Imagine you're part of a team deploying a conversational AI model that can reason across sensitive topics. During internal testing, you discovered that the model usually gives overly confident but factually wrong answers in high-risk contexts. How would you investigate and mitigate/solve that problem?