Update in 2026: In December 2025, Google rebuilt the text translation in Google Translate on its Gemini models. Google says the new version handles idioms and slang much better than the old word-for-word output, starting with English and about 20 languages that include Japanese and Chinese. AI assistants have also become a common source of auto-translated manuscripts. I haven’t re-run the tests below, which date from 2018, so read the grades as a snapshot of that period. The advice at the end still stands: nobody should publish machine output without review, and an editor can’t check terminology without the source text.
Update in 2021: Since I wrote this in late-2018, it has only gotten worse. More than ever, I see the use of Google Translate. I know this because I get the original text and drop it in Google Translate to be sure. Sure enough, it’s exactly what’s on the paper I was sent for editing.
It’s more cunning than ever because the tool seems especially well-trained for Chinese. Not sure why, because I don’t speak Chinese. Perhaps it’s the amount of data being input. Perhaps it’s the more-consistent patterns in the language. Also, compared with Japanese, there is only one reading of Chinese characters, which leaves much less guesswork for the machine learning. So, as a consequence, more than ever, Google is spitting out text in which a non-specialist would not spot terminology errors. The seemingly good news for me is that specialist editors should be in higher demand than ever. But with Fiverr pseudo-self-appointed editors and increasingly prevalent open access with lower standards and rapid publication, only the smartest and most ethical clients use a specialist.
If it’s epidemiology, economics, or some sort of population study, I can usually spot the problems quite easily. These are the sorts of things I detail in this article. I may trial my theory with some experts in areas like chemistry and physics, where I’m quite sure my skills are limited.
—
The telltale signs are there: the same word written three different ways, an acronym half the time and spelled out in full the other half, and comically literal terms.
This paper’s got Google Translate written all over it.
But I confess…
Google Translate’s gotten much better
Occasionally its translations need little or no correction. This can happen with stock phrases, simple grammar, and in highly formulaic writing such as common scientific methods.
Where it continues to fail is just about everywhere else, especially for (1) the conversational tone that modern-day marketing copy needs, and (2) the precision and clarity that scholarly writing needs.
If you only use Google Translate to point you in the right direction, to give you the gist, it can do the trick probably three-quarters of the time.
So, good job, Google. But you are a long way from replacing human translation, and a much longer way from replacing source language written from scratch.
Meanwhile, every other translation tool I know is poor for giving the general idea of most text, and is unspeakable for creating publishable text.
I should also mention that the Japanese–English pair is my basis for this. There may be better progress in other pairs with more similarly structured languages.
Takeaways
- Google Translation is good for getting an idea of what’s going on and as a start when doing direct, literal translations.
- All other translation tools are good for little more than looking up words and short phrases.
- Output from no translation tool should ever be used for published material.
