Given that the AI that kicked off the AI boom in 2022 is essentially a chatbot, people suddenly thought that GenAI can do anything that’s related to text. But can it really? What about translation, for example?
DISCLAIMER: This article was written in January 2025. It might be that some of the capabilities of AI as outlined in this article have changed if you’re reading it later than that.
When looking at the question whether AI can translate or not, it’s important to keep in mind that there’s nothing intelligent about AI. What we call Artificial Intelligence is really just a Large Language Model (LLM) with good marketing – it’s not based on some kind of understanding, but just calculates what’s most likely to come next.
I won’t go into the details about this here as I wrote another article looking at what AI really is in preparation for this one, but I will say this: AI is not intelligent. Think of it as a super fancy autocomplete rather than a smart entity. It doesn’t understand what you’re saying, it just calculates the probabilities and tells you what’s most likely to come next.
If you don’t believe me, ask AI how often a specific letter occurs in a word. Spoiler alert: they don’t know. While they do seem to get it right every now and then (lucky guesses? All engines I asked seemed to agree that almost all words contain any letter twice … 🤔), most answers were wrong:
- No, Claude, the word tailoring doesn’t contain 2 rs (even if you tell me that the first r is in the middle of the word and the second one at the end).
- And no, Gemini, there are not just two bs in the word bubble.
- There was an example here telling ChatGPT, that the word strawberry doesn’t have 2 es in the first draft of this article, but it looks like ChatGPT learned to count letters in the last few weeks. Good for you. 😅
So, keeping in mind that AI is based solely on statistics and probabilities, let’s have a look at the question this article poses – can AI translate? The quick answer is, yes. In theory, GenAI can translate and similarly to certain pre-AI machine translation (MT) tools it’s actually not half bad.
But don’t confuse that to mean it’s good.
AI in translation
Translation is a very complex process where you need to have a very good grasp of the source and the target language, add to it your knowledge of the world in general and the specific cultures of both languages involved.
You have to bundle all that and combine it with your copywriting skills to create an idiomatic copy that nails what the text type in question has to achieve and tailor it to your specific target audience (which is usually a lot narrower than “speakers of language C”) – all while maintaining the messaging of the source material.
So, in short, you need:
- a lot of know-how about two languages, two cultures, common text types as well as specific target audiences
- tons of creativity to work around untranslatable words, character limits and other challenges
- great skills when it comes to researching certain topics and writing about them (which is a lot harder now that Google slapped AI onto their search engine. Thanks for nothing.)
If we combine that with what we just said about AI being less like an intelligent entity and more like a very potent autocomplete, it becomes obvious that AI can’t translate. Not really, anyway.
That doesn’t mean that it can’t be used for certain translation tasks though – if you just need to get the gist of something, for example, it’s perfect. I also heard some legal translators say that apparently one particular model is quite good at translating contracts, so a lot of legal firms use that now, which makes sense because a lawyer has to check the translated legal documents anyway, no matter who or what translates it.
And this brings me to the next point: if you want to use AI for translation, you have to keep in mind that the savings aren’t actually that big, because if you want to use the translated copy for anything more than internal FYI purposes, you have to have a professional proofread it!
AI in translation still needs human support
This specialised service is called PEMT/MTPE (post-editing machine translation/machine translation post-editing) and it’s a step you can’t skip if the material is to be used in any official, client-facing capacity.
Against common belief, PEMT is actually quite time-consuming and takes a lot of effort, because when you’re reading tons of “good enough” text, the lines start to blur and you stop seeing the real issues.
A lot of professional linguists (such as myself) happily offer this service, but exclusively at an hourly rate – simply because reading 1k words takes the time it takes. That time grows when the quality of the material is bad – and with PEMT, you have to take it even slower if the quality has to be good in the end, not just good enough.
We professionals are currently fighting what seems like a losing battle against big translation agencies who are convincing clients to use AI-powered MT because it’s so cheap, saying that they have a professional proofread it, and then paying those “valued professionals” 1-2 cents per proofread word (which is € 10 for 1k words, which takes at least 1 hour to complete – I’ll let you do the maths). I’m guaranteeing that no actual professional who’s worth their salt will look at your text at such prices.
So, let’s look at some of the problems that arise with using GenAI-powered machine translation, which I hope will make clear where its limits are.
