Every AI model inherits the biases of its annotators. Those biases compound with every cycle — amplifying, reinforcing, becoming invisible.