A language model works out the most likely next word. It does not know when it does not know, so it answers just as confidently when it is guessing. It is customers' most common concern and it is justified.
There are three ways to reduce the risk. Let the model answer from your own sources with RAG and show the source in every answer. Test with questions where the right answer is known. Ask the model to say it does not know rather than guess.
The risk never goes away completely. That is why a person should review answers that lead to decisions about customers, money or people.