This is going to sound maybe darkly humorous, but one of the reasons we haven't been able to make much use of AI is because the natural language content is itse

This is going to sound maybe darkly humorous, but one of the reasons we haven't been able to make much use of AI is because the natural language content is itself nearly-universally incorrect, incoherent, or just plain gibberish. It's true even in published documentation - they are full of garbage. The garbage that gets tossed back in the AI team's face is a very similar flavor to when I first started doing doc analytics but with older techniques (topic modelling, embedding, ontology, clustering, some xquery/jquery/mongo crap, good ol' regex). We saw things like single digit percentile commonality between configurations that were, on-paper (and in the logistics system!) identical, or landing gear classified as a fuel system component, or whole assemblies without fasteners. What happened from that great revelation from back in the day? Absolutely nothing. There was no funding to fix any of it, because dollars and charge codes came from new projects. The directive remained: just shut your mouth and keep churning XML for the latest gee-whiz kludge sold to some sorry USMC procurement department head. No one reads the damn thing, no one reviews it, and it'll take an officer dying from your product before anyone notices. The absolutely terrifying take away I have from this is that some bright spark upstairs is going to realize they can generate garbage all day long so much more cheaply with generative AI, rather than taking a biiiiiiig step back and consider whether publishing garbage was a good idea in the first place. I'm going to put money on "Nope!", which I guess is the place to make money right now.