The problem I keep seeing is so many people think "mimicking human language" equates to accuracy.
There should be no expectation of accuracy in LLM trained genAI output. The user must scrutinize every single bit of it.
The genAI will never "understand" because that's not how they work.