Lost in Translation: The Importance of Human Expertise in a World of AI-Driven Translations

If we can choose one business or technology trend that defined 2023 so far, it’s definitely ChatGPT. With artificial intelligence picking up more and more attention, business leaders are putting out more and more content on how to transform a business using AI. One of the recommendations made by some of the biggest names in the marketing world, including Gary Vee, was to use ChatGPT to translate websites and marketing collateral into other languages.

I am a skeptic. While I've personally used machine translation softwares, such as Google Translate, to translate simple sentences between French and English, I have not been satisfied with their output for Arabic overall or for complex paragraphs in European languages. In fact, one of the key pieces of advice I give clients when asked about ways to identify if a translator they've worked with is good is asking them to run the translation through Google Translate back into the original language. If the output is not a grammatical match to the original text, then the translator generally should be doing a good job.

You’ll come across countless stories on social media about terrible translations that made it onto advertisements, signboards, and even tattoos. A lot of the wilder instances are deliberate manipulation of a machine translation that relies on crowd-sourced input, such as Google Translate, which allows users to suggest translations. If enough users suggest an incorrect translation, it becomes the translation that shows up for other users.

Some of the less wild instances are simply because machine translations generally translate word-for-word and ignore the surrounding context. AI translations are only slightly better because they're able to take the context into consideration. However, they only consider simple sentence/paragraph-related context. Additionally, due to limits on token counts, and higher creativity, AI can sometimes translate the same word in an article in two completely different ways.

Another downside of artificial intelligence translations is how differently they interpret the same abbreviation. You could ask ChatGPT to translate this sentence: ‘I watched a movie on MBC 2 today.’ It may be translated once as the TV channel we all know here in the Middle East, and in the second instance, it could be translated as Mary Baldwin College.

Let's say AI or machine translation improves a bit more so it avoids the pitfalls outlined above. This could take a few weeks to a few years, depending on how rapidly the machines learn how to think in different languages, simultaneously, to best translate any given piece of content. Does this solve all the issues with non-human translations? Does it eliminate the need for people like me and my team?

The answer, unfortunately, for those looking to cut down on costs, and fortunately for us translation specialists, is no. Because, simply put, translation isn't just about taking a few words and making sure they sound right in the other language. Translation has to take into account cultural differences, an understanding of the target audience, and a proper interpretation of the idea that the words are trying to convey.

An example straight from the early days of my career, when I was working as a freelance translator, is when I was working on helping a well-known language services provider with some patient surveys. These surveys were to be used in cognitive debriefings as part of clinical trials being run worldwide. The original text supplied by the pharmaceutical company was in English, and the requirement was to translate them into Arabic for use in Saudi Arabia.

One of the questions that were on the list tackled improvements in fine motor skills, and to make it simple for patients of varying educational backgrounds and ages, the question gave opening a milk carton as an example. I shot an email to the project manager asking for a bit of clarification about the type of milk cartons intended in the survey, and they quickly confirmed that it was the type that is used generally in the US, where you have to hold the top part and pull the carton open.

I requested permission to modify the example, as this type of milk carton was not common in Saudi Arabia at the time, and I changed it to peeling an orange because it involves the usage of both hands. One to hold and keep the orange steady, and the other to either peel by hand or with a knife. The same motor skills are used in opening milk cartons.

I was able to do this because 

a) I understood the idea behind the words ‘opening a milk carton’

b) I knew the local market in Saudi Arabia well 

c) I understood we needed a simple and common enough task that the target audience would understand.

Once, I was in a supermarket in a well-known shopping center in Dubai. There was an aisle that said in English ‘oatmeal’. The Arabic translation for that literally meant ‘meals made of oats’. I'll bet that was definitely an AI or machine translation.

When it comes to ensuring that a translation is accurate in terms of conveying the right meaning, artificial intelligence will not replace human intelligence anytime soon. We may use it to assist in translations and to cut down on time, but ultimately, there has to be a human revising and even rewriting the output given by machines. For me, one human isn't enough. Every translation we work on has to have a minimum of two pairs of eyes, to avoid the inevitable ‘human errors’.