The only areas of machine learning that I expect to live up to the hype are in areas, where somewhat noisy input and output doesn't ruin the usability, like image and audio processing and generation, or where you have to validate the output anyway, like the automated copy-paste from stackexchange. Anything that requires actual specifity and factuality straight from the output, like the language models attempting to replace search engines (or worse, professional analysis), will for the foreseeable future be tainted with hallucinations and misinformation.