What it means
- Test what the AI actually decides or produces, and check whether outcomes differ unfairly between groups of people.
- For decision models: compare approval rates, error rates and pricing across groups (gender, age, region, language and similar).
- For GenAI: test prompts across groups and languages for stereotyped, offensive or lower-quality responses.
- For automated content moderation tools: check they do not wrongly flag some communities more than others.
- Done before go-live and periodically after, because fairness can drift.
Why it matters
- MeitY's 2024 advisory says AI must not permit bias or discrimination; the 2026 IT Rules require periodic review of automated detection tools for fairness and bias.
- RBI FREE-AI requires fairness and bias checks in product approval by a team independent of developers.
- Unfair outcomes in lending, insurance or hiring are the most visible and reputationally damaging AI failures.
Checklist
- Choose fairness metrics suited to the use case (for example approval rate ratio, false positive rate gap) and set tolerance thresholds.
- Define groups to test, using proxies where direct attributes are not collected, and document the method.
- Run fairness tests pre-deployment as part of SC-AI-015 Pre-deployment testing and validation.
- Have someone independent of the developers review results for medium and high risk systems.
- Record any gaps, root cause, and fix or justification.
- Re-test on a schedule (for example quarterly for high risk) and when monitoring shows shifts.
- Include automated moderation and detection tools in scope.
Evidence examples
- Fairness test report for the personal loan model: metrics by group against thresholds.
- GenAI bias prompt suite results across English, Hindi and two regional languages.
- Independent reviewer sign-off in the product approval record.
- Quarterly fairness re-test results and remediation log.
Common pitfalls
- Testing overall accuracy only; a model can be accurate on average and unfair to a group.
- Picking a metric after seeing results.
- No thresholds, so nobody knows what counts as a fail.
- Forgetting moderation and fraud tools, which also affect people.
Satisfies
- VoluntaryISO/IEC 42001
- VoluntaryNIST AI RMF and GenAI Profile
- BindingNYC Local Law 144 (AEDT)
- BindingTexas TRAIGA
- SupervisoryMeitY AI and Deepfake Advisories
- VoluntaryRBI FREE-AI
Also relevant to
- India AI Governance Guidelinesno numbered obligations; mapped in its practical notes (voluntary standards on fairness)
- ICMR AI in Healthcare Guidelinesno numbered obligations; mapped in its practical notes (fairness)
- 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 (pre-deployment bias scoring)
- 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
- 6.1.4 AI system impact assessment
- A.5.4 Assessing AI system impact on individuals or groups of individuals
- A.6.2.4 AI system verification and validation
ISO/IEC 27001
No direct equivalent
SOC 2
No direct equivalent
Key takeaways
- Test outcomes by group with pre-set metrics and thresholds, before launch and on a schedule.
- New for AI, not covered by ISO 27001 or SOC 2.
- Independence matters: someone outside the build team should review results.