Having trust issues now with so many AI in market. All seems confident and act like they know what I am searching in document , better than myself.

I’ve been experimenting with AI-based document/contract analysis recently, and one thing keeps bothering me: Most AI contract reviewers feel like black boxes. I barely know how exactly they perform analysis and the weird things is every llm has something new pointed out apart from common issue. Now I don't really understand that if other missed this finding as mistake or it was desgined to ignore few things. Honestly , I am so confused sometime how to use these gpts. I mean sometime , I use a neutral Claude for judge and then give all other llm answer and then pick up the facts that are common blindly and then I read the unique findings and then make a decision to include it or not in my final findings report. For small documents it feel like , it over doing as now I am reading summary of like 3 llms , which is more than reading time of small document. For larger one , although it helps and usually accelerates my speed of reviewing or analyzing. I wonder how you guys are using it ?? Thought process are useful but it even more bigger to read. Do you guys feel same tension using llm . This is my theory but I think honestly the llm are being used in wrong way. Its really good extractor tbh , but its not a best responser . What I mean is to generate response it just predict what words should come after and thus with all that pattern matching , it eventually make sentences which are contextually aligned and make sense, but there is no way I can honestly say, yeah it has thought all the angles. Especially for high stakes contracts . I think we should restrict use of llm blindly , and more use deterministic rules as brain rather than llm as brain , which I am seeing most document analyzer are doing. Its still really useful with its sheer speed of processing document. If a dumb job needs to be done then llm is good where stakes are low. Are you guys also feeling this pain ? I built this wrapper doing exact things, where I believe analysis done for any document is backed by * exact evidence, * why the conclusion was reached, * what source was used, * what the AI could NOT verify. And all these with confidence level. Not some wrapper where llm produces ai slob with no meaningful information and sounds like OG who knows what I want more than me. Sometime they act stupid. Anyway what you guys take on this ?? Do you think necessity of this ?