I have just seen a good reason why AI can give erroneous answers, on the basis that it will not understand context.
I follow Mark Gibbings-Jones on
twitter, his blog
Broken TV researches "The 100 Most-Broadcast BBC Programmes Of All Time" - It appears he has used info from the BBC "genome" project - that has details of nearly
every programme listing in the Radio Times. However, because Optical Character Recognition (OCR) was used to "read" the listings and lack of resources mean that the OCR has not been error checked, there are discrepancies, like the OCR mis-reading characters, so a letter "t" might be seen as an "i" or "l" by the OCR, meaning the programme title would be missed in the count of appearances. That in itself creates potential for AI to mis-represent facts.
In this instance, this had the effect of meaning a particular programme "appeared" to have been listed a different number of occasions. So, in the detailed listing in the blog article there is the "original" list of entries from 16th most watched to 11th, and an amended list of 15th to 10th place based on the corrected numbers. So, in the same article 6 shows are listed with correct and incorrect positions. Let's see an AI sort that out!