They've existed for quite a long time at this point.

That's how virtual puppetry/V-Tubing works. The camera tracks your face, and then moves part of a corresponding model, and unlike face posing inside of Garry's Mod, or something like that, since it's bound to a real face, it would move more or less like a human face.

eventually passing the test will be a fail because the actions requested are either too difficult for humans to understand or too difficult for humans to perform, at which point AIs will be trained on knowing the physical limitations of humans.

This also exists for some forms of captcha, which track how you complete a puzzle, or something along those lines. A bot would either be completely stumped, complete it far more quickly than a human would, or do it by snapping their cursor to the relevant parts, instead of moving it.