Something like, "AI generated images present a difficult challenge for applications that require that sensitive content be removed (e.g. Tumblr) or hidden (e.g. Discord) due to the fact that the abstract nature of the generation can trick most digital 'eyes'. To that end we've invented new measures based on already established content warnings and created a large corpus of prompts that have semantic relationships with those warnings. We've asked our labelers to measure content in the context of those content warnings; for example, an image might be labeled "is of a sexual nature", or "is of a violent nature." These responses are summarized as the scores on Figure 1. Going forward we will be making continuous improvements to our moderation API to identify such images with higher accuracy and future models will be tuned to generate them with lower frequency. We understand that a portion of our existing user-base desires the ability to generate sensitive images specifically but it is our goal as a company to build a product that can be deployed in a wide variety of applications without the need for human supervision. At some point down the road we may look into releasing models that cater to that use-case but at present our efforts are elsewhere."