A model trained with a frequentist objective using i.i.d. observational train and validation data is almost surely converging to the non-causal model of the underlying data generating process. That is because using short cuts/spurious correlations leads to a lower loss than not, i.e. you are actually specifically incentivizing your model not to be causal. "Emergence" believers' core fallacy is the assumption the causal model would have the lowest error, and therefore better optimization would eventually converge to the causal model. The opposite is true - the best scoring model (per frequentist objective) is actually the one that is best at exploiting spurious correlations.