Problems when using AI for translation
One thing AI finds particularly difficult are nuance and mood. Recreating a certain feeling in a translation is quite complex and the result of careful word choice, sentence structure and rhythm. All that is much less tangible than the simple replacement of one word with its literal translation in another language.
These issues are especially noticeable when it comes to marketing content that’s supposed to evoke emotion and create a connection.
To illustrate this, let’s have a look at the translation Claude provided for a section of my copywriting page:
| English source text | Claude’s German translation | Back translation of DE version |
| I want to help your company succeed through truly excellent marketing content with the right balance of wit, emotions and word confetti. | Ich möchte Ihrem Unternehmen zum Erfolg verhelfen durch (1) wirklich hervorragende Marketinginhalte mit der richtigen Balance (2) aus Witz, Emotionen (2) und Wortglitzer (3). | I want to help your company succeed through (1) really excellent marketing content with the right balance (2) of wit, emotions (2) and word glitter (3). |
| When your customers read about your latest ice cream flavours in your newsletter, their mouths should water. | Wenn Ihre Kunden in Ihrem Newsletter von Ihren neuesten Eissorten lesen, sollten ihnen die Mundwinkel vor Verlangen nach unten laufen (4). | When your customers open your newsletter and read about your new ice cream flavours, the corners of their mouth should melt downwards with desire (4). |
- Not idiomatic: The first part of this translation sounds very awkward in German because it uses the English sentence structure.
- False Friends/Mistranslations: In the second part of the sentence, “balance” and “emotion” have been translated literally, when “combination” and the synonym “Gefühl” (feeling/emotion) would’ve been the better, more idiomatic choices in this context.
- Mistranslation: Conversely, the last word of the first sentence, “word confetti”, doesn’t exist, so a literal translation would’ve been better to preserve the original intent than the new creation “word glitter” – especially because glitter has a certain girlish connotation that confetti doesn’t have.
- Mistranslation + wrong grammar: In the fourth example, Claude completely messed up by saying that the corners of people’s mouth should melt downwards with desire. Three problems with that: This is not a collocation in German. This evokes the creepy image of a face melting. And the word “Verlangen” (desire) has two connotations in German, neither of which works in this context – it’s usually either sexual or used in archaic language like in fantasy settings.
And if you’re wondering, while Gemini and ChatGPT did better with issue #1, they had the same problems with #2 and #3. For issue #4, they both used the correct collocation of a mouth watering, but they both introduced a grammatical error. I found it interesting though, that neither of the three models provided a literal translation for “word confetti” (ChatGPT turned it to “word acrobatics” and Gemini to “word fireworks”).
And if you’re now curious about how a skilled transcreator localised this, here’s my version. 😉
| English source text | Claude’s German translation | Back translation of DE version |
| I want to help your company succeed through truly excellent marketing content with the right balance of wit, emotions and word confetti. | Ich möchte, dass dein Unternehmen auffällt – und zwar durch richtig gute Marketing-Texte mit Witz, Gefühl und Sprachkonfetti. | I want your company to stand out – through really great marketing content with wit, emotion and language confetti. |
| When your customers read about your latest ice cream flavours in your newsletter, their mouths should water. | Damit deinen Kund:innen der Geschmack des Sommers plötzlich wieder auf der Zunge tanzt, wenn sie den Newsletter zu deinen neuen Eissorten lesen. | (So that) the taste of summer dances on your customers’ tongues when they read about your new ice cream flavours in your newsletter. |
That the three GenAI tools yielded so similar translations brings me to another point I personally am really struggling with: Machine Translation, even when it’s supported with GenAI, usually all sounds the same – the tone I used in my English source text hasn’t been recreated very well and the word choices are practically identical in all three models. All the personality I poured into my website copy has been removed.
And yes, I know, you can tell AI to use a certain style, form of address, etc. – but most people don’t, so everything ends up sounding the same. Linguistic variety is a beautiful thing, and GenAI is killing it.
Even if you do PEMT properly and have a professional proofread your AI translation, this regression to the mean is preserved to a certain extent. This is because when you read a translation for something (no matter if it comes from a human or a machine), your brain is already steered into a certain direction. It’s still possible to come up with something completely new when that happens, but it’s very, very hard.
So, to sum all of this up, I’m not opposed to AI-powered MT, but there’s a time and a place for it. Carefully evaluate which of your material is fine with an okay-ish translation and what copy needs to be spot-on.
And if you’re in doubt, contact a professional, they’d be happy to help you with that choice. 😉
