TL;DR: IAMAE and can't find a decent source because the internet is shit, but a good LLM won't modify your source data source and won't hallucinate an answer for you, it will give you the prompt. The bubble technology and how most people are using LLMs, and how capital wants you to use LLMs, is backwards. It's like going to a methed up doctor with a diagnosis and asking the doctor to prove your diagnosis to you, instead of the doctor soberly observing your health and telling you what if anything is wrong based on what they actually observe.

Talk to anyone who worked with LLMs before ChatGPT was a household name and didn't work for OpenAI or Anthropic. like before 2019. This person will likely sigh deeply in anguish when you bring the subject up, because their whole career and practice has been bastardized. See https://www.smithsonianmag.com/smart-news/researchers-translate-bat-talk-and-they-argue-lot-180961564/ from 2016.

The first use case for LLMs was basically analyzing large sets of structured (like a table with headers and consistent values) data. Like audio recordings of bats making noises. Or cancer biopsy results. With the right training and data, a LLM can identify patterns that a human likely couldn't based on looking at characteristics of the data that just wouldn't be apparent or the data set too large for a human to comprehend.

The second use case for LLMs was analyzing unstructured data, like all the Discord messages with freeform text. Or pictures of fruit on a conveyor belt. A trained LLM could tell you if people were talking about agitprop or movie reviews, or if a piece of fruit is above or below a quality threshold because of it's size, coloring, shape, etc.