We present a procedure to automatically generate natural language descriptions of neurons in computer vision models. These generated descriptions support important interpretability applications: we use them to analyze neuron importance, identify adversarial vulnerabilities, audit for unexpected features, and edit out spurious correlations.