Abstract

Responsible AI is often discussed in terms of fairness, transparency, accountability, robustness, and trustworthiness. The question, however, is how these ambitions can be achieved in practice. In this presentation, we argue that Responsible AI requires a rigorous methodology that connects high-level principles to concrete choices in data, modelling, evaluation, implementation, and monitoring. A methodological framework helps to identify these concrete choices by identifying the problems, assess their consequences, and determine what is needed to make AI systems reliable, valid, and fit for purpose.
Using examples from applied AI in official statistics, we show how the Total Machine Learning Error framework can help close the gap between Responsible AI as a guiding principle and the resulting implementation.