Agree with you on that, it was my concern too, but the way I think about it is access to information, the goal is not to provide hallucinations with a straight face (aka GPT), but rather use it as a way to extract necessary information fast. For instance, I have a built-in RAG that reads of growing collection on books on medical, survival, etc. (https://github.com/dmitry-grechko/waycore-knowledge) that AI agent is using to answer questions. Moreover, it has a built-in safety loop to always inform users on the accuracy of the information, but also if the information request has an impact on health & safety, it will warn users about it too.

So, I certainly see the inherited risk and problems, but mostly think about it as a means of information extraction