On the (Im) Possibility of Assumption-Free Inference for Causal Effects Estimated with ML