The research design of machine learning integrates a whole series of design decisions, each of which both enables and restricts the output. Many of these decisions concern trade-offs. ML applications basically frame whatever they ‘perceive’ and thus also frame the behaviour of the application. In this talk she will discuss the productive nature of the bias that is inherent in machine learning development, and its relationship with unfair and unlawful bias in subsequent applications. She will discuss this in light of the resilience, robustness, reliability and responsibility of AI applications and the need to better communicate about both the inherent and the problematic aspects of bias in ML.