What it means
- Before training, check whether the data fairly represents the people the AI will serve, and whether it carries historic bias.
- Look at how groups are represented: gender, age, region, language, urban or rural, income band, and proxies that stand in for protected traits (PIN code, surname, device type).
- Check labels too: if past human decisions were biased, a model trained on them learns the bias.
- This is about the input data. Testing the model's outputs for fairness is SC-AI-017 Fairness and discrimination testing of outputs.
Why it matters
- RBI FREE-AI expects data that is relevant, fairly representative and ethically sourced.
- MeitY's 2024 advisory says systems must not allow bias or discrimination.
- A model trained mostly on metro, English-speaking customers will perform worse for everyone else, which is both a fairness and a business problem.
Checklist
- Define the groups that matter for each AI use case and which attributes or proxies to check.
- Profile training data: group counts, outcome rates by group, missing data by group.
- Compare training data distribution against the real customer or applicant population.
- Look for proxy variables strongly correlated with sensitive traits; justify or remove them.
- Document gaps and the fix chosen (more data, reweighting, exclusion of the use case for some segments).
- Repeat on each data refresh.
Evidence examples
- Data representativeness report for the loan model training set, with group breakdowns against the applicant base.
- Proxy variable analysis and decision log (for example PIN code dropped as a feature).
- Sign-off by the model owner and an independent reviewer.
Common pitfalls
- Not collecting any group data, then claiming you cannot check bias.
- Checking headline counts only, not outcome rates by group.
- Treating Indian diversity (language, region) as out of scope.
Satisfies
- VoluntaryISO/IEC 42001
- BindingEU AI Act
- SupervisoryPRA SS1/23
- BindingNYC Local Law 144 (AEDT)
- BindingTexas TRAIGA
- SupervisoryOSFI E-23
- SupervisoryMeitY AI and Deepfake Advisories
- VoluntaryRBI FREE-AI
Also relevant to
- ICMR AI in Healthcare Guidelinesno numbered obligations; mapped in its practical notes (representative training data)
- NITI Aayog Responsible AI for Allno numbered obligations; mapped in its practical notes (equality and non-discrimination)
- Tamil Nadu Safe and Ethical AI Policyno numbered obligations; mapped in its practical notes (diverse training data)
- AI (Ethics and Accountability) Bill 2025no numbered obligations; mapped in its practical notes (bias audits)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- A.5.4 Assessing AI system impact on individuals or groups of individuals
- A.7.4 Quality of data for AI systems
ISO/IEC 27001
No direct equivalent
SOC 2
No direct equivalent
Key takeaways
- Biased data makes biased models; check representation and labels before training.
- New for AI, not covered by ISO 27001 or SOC 2.
- Pair with output testing in SC-AI-017 Fairness and discrimination testing of outputs.