That would require a lot of data privacy concerns to be addressed. Even if it’s an explicit opt-in. The current method uses sample text which can’t include PII. Using user supplied text would almost guarantee they’d get names and other PII in their data set.
I also imagine it’s harder to train the model when you don’t know exactly what the user was trying to type. I.e. Was the swipe detection wrong, or did the user delete the word because they changed their mind on what to write?
I’ve got a few old PCI cards around somewhere. I should pull one of them out and give them a try at this.