With the largest pile in the world of written natural language representing dynamic human communication, Google is able to use deep learning, combined with the

With the largest pile in the world of written natural language representing dynamic human communication, Google is able to use deep learning, combined with the methods of Natural Language Processing (NLP) to draw inferences about the essential content and intent of your emails and sort them into types. Using these same machine learning methods, a trio of possible responses to a given type of email message is generated and tested. Every time a human uses one of these three responses, a datapoint is supplied to Google that says “given message type A, a human has chosen response X as an appropriate one.” Multiply this last step by a thousand or a million or a gazillion, to the point where a clear statistical pattern emerges, and Google can conclude with some confidence that when a person expresses ideas, thoughts, feelings, or questions that can be classified as type A, it’s reasonable for another human to respond with utterance X. This is usual attribution of magical powers to "AI" (and really they mean "ML") by the people who don't know how it actually works. In this case: * Sensitivity detector checks if email is sensitive, smart reply is turned off for sensitive emails (obviously didn't trigger in this case). I personally think this is questionable feature - imagine your keyboard turns off if it thinks topic is sensitive. * Encoder model produces embedding which is compared with pre-computed and pre-clustered embeddings for whitelisted replies. Top replies for three top clusters are selected (selecting replies without clustering will produce very similar replies). All potential replies are white-listed for obvious reasons. While clicks produce useful metrics (really, this is the target metric), they don't produce useful training data for two reasons: (roughly speaking) you can't train a model on it's own output and there is no way for model to learn to suggest something new. Fundamental challenge for training such models is that production metric is different from training metric: model is trained to select likely reply that was typed, production metric is reply that is clicked. There is bunch of blog posts and papers from Google Research describing how it works.