How good are AI translations really?

The world of business and the translation industry are embracing AI translations as a promising development for the future. But how does AI translation measure up against human translation, and is it the right solution for every context?

AI translations are increasingly being used by translation agencies and clients alike as a rapid, more cost-effective alternative to human translation. Google launched its first statistical machine translation in 2006, but it was not until ten years later that Google switched to neural machine translation technology (NMT), which analysed the meaning of entire sentences instead of translating one word at a time. NMT has been refined extensively in the years since then and this has triggered the boom in demand for AI translation.

The major players offering AI translation

There are currently at least sixteen providers of AI translation software on the market, but the major ones are DeepL, ChatGPT and Google Translate. DeepL is based on the Linguee online dictionary and is built on a convolutional neural network, which progressively refines its own output using filters to remove inaccurate results. Below is a screenshot of DeepL’s translation of a paragraph from an article in the Süddeutsche Zeitung about this year’s Nobel Prize for Chemistry:

DeepL has correctly identified that the excerpt was written in German and provided a fair translation of it into English. The same text run through Google Translate looks like this:

Unlike DeepL, Google Translate does not allow users to choose between US and UK English, so the ‘Was’ after the colon is capitalised, which may seem strange to UK English readers.

ChatGPT (the ‘GPT’ stands for ‘Generative Pretrained Transformer’) was initially developed by OpenAI as a chatbot and subsequently given a translation function.

Interestingly, ChatGPT is the only one of the three AI engines to give the correct idiomatic translation of ‘doch eine Falschmeldung’ (‘just a false report’).

AI translation is not confined to the written word. HeyGen, founded by Joshua Xu and Wayne Liang in 2020, is an AI video generator. The AI records the user’s voice and their lip movements and uses them to create videos where the user appears to be speaking themselves but their words are taken from a text file. The technology can even film a person speaking in one language, translate their words into another language and adapt their mouth movements to match the sounds of the translated text. It is obvious how this tool can be used to spread misinformation: anyone could scan in a video of a politician or celebrity and use it to create a clip of that person making whichever statements the person wanted. At the moment, the technology is still ahead of the regulatory curve, so it remains to be seen which action will be taken to govern the use of this AI feature and how people’s intellectual property in their own image will be protected.

Machine translation is trending

Somewhere between AI translations and human translations is the so-called MTPE process (Machine Translation Post-Editing). This involves having a computer translate a text, then a human reviewer checks the translation for accuracy, completeness and flow. It is similar to the traditional process of translation and proofreading, except the translation step is done by machine. At SwissGlobal, we have witnessed a sharp upturn in demand for MTPE.

MTPE is most well-suited to projects intended for information only (internal company documents, anything not destined for publication, and so on) and for when the deadline and/or the budget are tight: the result is not as impeccable as a purely human translation would be, but it might be deemed an acceptable alternative if the resources needed for a human translation are lacking. AI translation providers are actually starting to offer this added human element as an option: ModernMT proposes standard AI translation as its basic package, and an engine that learns from human corrections in real time as its premium ‘Human-In-The-Loop’ product.

Another distinction between machine translation and AI translation is that the machine translation engine does not learn over time, so the results after five years of use will be of the same quality as the results produced on the first day.

Knowing the limitations

AI translation is increasingly being hailed as a miracle cure but it still has several limitations. It is still not as accurate or as idiomatic as human translation and can fail to notice subtleties like wordplay, not to mention the fact that it will never identify the need to localise content and adapt cultural references to suit the target audience. There is also the issue of data security: Most AI translation engines populate themselves with text found online, and in turn release the results of their own translations online (although DeepL promises not to release your translations online if you pay for DeepL Pro). If the text you need translated is in any way confidential, you cannot afford to have it ending up on the public internet for the world to see. SwissGlobal’s keen interest in AI translation and our many years’ experience in working with data security solutions puts us in a unique position in the translation industry of being able to address both issues simultaneously. We have the human experts on hand to check the quality of your translations, and all of your data is stored and processed in data centres within Switzerland.

Machines, humans and collaboration

The quality and performance of AI translation are improving at an impressive rate, but the technology still cannot replace human translators. What it can do is save time for human translators and money for translation clients: AI translation provided and verified by a secure, professional provider like SwissGlobal can offer the best of both worlds.