> even in scenario with a completely untrusted client receiving only final video output, there's no guarantee that the actions being input are human.

I imagine it would be a fairly trivial application of computer vision to make an aimbot for first person shooters that works most of the time.