Even GPT-4 is not actually that good at it, and breaks down past a few simple rules. If you want an LLM to actually learn a language, today, you have to include enough of that language in its training data to begin with (I'm not sure if anyone tried it with fine-tuning).
Has that trend taken over more established conlang? Would it make sense for those communities to train an LLM to speak that language as fluently as possible to help materialize it: build lessons, have a conversation partner at any time, and possibly discuss grammar?