Machine learning applications are increasingly being implemented in socio-technical safety-critical systems, but the safety of these applications is not well understood. This thesis used a mixed-method design and applied four methods to explore the field of machine learning and safety. With the results of these four methods, a framework was constructed for the risk analysis of machine learning applications. All findings were synthesised in two ways: by drawing conclusions about machine learning capabilities, and by drawing conclusions about organisational capabilities. This approach provided a way to think conceptually about how far we can take machine learning to increase safety in socio-technical safety-critical systems.