An Apple study has found that artificial intelligence models get confused by irrelevant information in math problems.