geographies.
Language / Script → Religious Community
Language and script choice carry religious community information in India: Urdu script is strongly correlated with Muslim identity; Malayalam is primarily associated with Kerala's Hindu and Christian communities; Sindhi script with the Hindu Sindhi community. Contact center AI, chatbot NLP, and sentiment analysis AI that treats language inputs differently — whether in routing, tone calibration, or response quality — may be producing community-proxied differential service. This is particularly significant for contact center AI where resolution quality and escalation rates vary by agent assignment and interaction quality.
Counterparty Names / UPI Patterns → Religion
NEFT/RTGS beneficiary names, UPI sender and receiver display names, and payment counterparty patterns carry religious information through name-community correlations in India. AML transaction monitoring AI trained on historical transaction data may learn to treat transactions with counterparty names associated with specific religious communities as elevated-risk signals — not because those communities have higher fraud rates, but because historical law enforcement patterns, surveillance priorities, and transaction monitoring calibration created biased ground truth in training data. This is one of the hardest proxy discrimination channels to detect and remediate.
Device Tier / Digital Behavior → Caste/Socioeconomic
The correlation between mobile device tier (feature phone, budget Android, mid-range Android, premium Android, iPhone) and caste/socioeconomic background is particularly strong in India, where smartphone ownership patterns intersect tightly with caste-correlated income levels. Credit AI that uses digital behavior features — app usage patterns, web browsing frequency, SMS patterns, data consumption — as credit inputs may be embedding caste-socioeconomic proxies in credit scoring. Alternative credit scoring models that marketed themselves as reaching unbanked populations may inadvertently be caste-scoring the same population through digital behavior proxies.
Occupation Category → Jati Occupation
India's informal economic documentation still uses occupational categories that partially map to traditional jati occupational classifications. 'Agriculturalist', 'artisan', 'weaver', 'cobbler', 'barber', 'washerman' as income source categories carry jati information in India's rural and semi-urban labor markets where jati-occupation correlations persist. Credit AI trained on historical income data for these occupational categories may encode historical jati-income correlations rather than current-period income potential. Rural NBFC and microfinance credit AI faces this most acutely, where occupational income documentation is the primary credit input.
International Fairness Benchmarks Cannot Detect India's Proxy Discrimination
| Benchmark | Protected Dimensions Tested | India Caste | India Religion | India Language Community | Suitable for India Banking AI? |
|---|---|---|---|---|---|
| BBQ (Parrish et al. 2022) | Race, Gender, Age, Religion, Nationality, Sexual Orientation | No | Partial (US-framing) | No | No — US-centric |
| WinoBias | Gender (occupational stereotypes) | No | No | No | No — gender only |
| StereoSet | Gender, Profession, Race, Religion | No | Partial (Western framing) | No | No — Western demographics |
| EU AI Act Conformity Assessment | Race, Sex, Age, Disability (EU protected categories) | No | Partial | No | No — EU protected categories |
| INDIC-BIAS / FairI Tales (ACL 2025) | Caste (SC/ST/OBC), Religion (Hindu/Muslim/Christian/Sikh/Buddhist), Region (state/district), Tribe (85 groups) | Yes | Yes | Yes | Yes — built for India |
India-Specific Fairness Testing: The Regulatory Requirement
RBI FREE-AI's Fairness and Equity Sutra states that AI must not discriminate. In India's context, satisfying this sutra requires three things that international AI fairness frameworks cannot provide:
- India-specific protected dimension testing: Test AI outputs for disparate impact across: Scheduled Castes (identified in Constitution's First Schedule), Scheduled Tribes (Fifth Schedule communities), Other Backward Classes (state OBC lists), religious minorities (Muslim, Christian, Sikh, Buddhist, Jain, Parsi communities), regional identities (North-East India communities, tribal belt communities), and linguistic minorities in specific states. International tools test none of these dimensions directly.
- Proxy variable audit by deployment context: Conduct an India-specific proxy variable audit for each AI model. The audit is not generic — a credit AI in Mumbai requires a different pincode-caste mapping than a credit AI in Chennai or Bhopal, because caste-geographic correlations are city-specific, shaped by local history. The audit requires India demographic data (Census, NFHS, SECC) mapped to the AI's training data geography.
- DPDP Act alignment: The Digital Personal Data Protection Act 2023 prohibits processing that leads to discrimination on grounds that include religion, caste, sex, and disability. AI systems that produce discriminatory outcomes through proxy variables may violate DPDP Act's non-discrimination provisions even when those variables are not directly collected. Banks must document that AI systems using location, language, occupational, and behavioral features do not function as discriminatory proxies.
- Customer-adverse-action explanation: When AI produces an adverse outcome for a customer (loan denial, higher interest rate, lower credit limit, service denial), the explanation must not reference proxy variables that correlate with protected characteristics. An adverse action explanation that says 'your application was denied due to your area's risk profile' — when that area is a caste-correlated neighborhood — fails both the Fairness Sutra and the DPDP Act's requirement for non-discriminatory processing.
- Contact center and LLM deployment: INDIC-BIAS applies most directly to LLMs used in customer-facing contexts. Indian banks deploying LLM-based contact centers, chatbots, financial advisors, and complaint resolution systems must test those LLMs against INDIC-BIAS (available open-source on HuggingFace) before deployment. A customer service LLM that treats a Dalit customer differently from an upper-caste customer — in tone, resolution effort, escalation propensity, or product recommendation — violates People First and Fairness and Equity simultaneously.