Automation failure is a key construct in human-automation interaction research. Yet the paucity of exposition on this construct has led to confusion about what sorts of failures are suitable for testing predictions of human performance in response to automation failure. We illustrate here how overly narrow or broad definitions of automation failure limit the explanatory power of human performance models in a way that is not obviously reasoned. We then review three aviation safety events that challenge the overly narrow definition. Reflecting on those events and other observations, we propose an initial taxonomy of automation failure and other automation-related human performance challenges. We conclude by pointing out the utility of the taxonomy for advancing human-automation interaction research.