I would argue (speaking as a linguist) that the evidence is that these apps are good for sentence-level translation but still reliably produce incoherence at the passage level. The more complexity you introduce, the higher and higher the likelihood that you encounter a construction that requires semantic knowledge…