It's called "constrastive negation", and both GPT and Claude tend towards it.