The company I work for is currently developing an AI based system to “read” legal documents and boil them down to a few pieces of info to shove in a spreadsheet. One of the people I work with has to make a system that tests the quality of that AI. So far he’s found that to be an impossible task because the info he’d check it against, that in which humans did the same task, is also super fallible and basically he’s found that they both suck in different ways.

Which means the AI is just being used because no one can say its wrong, since no one can say its definitively right either… but it’s cheaper.