As someone who regularly purchases work from MTurk, some rando's crumby bot is absolutely not what I'm looking for. We're training our own AI from MTurk responses, and we're sending to MTurk difficult edgecases that our own AI gets wrong.
Such a combination is what Unbabel does in translation. ML-driven automated translation followed by crowd-sourced human review and corrections. It's a neat idea and could be the future of translation (as opposed to full automation).
I would assume that Amazon already does this. If a task is trivial enough for an AI, they would do it themselves and get the posters money before the job gets forwarded to the Turks.
Sure, Amazon have all sorts of ml APIs. But a task platform where anyone can complete the tasks using ai, humans or a combination that f both could be substantial impressive.
Say you need to look through 20k images and determine if they contain hipsters. There may not be an off-the-shelf hipster detector. Maybe you could make one, or use humans or a combination by training an NN.. etc.