It has a lot of drawbacks, that make it easily noticeable that something is not right. (Browser autocompletion not working for example.) Though it's an interesting concept.
Interfering with focus might be more noticeable, although Windows 10 has made this much less obvious than it used to be.
It does rather drive home the extent to which the "desktop" model is at odds with the user being able to run multiple mutually hostile applications from different sources. And sometimes the ability to read and inject keyboard input is legitimate.
Interfering with focus is what breaks autocompletion. (This drawback is explained in the paper itself.)
IOW, web designers have already set the precedent for broken autocomplete, no keylogger required.