BACK TO SCHEDULE SPECIAL

Session 16

Quality Dimensions of Machine Learning and Artificial Intelligence in Official Statistics

4 June 2026
11:00 – 12:30
ŠIBENIK II

Presentation title
From Rules to Algorithms: Rethinking Quality Assurance in the Age of AI
The integration of Artificial Intelligence (AI) into statistical production is reshaping how official statistics are designed, produced, and communicated.

Read more Read less

This paper examines how algorithms in general, and AI specifically affects the quality assurance framework that underpins European official statistics and outlines areas for targeted adaptation of the European Statistics Code of Practice (CoP). It analyses how AI impacts the three CoP pillars — institutional environment, statistical processes, and output quality — introducing new risks such as opacity and dependence on external providers, while also creating opportunities for greater efficiency, accessibility, and timeliness. The paper further provides a list of existing or forthcoming resources already available to support the extension of the ESS quality framework, including guidance on legal compliance, organisational governance, algorithmic transparency, and AI model lifecycle management. It tentatively concludes by proposing guiding principles and adaptation pathways for the future evolution of the CoP, ensuring that European official statistics can harness AI’s potential while upholding their core values of trustworthiness, transparency, and sustainability.

Main author / Presenter
Jean-Marc Museux
Eurostat

Read more Read less

Jean-Marc Museux is a Senior Expert and Enterprise Architect at Eurostat, the statistical office of the European Union. He is actively involved in innovation and capacity building across the European Statistical System, with a particular focus on the reuse of non-conventional data sources and the integration of emerging technologies such as artificial intelligence in official statistics. With extensive experience in international cooperation and statistical modernization, he contributes to shaping innovation initiatives that enhance the quality and relevance of official data in an evolving data ecosystem. A statistical methodologist since 1997, Jean-Marc began his career at Statistics Belgium and has been with Eurostat since 2001. He holds a ph.D in Physics from the Free University of Brussels.

CO-AUTHOR:

Presentation title
From the CoP to a quality concept for AI/ML in official statistics and back
Nowadays, it is common practice to use artificial intelligence to produce official statistics.

Read more Read less

This often involves "traditional machine learning (ML)", i.e. mostly non-parametric statistical methods which, to name a few examples, separate relevant units from irrelevant ones in the context of data collection, contribute to the (partial) automation of classic processing steps such as classification and coding, assist us with record linkage and statistical matching, or enable further analysis of official data. The use of large language models in working directly on the data is also increasingly being discussed. The use of ML on the data has a direct impact on the quality of statistical processes and statistical outputs, for which the European Statistics Code of Practice and the Quality Assurance Framework (QAF) already provide principles (including sound methodology, appropriate statistical procedures, and accuracy and reliability) and associated indicators (e.g. that sampling errors and non-sampling errors are measured and systematically documented according to European standards). However, it remains unclear what these principles and indicators specifically mean for the use of ML in the production of official statistics: Which requirements from the QAF can be derived for ML? How can the requirements for processes and outputs be translated into requirements for ML in sub-processes? Based on preliminary work at the level of the UNECE and some ESS member states, a proposal for quality dimensions and quality guidelines for the ESS has been developed with the aim of enabling structured decisions to be made about the use of ML, balancing these quality dimensions.

Main author / Presenter
Florian Dumpert
Destatis

Read more Read less

Florian Dumpert heads a unit in the Federal Statistical Office of Germany (Destatis) that develops methodological and technological solutions and architectures for statistical production. The unit works on topics related to metadata, quality, digitisation, standardisation and automation of official statistics. His research interests include in particular machine learning, statistical data processing and imputation.

Presentation title
Technical perspectives on quality of the use of AI in official statistics
This contribution covers practical aspects to ensure that integrating AI and ML into statistical production leads to sustainable and transparent solutions.

Read more Read less

It highlights the technical building blocks and processes needed for robust data pipelines and ML-Ops, and the additional components required when extending to generative AI, including context provisioning, sound architectural choices, evaluation and monitoring, and human-in-the-loop structures that preserve auditability and confidentiality. These building blocks and ways of working require new skills across statisticians, methodologists, and IT. Examples include reproducible production chains, pipeline engineering, CI/CD and use of model registries, observability, and privacy and security practices. The contribution also shows how these building blocks and skills can raise the quality of official statistics even when traditional methods, rather than AI or ML, are used.

Main author / Presenter
Jakob Engdahl
Statistics Sweden

Read more Read less

Jakob Engdahl: Jakob is a strategist focusing on AI and innovation. He is involved in several expert groups on AI on national level as well as international initiatives within UNECE and ESS aimed at creating guidance on AI as well as reference architectures. Jens Malmros: Jens Malmros is a Methodologist at Statistics Sweden. His work concerns the use of machine learning and AI for official statistics production with a focus on processes and applications. Jens leads the machine learning group at Statistics Sweden. He holds a PhD in mathematical statistics from Stockholm University.

CO-AUTHOR:

Presentation title
The methodological view: current developments of the Total Machine Learning Error Model
As machine learning increasingly enters the domain of official statistics, it becomes essential to understand ML models not merely as algorithms but as statistical measurement instruments embedded in a broader inferential system.

Read more Read less

The Total Machine Learning Error (TMLE) Model provides such an understanding in a way that is familiar for official statisticians. It offers a principled framework for describing how errors arise, interact, and propagate throughout the life cycle of an ML-based process, linking model behaviour to representativity, population structure, data quality, and modelling assumptions, with the aim of achieving robust and reliable model performance.

Within the ESS quality context, the TMLE model can be seen as a foundational instrument for understanding and assuring the quality of AI-enabled statistical production. It provides the conceptual and analytical structure needed to integrate machine learning responsibly into official statistics while preserving transparency, interpretability, and trust.

In this presentation, some examples will be given to show how the TMLE helps to guard the high quality standards of official statistics.

Main author / Presenter
Marco Puts
Statistics Netherlands (CBS)

Read more Read less

dr. Marco Puts is a methodologist and coordinator of the research topic "applicable AI" at Statistics Netherlands. His main concern is how we can use Artificial Intelligence, and specifically Machine Learning, within the very strict setting of Official Statistics. Marco has a background in computer science and cognitive science.