Technical validity
Accuracy, reliability, robustness, calibration, failure modes and reproducibility.
AI Methodology
A public-interest method for evaluating AI before adoption, during use and after impact.
This framework helps teams examine an AI system as a complete socio-technical intervention: problem, data, model, user, decision, institution, risk, alternative, outcome and accountability.
Eight-stage assessment
A controlled path from claimed need to continuing oversight.
Select a stage to see its decision question and minimum evidence.
Stage 01
Validate the problem
Define the affected people, present process, decision need, harm, root causes and why an AI system is being considered.
EvidenceBaseline process and problem data.
QuestionIs AI relevant to the actual cause?
OutputScoped problem and decision statement.
GateProceed, reform process or stop.
Evidence dimensions
One score cannot represent every kind of AI risk.
Data validity
Provenance, relevance, representation, legality, quality and retention.
Human factors
Understanding, workload, overreliance, accessibility and ability to contest.
Distributional impact
Who benefits, who bears risk and how outcomes vary across groups.
Institutional readiness
Skills, authority, procurement, maintenance, incident response and ownership.
Public value
Real outcome, cost, opportunity cost, rights, trust and long-term dependency.
Decision record
Every deployment should leave an auditable answer.
| Field | Required answer | Evidence | Decision |
|---|---|---|---|
| Purpose | Exact user, task and decision being supported. | Problem baseline and process map. | Appropriate or mis-scoped. |
| Alternative | Non-AI and lower-risk options compared. | Cost, quality and feasibility comparison. | AI justified or unnecessary. |
| Performance | Results in relevant Indian conditions. | Disaggregated test and pilot evidence. | Acceptable, conditional or failed. |
| Rights | Privacy, consent, explanation and appeal. | Controls, notices and redress testing. | Protected or unresolved. |
| Outcome | Effect on people and public service. | Baseline, comparison and follow-up. | Scale, modify, pause or stop. |
Confidence label: distinguish verified evidence, partial evidence, modelled assumption, vendor claim and unverified observation.
Stop conditions
Some failures require pause, rollback or withdrawal.
Uncontrolled harm
Material safety, rights or discrimination risk without effective mitigation.
Uncontestable decision
Affected people cannot understand, correct or appeal a consequential result.
Performance collapse
Real-world accuracy or reliability falls below the approved threshold.
Purpose expansion
Data or model used for a new purpose without fresh review and authority.
Vendor dependency
Essential service cannot be maintained, audited, exported or safely exited.
No public benefit
Activity rises but the defined public outcome does not improve.
Evaluate an AI system with context, evidence and accountable decision gates.
Use the methodology for education, health, governance, infrastructure or another public-interest domain.