LLM hallucinations remain an unsolved problem. One interesting observation made by some Anthropic Fellows was that the more capable the model gets, the less the errors correlate with bias, which makes the errors they produce very hard to predict [1]. This could be part of why LLMs are not able to ground knowledge - they themselves have no means to effectively evaluate whether the information they're providing is true or false.
What this means in practice is that for any kind of high-stakes application, you as the author should always do a quick glance over any materials you are publishing before you release a final version.