> We don't have a number for the percentage of usage for what the user specifically types, but it sounds like definitely something we should have

It sounds like a violation of user privacy. Data-collection and privacy is a fine line, and I'm pretty sure any numbers related to specific text users type in private communication being surfaced up to humans is across that line.

I'm fine with my phone keyboard suggesting next words. I would not be fine with a human looking at the data model for specific sentences and words, even in aggregate.