Responsible AI
AI at the Data Lab works under human oversight matched to the task. Where AI drafts or explains, experts decide what to use. Where it classifies millions of records, experts check the cases that matter most. Responsible AI is not a separate activity for us. It shapes how each tool is designed, tested and used.
- Oversight matched to the task: experts review drafts, and check the high-impact cases of large-scale classifications
- Grounded and traceable: answers rest on data and evidence, with sources cited
- Data kept in-house where needed: open-weight models on FAO infrastructure
Our work follows the Fundamental Principles of Official Statistics and the Principles for the Ethical Use of Artificial Intelligence in the United Nations System.
Principles in practice
| Principle | What it means | How it shows in our tools |
|---|---|---|
| Human oversight, matched to the task | The level of human review fits the scale and impact of each task | Analysts edit and finalise SCOPE Report Generator drafts. Experts rate the categories the AI assigns to FAO projects. In the classification of official development flows, which covers millions of projects, experts check the largest projects, where a misclassification would most affect the totals. The LUMINA-AI assistant explains results and evidence; decisions stay with its users |
| Grounded outputs | Where AI generates text, it rests on data and evidence, not on the model's general knowledge | Indicator descriptions state only what the data show, with no causal claims. The LUMINA-AI assistant says when no evidence supports an answer |
| Transparency and traceability | Users can see where every statement comes from | Each indicator carries its data source; document-based answers cite their passages; the project classifier records a written rationale for every category; the LUMINA-AI assistant cites the literature it draws on |
| Validation | Outputs are checked against data, rules and expert review | Automatic checks catch number-formatting errors independently of the AI model; analysts review drafts before use |
| Clear labelling | Users always know when content is AI-generated | Reports from the SCOPE Report Generator are flagged as AI-generated and to be verified |
| Separation of roles | AI explains; statistical models estimate | In LUMINA-AI, every risk figure comes from the Bayesian models, and the AI assistant explains them |
Privacy and security
We match each tool's deployment to the sensitivity of its data. Where appropriate, the Data Lab runs open-weight large language models on approved virtual machines within FAO-controlled IT infrastructure. This brings the model to the data: information is processed and stored inside the FAO environment rather than sent to an external LLM provider. For example, FAO project documents are classified this way in our technology and innovation mapping work. This approach gives FAO greater control over data flows, model versions and access, while still requiring appropriate approval, access controls and security safeguards.
From practice to organizational governance
The Data Lab contributes practical experience from designing and applying AI in statistical and analytical workflows. Its work turns responsible AI principles into concrete methods, including human-in-the-loop review, evidence-grounded outputs, traceability and validation, and provides lessons that can inform wider guidance and safeguards.
At the organizational level, FAO's cross-divisional AI Coordination Team (ACT) supports the operationalization of FAO's Responsible AI principles. It develops guidance, supports the governance framework and the risk and impact assessment process, maintains visibility across AI initiatives, and promotes learning and responsible innovation. The Data Lab contributes its statistical, methodological and implementation expertise to this collective effort.
Building institutional capacity
Adopting AI well depends on people and skills as much as on technology. The Data Lab co-organised the webinar series AI Fundamentals, with the FAO AI Community of Practice, where non-technical colleagues to AI, including the governance of responsible AI adoption.