In A Brief History of Intelligence, author Max Bennett tell us how, as a tech guy, he has had many discussions with companies about AI applications over the past few years. The crux of the matter? These companies would often ask for solutions to tasks that, while seemingly very straightforward to us, prove immensely difficult for artificial intelligence. Or, as he so aptly puts it himself: ‘How is it possible that AI can beat a grandmaster at chess but can’t effectively load a dishwasher?’
A bit of empathy, please
For us language professionals, AI is far from a simple topic. Why? To give you an idea, let’s expand on the dishwasher analogy. When it comes to seemingly simple tasks such as text creation, transcreation and translation, expectations of what AI can do are often extremely high. Many companies ask us whether they couldn’t save costs by ‘just throwing texts into ChatGPT’. Or, worse still, they simply do it, with dreadful texts as a result (albeit ones that often sound quite natural to non-native speakers).
Now what do DeepL, ChatGPT and the like have to do with a dishwasher? Well, so long as these tools are not familiar with the discussions on how best to load the thing, have never fished the clogged filter out of murky water while holding their noses, or have never perilously stacked that one last jar on top of the already bulging drawer, their texts will continue to be missing ‘something’. This is because our language experts have something these tools simply don’t: empathy. Couple that with the necessary dose of creativity, as well as a whole host of other language skills, and you get that certain magic that turns collections of words into well-formulated copy.
So, should we remain steadfast in ignoring AI? No. After all, why not expect a certain level of assistance from a technology that can indeed beat a grandmaster at chess, analyse medical data and imaging, detect diseases in no time, recognise birds by listening to only a second of chirping and drive cars better than the average Belgian?
Machine translation over the years: from awful to good enough
Testing out new translation tools has become a regular fixture at Blue Lines over the past few years. Until recently, the conclusion was always that the result was not of sufficient quality to meet our stringent requirements – after all, quality is kind of our thing. But now, the quality of machine translation tools – which today also work using a form of (generative) AI that allows them to be constantly learning – is in some cases good enough to be used as an initial basis.
But what’s the departure point from ‘good’ to ‘good enough’? How do we decide when to deploy machine translation (MT)*? And what about our own offering? Let us explain!
Machine translation: do or don’t?
To use machine translation, or not to use machine translation? The starting point to solving this riddle is actually quite simple. Back to the above introduction for a moment: while AI is perfectly capable of solving groundbreaking mathematical formulas, it lacks empathy and common sense. And you can carry this idea through to translations: for drier material such as lengthy manuals, contracts, minutes, etc., machine translation may provide an excellent basis from which a translator can subsequently work on as a post-editor. But for creative texts, marketing material or commercial information, it doesn’t really cut the mustard. Yes, a bit of brainstorming using AI can yield some inspiration, but that’s as far as it goes.
Another factor to consider is the format of the text itself. We were recently asked to translate some ultra-short pieces of text to be used in an app. While this is pretty straightforward stuff, machine translation often misses the mark when it comes to this, too. Because a translator can decipher that, for example, the word ‘dismiss’ in that context does not mean ‘fire’, whereas a machine cannot. As a company, it would perhaps be better if your app did not encourage users to start firing colleagues…
However, we can equally say the same the other way round. For one of our clients, we translate a quarterly trend report. This is a fluently written piece, one that would generally fall into the ‘non-MT’ category. However, the report is of such a large volume that the translation would cost almost as much as the report itself. Because this report is only intended to be read internally, and because we only have a few days to translate it in its entirety, here we opt for machine translation and post-editing. The client pays less, gets the translation quicker, but is aware that the quality is lower than what they would get from us for, say, a blog post, a social media campaign or an e-book.
Using AI or MT on the sly? Honesty is the best policy
Since starting out some 20 years ago as a language service provider, we have always prided ourselves on the quality of our work. Even if we use machine translation, we want the quality we can offer to be better than from any other translation agency. Therefore, we will never – and I really mean never – send our clients a text that has not passed under the nose of at least one post-editor or translator.
We make it a point to communicate openly about what you can expect for the price you pay with us. We also demand this transparency from any companies who want to work with us. Very frequently, we get sent texts ‘for simple revision, because we’ve already translated it in-house’. If we find that MT has been used for this, we will re-open the conversation. Do they know that MT or AI has been used? What exactly is the text for? What is their available budget? Sometimes we end up doing a traditional translation, while other times we will do a machine translation that we then post-edit.
Where exactly we are going to end up with these innovations in our sector is, just like in other sectors, difficult to predict. We have long been convinced that such a revolution would result in all of us losing our jobs… For me, if there’s one takeaway from this discussion, it’s this: just as loading a dishwasher seems simple but requires a lot of complex skills and insights, crafting a (translated) text is far more difficult than it seems. So if you want to be taken seriously as a company or organisation, this is something you should bear in mind.
* For the sake of simplicity, I am using machine translation and (generative) AI interchangeably here. There are, however, some differences. For example, while DeepL is a machine translation tool that can also use AI to further develop and improve itself, ChatGPT is a text generator that uses AI but is not specifically designed to translate. If you’re interested in how generative AI performs against machine translation tools, check out this useful article. Spoiler: if you want to avoid hallucinations (i.e. errors), stay away from ChatGPT and co. for translation work.