Abstract

Ethical Governance and Accountability in the Use of AI in Official Statistics

AI is moving into the operating model of national statistical offices, and not only into their production processes. Most of the current debate concentrates on models and methods. The contribution addresses a different question, and it is organisational: when an AI system has contributed to an official figure, who answers for it?

Responsibility cannot be moved onto a model, a supplier or a methodology. It stays with the organisation, and it starts at the top. Leadership sets the tone, decides what is acceptable before anyone is asked to decide it alone, and owns the consequences. Below that level, responsibility has to be visible in the job descriptions and competence requirements of the people involved.

Two reference texts frame the discussion from opposite ends. Ethics in Official Statistics (ECE/CES/STAT/2025/6) comes out of a UNECE task team that surveyed more than fifty statistical offices, building on earlier HLG-MOS work on risk management and human resources. Its central distinction is between what the law permits and what an office should actually do, and it works outwards into the instruments that make that distinction operational: values statements, codes of ethics and conduct, due diligence, whistleblowing channels, and a risk management process applied to ethical risk.

The HLG-MOS Framework on Responsible AI for Official Statistics (2025) takes the technology as given and asks what has to hold around it. It sets out six guiding principles — ethical purpose and public good, accountability, transparency and explainability, fairness, privacy and security, validity and robustness — and attaches to each concrete guidelines, from assigning ownership at every stage of the lifecycle to justifying in writing the choice of an opaque model over an interpretable one.

What the documents do not settle is whether their principles survive contact with the instruments an NSO already uses, such as ISO/IEC 42001:2023 or the risk cycle of ISO 31000 and COSO ERM, without ethics collapsing into a compliance checklist.

The organisational risks are concrete. Ownership across an AI lifecycle is one: many functions contribute and it is often unclear who holds the outcome. The boundary between methodological authority and institutional responsibility is another. A third concerns human resources, where skills, workload and the division of labour shift faster than roles and training are updated, and where staff may be asked to validate outputs they have not been equipped to question. A fourth is whether human validation is still judgement, or has quietly become endorsement.

Behind all of this sits the problem of ethical dilemmas. A dilemma is not a mistake and not a breach. It is a situation in which two defensible values pull in opposite directions, and in which the law may well permit something the organisation should still decline to do. Statistical offices meet these situations regularly and often resolve them informally, which means invisibly. Whether an NSO can recognise them, surface them and deliberate them is the real measure of its ethical capacity as AI spreads.