It would be helpful if those using MT (Machine or Online translation) have some grasp of how it produces the resulting translated texts. Most people expect the MT to produce translations based on knowledge of linguistics and a memory dictionary. This used to be the case, but MT has now become much more like a search engine. When a text is entered into the API (Application Programming Interface) its output translation is mostly based on the statistical chances that a succession of words will follow the input succession of words. In other words, a string is supplied for the translation with the most likely string of translated words supplied in the translation response. This is a rather simplistic description of the process of MT to date.
Linguistic Depth
MT systems may be very sophisticated but still lack the linguistic depth needed for certain translation applications and jobs. Colloquially, the machine translations software and dictionary may not have street smarts. The software system and dictionary may have knowledge of technical and legal phrases and texts of translations, but not some common every-day use translations. Another example would be incorrectly entered texts or texts with errors. The API would not have many (if any) stored responses to return a correctly translated response. This highlights the importance of using the machine translation corpus (body of translated texts) as a theoretical or research tool and not so much a 100% accurate translation tool.
MT Awareness and Usage
The greater public seems to have some errors in perception about machine translation and education must be the key to changing that. A brief study of the use of an English-Espanol MT system showed possible confusion of the people between it and a language dictionary. MT users are prone to (large) errors in translations when submitting translations for homework or simple chat sessions. Dictionaries are generally best for word-in-context definitions. MT is best for translators and translations of whole sentences. MT users would benefit for better comprehension of their translation inputs and outputs, naturally. MT programmers can help by developing better MT to meet those needs. And, language courses might help by including material on the last two points.