Their keyboard predictions have gone from "OK" to "Amazing, we live in the future", and over the past couple years to "of course I didn't mean 'aaAAaAAnd', wtf were you thinking".

I frequently suspect they're starting to optimize more for $ than they were before, and ML just gives them more ways to make that number go up another % or so... but it often comes with impossible-to-predict and wildly inhuman edge cases. It's a pretty common trend when companies start focusing on small number increases - each A/B test shows improvement, but the product as a whole worsens and it drives people away in time.