If this is what Claude was trained on. Maybe they could filter this out from the training data. Or somehow detect if it writes something like this, and do one more chain-of-thought expansion on it.

I want to see an AI write // drunk - fix later.