(Edit: it looks like you CAN turn it off in the settings.)
Here's some info on how it works if it helps address your concern: We've transformed the language generation problem into a labelling problem. From a systems point of view it's not any different from spam filtering, or tabbed inbox, and we use the exact same features. Smart Reply uses a whitelist of around 30k independently generated & sanitized phrases. You could think of each whitelist entry as a label. So when an email comes in on the delivery pipeline we just label things, like we've always done. We label (spam/not-spam) for spam filtering, we label (promo/social/update/forum) for tabbed inbox, and we label (Thanks!/Sounds good!/etc/etc) for smart reply.
Do you record what responses people choose? If so is that data used to personalize the responses or just stored in aggregate?