Microsoft is capturing text input and handwriting from touch input interfaces the same way Google's Android keyboard , Swype, or Swift Key do to improve predictive input, spelling correction/suggestions, personal dictionaries etc. Remember Windows 10 is touch enabled operating system like iOS or Android.

I'm not as familiar with the Apple Keyboard but I would be surprised if it didn't do that too in some form. Their Quick Type feature states it performs heuristics locally but it doesn't mention other aspects. Their policy states they don't collect personal information or conversation history but that doesn't mean they aren't capturing corrections or things they deem to be non-personal but sound pretty personal to me (e.g. occupation, language, zip code, area code, unique device identifier). Remember how researchers were able to identify Netflix users based on lesser anonymized meta-data? That non-personal information sounds pretty damn personal in comparison.

I hope you're not using a 3rd party keyboard on iOS with Full Access enable because if so then you agree to the same thing Microsoft is asking for.

I'm surprised you're ok with Apple recording Siri Queries and sharing them with Walk N’Talk Technologies who has humans listening to them but you're opposed to Microsoft doing anything similar in Cortona. How do you feel about Google Now? Did you ever use Google 411 because that was just a quick way to get a massive archive of audio samples.

Ever wonder how Google was able to catch up and perhaps surpass Microsoft, IBM, and everyone else in the Voice Recognition field so quickly? It wasn't because they came up with some revolutionary algorithm overnight. It was because they very quickly amassed an archive of transcribed audio samples. How did they do that? Very cleverly with Google 411.

If you ever used Google 411 you might have noticed it worked slightly differently from regular 411. You spoke your query, the voice recognition software spoke back what it thought you said and asked if that was correct. If you said no, or it couldn't understand your reply, it connected you to an operator who first listened to what you'd said and then repeated the confirmation process with you again whilst inputting what you actually said into the system. This created a transcribed audio sample that Google could use as a test case for their voice recognition software. This allowed them to iterate much faster than other companies.