That's not saying it's perfect. It still makes mistakes and uses weird wording sometimes, but it needs way less correction.
The usual caveats of automated translation apply though. "Take it" will be translated as "tomem-no" in Portuguese, "возьми его" in Russian and "prenez-le" in French, every time assuming that the "it" is masculine (or possibly neutral, in the Russian case). Depending on context that could be confusing if the "it" is feminine in the target language's grammar.
There's also a bit of a mess with the conjugations. The Portuguese version uses "tomem", which is imperative plural (y'all take it), the french version uses "prenez" that's either imperative plural or formal/polite imperative singular, while the Russian version uses friendly/familiar imperative singular (thou, as English used to use) instead of the more polite and/or plural возьмите.
Of course those are all valid translations for "take it" in English, but in a given context only one of them will be intended by the speaker.
And those are all indo-european languages which share a common (if sometimes distant) heritage when it comes to grammar and social customs. If I spoke Japanese (which uses a complex system of honorifics) I'm sure I could come up with much more awkward translations.
In Deepl's defense however it manages to piece it together if I give it more context, for instance "The box is on my desk. Please take it." is translated as "Коробка на моем столе. Пожалуйста, возьмите ее." where it uses the correct её for the (feminine) box, and the adding "please" made it switch to the polite form of брать.
But of course in casual conversation through some online messaging application piecing together the context will be a lot harder in many cases. You have to deal with slang, incomplete sentence fragments, typos etc...
I found this article about the interaction of Chinese and Mongolian workers at a fluorspar mine particularly illuminating in that regard: https://madeinchinajournal.com/2021/01/25/the-double-tongued...
Excerpt:
Xu Jun, the crushing section lead, recalled: "Once I asked my Mongolian assistant to make a slanted connector for the ore crusher. I showed him many times the angle of slant with my gestures, but he just did not get it. Do you know what he gave me in return? A straight angle and a blank look. In the end I had to make one myself to show him what I meant. He instantly understood and raised his thumb, saying ‘Sain, sain’ [‘good, good’]. I didn’t know if I should laugh or cry."
He blamed the incident on the Mongolian worker’s ‘inflexible brain’ rather than acknowledging a communication failure for which both were responsible. Such refusals to share responsibility were not uncommon among monolingual workers. Linguistic miscommunication without timely mediation easily perpetuates essentialist stereotypes. After the incident, Xu Jun would always ask a bilingual person to bridge the linguistic gap. Even though the translation may be imperfect as the intermediary might not know particular technical terms in Mongolian, their colloquial translation of the key actions involved could help the listeners understand.
Lol, it's like he works for the 1700s East India Company
Using Google's translation API, the number of mistakes is reasonably low, and it's absolutely possible to have a comfortable meeting even when technical topics are discussed, as long as participants are aware of the possibility of mistranslation and take care to avoid resulting miscommunication. It's far from perfect, but it's good enough with those caveats. We have plans to test out several others including DeepL.
The output just barely manages to get the broadest meaning across. If even that.
If it is also supposed to be real time and able to handle a conversation at a realistic pace... just forget it.
This is just too be able to list it as a feature, it's not actually going to be good or work.
When the participants in a back-and-forth e-mail exchange, for example, do not share a common language and one or both use translation software to communicate, are they able accomplish what they need to do reasonably well? Or does the communication break down at some point?
Can live meetings be conducted successfully when one or more active participant uses MT interpretation? Or are there too many cases of faulty voice recognition and mistranslation for the meeting to proceed comfortably?
I ask because I have been thinking and writing for several years about the implications of improved machine translation for foreign-language education. I know how accurate, and inaccurate, MT can be at the sentence-by-sentence level, especially between Japanese and English, and researchers I know are starting to study how some language learners use MT. But it’s been hard to get a handle on the extent to which the software is actually used in real-life contexts.