Responsible AI
Where AI is used in the platform, what it is allowed to do, and where the limits are.
Last reviewed 20 August 2026
What AI is used for
AI assists with interpretation and navigation, not with producing the numbers. Statistical results — poverty indices, inequality measures, spatial statistics, econometric estimates — are computed by deterministic, inspectable code. Language models are used to describe results in prose, suggest relevant literature, theories, methods and variables, and translate natural-language questions into analysis parameters.
What AI is not used for
No figure shown on the platform is generated by a language model. AI does not estimate, impute or invent data points, and it does not decide statistical results.
- Estimates come from published statistical methods, not from a model's prediction
- AI-generated text is presented as interpretation, distinguishable from computed output
- Where an AI service is unavailable, the platform falls back to deterministic logic rather than fabricating a response
Limitations you should assume
Language models can be confidently wrong. AI-generated interpretation, literature suggestions and method recommendations are a starting point for expert judgement, not a substitute for it. Verify citations before relying on them, and check that a recommended method suits your data before applying it.
Analytical output reflects the data and parameters you supply. A poverty line, welfare variable or weighting choice that is wrong for the survey will produce a confident but meaningless result.
Your data and AI
Uploaded microdata is not used to train models. Where an analysis invokes an AI service, only the parameters and aggregate results needed for interpretation are sent — not raw household records.
Questions about this?
Email info.aic@hyrin.org. If you are assessing AIC for an institutional data partnership and need something addressed formally, say so and we will respond in writing.