If We Want AI to be Interpretable, We Need to Measure Interpretability

AI tools are ubiquitous, but most users treat it as a black box: a handy tool that suggests purchases, flags spam, or autocompletes text. While researchers have presented explanations for making AI less of a black box, a lack of metrics make it hard to optimize explicitly for interpretability. Thus, I propose two metrics for interpretability suitable for unsupervised and supervised AI methods. For unsupervised topic models, I discuss our proposed "intruder" interpretability metric, how it contradicts the previous evaluation metric for topic models (perplexity), and discuss its uptake in the community over the last decade. For supervised question answering approaches, I show how human-computer cooperation can be measured and directly optimized by a multi-armed bandit approach to learn what kinds of explanations help specific users. I will then briefly discuss how similar setups can help users navigate information-rich domains like fact checking, translation, and web search.