The translation service of Google, which works astonishing well in many cases, also serves as a nice example for a flaws in machine learning:

Neural Networks identify patterns in the data and might come up with “answers”, even if the input data is chaos. The results depend a lot on the data that was used for training.

An article at motherboard answers the question, why some of those “translations” often have a touch of religious wordings: In language with limited translated corpora, religious texts like the bible often serve a good part of the training data.

René from Nerdcore, where I stumbled upon this topic, has some more examples. But I wanted to try this for my own: