Room : P02

Artificial intelligence promises to make official statistics faster, more flexible, and better able to exploit various data sources. However, its adoption also raises fundamental methodological questions. What exactly does high quality mean in an AI-based statistical product? Do established ideas about validity, uncertainty, and statistical error need to be considered, or is predictive performance sufficient?

In this seminar, we discuss how artificial intelligence, specifically machine learning, is changing the way official statistics are created and evaluated in the future. As part of the Total Machine Learning Error (TMLE) framework, errors can arise as a result of the definition of the target population, the data and representations used, model development and validation, as well as model implementation.

A responsible adoption of artificial intelligence requires more than a better algorithm. It requires a broader approach to statistical quality from a methodological perspective.