When Google Translate is eerily better

One of the pleasures of being a teacher is getting surprised by my students, their responses, and their perspectives on things. For this blog entry, I got surprised with an assignment my students were working on and an eerie use of Google Translate.

This past week I had paired my students in my class so that each partner spoke a different first language. They were told to collaboratively write a story using English and to then translate it back into their first languages. My plan was to put the completed stories and translations on the bulletin board outside our classroom prior to an upcoming parent night.

One of my Chinese students, who I’ll call James, had finished writing his story with his Japanese partner and told me he was going to be using Google Translate to make his translation. I gave James my regular spiel, that this was not okay because Google Translations were never as good as human translations, and that even if they seemed good, it was probably because the student’s first language skills were not up to snuff. (I of course didn’t actually use spiel and snuff in my explanation). I told him to carefully check the Google Translation and then tell me if he could find any mistakes.

So he did. And he couldn’t find any mistakes.

Here is an excerpt of the original, unedited English version.

In the forest there is a different type of monkey. That monkey has one power, it’s power is speed. How did that monkey get its power, that monkey born in 1990 year october thirty,that time the monkey is sleep, that time the lightning hit the monkey, do monkey get power from lightning.

Of course there are quite a few errors in this passage, but since this is the beginning of the year, we’re not working on exhaustive editing. And as you will soon see, this is where Google Translate really went above and beyond.

Anyways, here’s the Google translated Chinese version in traditional characters with pinyin (romanization of Chinese characters) rendered below.

在森林裡有一種不同類型的猴子。 那隻猴子有一種力量,它的力量就是速度。 那隻猴子是如何獲得它的力量的,那隻猴子出生於1990年10月30日,當時猴子正在睡覺,那時閃電擊中了猴子,猴子從閃電中獲得能量。

Zài sēnlín lǐ yǒu yīzhǒng bùtóng lèixíng de hóuzi. Nà zhī hóuzi yǒu yīzhǒng lìliàng, tā de lìliàng jiùshì sùdù. Nà zhī hóuzi shì rúhé huòdé tā de lìliàng de, nà zhī hóu zǐ chūshēng yú 1990 nián 10 yuè 30 rì, dāngshí hóuzi zhèngzài shuìjiào, nà shí shǎndiàn jí zhōngle hóuzi, hóuzi cóng shǎndiàn zhōng huòdé néngliàng.

So I got James to read it to me, and he said, “See, Mr. Noble. It’s good.” I teased him and said that it was probably because his Chinese was not good, but then I scanned the pinyin. Even with my rudimentary Chinese, I could tell that this was not completely awful. It was certainly much different from a time eight years ago when a high school student of mine had used Google Translate to describe the point in Lord of the Flies when Jack and the boys kill the mother sow and all the piglets. Her translated version came out as, “Dozens of wrinkled puppies, helplessly killed.”

For a long time a hallmark of Google Translation was the careful capitalization, spelling, and punctuation of phrases. It’s easy to create rules for capitals at the beginnings of sentences and commas or periods at the end of phrases. But look more carefully and you’ll see many sentence fragments and other absurd word choices.

I’ve prided myself as a teacher at my ability to detect the work of Google Translate. So when James was insisting that his work was good, I thought I would retranslate his passage back into English to demonstrate that there was no way this translation was acceptable. This is what came out.

There is a different type of monkey in the forest. The monkey has a power, and its power is speed. How did the monkey gain its power, the monkey was born on October 30, 1990, when the monkey was sleeping, when lightning struck the monkey, and the monkey gained energy from the lightning.

And that’s when I received the shock. This retranslated English version is better than the first one that James and his partner originally wrote. Notice how this second version inserted correct pronouns, going form “that monkey” to “the monkey“. It fixed the date, “1990 year october thirty” to “October 30, 1990”. “that time the monkey is sleep” gets correctly placed into a subordinate clause using the past progressive tense, “when the monkey was sleeping”. Most notably, the second translation rendered the last clause perfectly, from “do monkey get power from lightning” to “and the monkey gained energy from the lightning“.

Of course the whole original piece was far from an excellent piece of writing, but for an average EAL student writing a rough draft it’s about right in terms of a high beginner low intermediate level of writing. Even if the retranslated version has too many clauses, it has suddenly become a whole lot easier to fix. I’d much rather have my students making the types of errors in the retranslated version compared to the first. In the past I found that Google Translate did far more harm than good, but now I am having second doubts about the ramifications of all this.

“So Mr. Noble. Is it okay?” It’s not too often that I’m a bit tongue tied, but I was a little stumped. The purpose of the exercise was to provide a decent translation of their work in order to validate and celebrate their first languages. I could have forced James to manually translate line by line, but as he said, this was a good enough translation for him. Eventually, I told him he could take a break and read a book while I emailed his passage to his Chinese teacher.

She got back to me later in the day and said that the Chinese Google translation wasn’t too bad, and that she would give it a 7 out of 10 for grammar for a Grade 5 boy. Now 7 out of 10 is not bad at all, especially for an average ability student. When another Chinese student saw that James was no longer working, he went across and inspected James’ work. He did find an error, but it was a minor one and certainly the type of error that James would make.

James’ Chinese teacher, intrigued a little, said she made a Google translation of the text of my return email to her and remarked that Google must have recently improved because the translated email was actually quite good. Here is part of my email to her, and the Chinese translation below.

That’s a bit spooky. Usually when it goes from Chinese to English it’s not good at all, so it was really strange when the retranslated English version was improved.


Tōngcháng dāng tā cóng zhōngwén dào yīngwén shí tā gēnběn jiù bù hǎo, suǒyǐ dāng chóngxīn fānyì de yīngwén bǎnběn dédào gǎijìn shí, zhè zhēn de hěn qíguài.

I know that Google Translate has started to make more use of AI and become rapidly more natural and accurate with its translations. But I hadn’t seen this kind of behaviour before.

I won’t be surprised if translation software will soon surpass human translation, but I don’t think it will remove the need for language teachers or the conscious interaction with language. Before there were bilingual dictionaries, language learning really was a tough slog, but the human brain still needs to do the heavy lifting when it comes to processing and retaining knowledge. I’ve often joked with my students about all the language learning options today with the net result often being, “Easy-in, easy-out.”

But after this incident with my student James, I wonder if there are applications for Google Translate using this retranslating process with student work. Your thoughts? Send me a tweet or respond in comments below.





  1. Maybe it is not so eerie when one considers that the “mistakes” makes in his original English story are actually grammatical constructs close to Chinese constructs. That would explain why it is relatively easy for GT to generate correct Chinese, and then translate back into “better” English.

    • Good point. It’s possible that the Chinglish inputted made it easier for GT to make sense of it in Chinese. He was working with a Japanese student and so theoretically it was their combined labours. But my understanding of GT then (albeit limited) is that it works on a probabilistic model, which in turn could reflect an underlying grammatical contract.

      But even with high quality English input, my students seem to find fewer grammatical errors in the GT Chinese output, than vice versa.

      The implication for Chinese English language learners might be that they can input their work into GT, retranslate it back, and then find it much easier to work with. In the students’ retranslated work above, all that’s needed is to improve the run-on sentence at the end.

      This was from two and a half years ago. We’re even more advanced now with GPT-3 and what impact that will have on translation.

      Thanks for the comment.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.