The fact that MTurkers are paid makes them unreliable for reflecting genuine human responses.

That they can’t replicate indicates that the sampling population isn’t consistent representation of population demographics and so there’s no real signal there that can be applied to a general population like most failure to replicate studies