Not What I was Trained for – Out-of-Distribution-Tests for Interactive AIs

Recent advances in AI demonstrate its capacities to not only automate more and more everyday tasks, but also make direct Human-AI-Interaction possible. When using AI models in production, we might face situations where the AI is confronted with input data that is very different from the data it was trained on. Such situations are called Out-of-Distribution situations and can result in misleading AI inferences. We argue that identification, handling, and prevention of Out-of-Distribution situations is key for creating production-ready interactive AI components. In this paper, we test the robustness of state-of-the-art AI/ML approaches in Out-of-Distribution situations and propose a research agenda to gather a deeper understanding of how to identify, handle, and prevent such situations in interactive applications.