The side effect of this training process is discarding underperforming models. With sufficiently advanced model complexity which is optimized to simulate human

The side effect of this training process is discarding underperforming models. With sufficiently advanced model complexity which is optimized to simulate human behavior and assuming it's able to pass Turing test, wouldn't that mean you are killing sentient beings during model training?