Most sources for machine translation from Arabic-English pairs are data from US Army / anti-terrorist data.
So the data are learned to translate terrorist messages very well, as well as army terms. However, it doesn't translate day-to-day use that well.
Machine learning is only as good as the source data - and Arabic-English language pair specifically has data from US military complex, that loves translating terrorist messages. So translating Al Qaeda messaging will be good, translating regular message boards will be worse. The machine learns to translate everything as an attack, since that is what it knows.
(I cannot give you exact sources, but I knew people from inside Google. Facebook uses Bing, but I think the data sources will be similar.)
So the data are learned to translate terrorist messages very well, as well as army terms. However, it doesn't translate day-to-day use that well.
Machine learning is only as good as the source data - and Arabic-English language pair specifically has data from US military complex, that loves translating terrorist messages. So translating Al Qaeda messaging will be good, translating regular message boards will be worse. The machine learns to translate everything as an attack, since that is what it knows.
(I cannot give you exact sources, but I knew people from inside Google. Facebook uses Bing, but I think the data sources will be similar.)