It looks like upsampling is actually done by the AMHD app, not by the OS. Of course, there is no way to know for sure since the app is closed source but my research suggests AMHD app resamples all incoming streams to fixed rate, highest one supported by the OS, and this is done within the Amazon software itself. That said, OS audio mixer and stack could of course intervene downstream as well, no guarantee this would never happen.
The reason for this design, I assume, is to simplify maintenance of the portable code that is required to run across numerous versions of several very different operating systems, by unifying its API as much as possible. It is clear that bit-perfect reproduction never was a priority for Amazon developers, in which case performing fixed rate upsampling in the app does make sense.
I doubt this type of upsampling introduces audible artifacts anyway. Actually, the effect is most likely unmeasurable.