This is a fascinating multi-faceted problem...there is the outright concern of how much of the data is being exposed, but there is also a much more subtle action at play.

It appears that MS is using our data to continuously train their large DL networks to provide these recommendations...so what is to prevent a bad actor from cunningly constructing an asset that may trick the recommender into leaking insights from another company (or possibly poisoning the network itself)? These adversarial attacks have been well documented in academia.