LLMs are perfectly fine, and cool tech. Problem is they’re billed as being actual intelligence or things that can replace humans. Sure they mimic humans well enough, but it would take a lot more than just absorbing content to be good enough at it to replace a human, rather than just aiding them. Either the content needs to be manually processed to add social context, or new tech needs to be made that includes models for how to interpret content in every culture represented by every piece of content, including dead cultures who’s work is available to the model. Otherwise, “hallucinations” (e.g. misinterpretation and thus miscategorization of data) will make them totally unreliable without human filtering.

That being said, there many more targeted uses of the tech that are quite good, but always with the need for a human to verify.