You can't fix it with a pure LLM, no. But very few of the major models are pure LLMs anymore. They've got LLMs built in to them as subsystems, but they've got control systems layered on top of those to adjust their behaviors. And those don't have to be trained purely with next word prediction.