In November 2015, Google introduced Smart Reply, a feature that suggested short responses to incoming email, first in its Inbox app. The research blog post announcing it was titled “Computer, respond to this email.” Smart Reply was, in Google’s words, “a deep neural network that writes email,” generating two or three candidate replies a user could send with a single tap.
Technically, Smart Reply was an applied sequence-to-sequence model: “a pair of recurrent neural networks, one used to encode the incoming email and one to predict possible responses.” The encoder turned the message into what the post called, borrowing Geoffrey Hinton’s phrase, a “thought vector” capturing the gist of the email, and the decoder generated likely responses from it. The engineer Anjuli Kannan led the work that turned this research idea into a production feature.
The team described a memorable failure mode while building it: the early model loved to suggest “I love you” as a reply, because short, common responses were safe bets when the model was unsure. They fixed it by penalizing responses that were merely common in general rather than specifically fitting the message. Smart Reply later expanded to Gmail itself and influenced the predictive “Smart Compose” feature.
For business readers, Smart Reply is an early, low-key example of generative AI inside a product millions used daily, years before ChatGPT made text generation a headline. It also shows the recurring challenge of generative systems defaulting to bland, safe output, and the engineering effort required to make suggestions actually useful.