Making the most of Generative AI

Translation

• Generative AI can be a very flexible tool for translation work, allowing you to access sources you would otherwise not be able to use
• Generative AI can generate errors and misinformation: this is especially problematic in translation work, as you may have no way of spotting any such mistakes when dealing with languages which you do not understand.
• Issues with fully representing creative voices and cultural background mean that AIs may not always produce the best translations of creative works.
• As a result, generative AIs are often best used as an aid to your own translation work, rather than being relied on to do all the translation themselves.

Benefits for translation

Increased access to sources

If you can type or copy-and-paste it, generative AI can usually translate it for you. This opens up a huge range of information and viewpoints you might not normally be able to access and use in your work, including sources which would not normally have a published translation.

Flexibility

Generative AI can be a very versatile translation tool: via a single AI, you can work with a large number of languages, either using it like a dictionary to help with individual words and phrases, or getting it to translate whole passages of text in the style you desire.

Checking your own translation

Generative AI can be useful for supporting your own translation work. Even if you are reasonably familiar with a language, there may be vocabulary, usages, grammar or syntax you are less confident in translating. If you are not quite certain what a passage says, or even what an individual word might mean, an AI can bridge that gap in your knowledge and confirm your translation.

AIs are most effective in this sort of context where you understand the language a little: the AI aids you in those areas of the language you still struggle with, and you should have sufficient ability to at least spot if the AI has made a mistake, even if you don't know the correct answer yourself.

Limitations for translation

Limited training data

Generative AI is only as good as the data it is are trained on and this can affect translations as much as any other form of work. If an AI doesn't have many examples of a given language to work from, it may be not be able to produce a good translation. Even if it has many examples of a language to base a translation on, an AI may struggle with rare or specific uses of a word because that word or usage doesn't appear often in its training data: examples might be specialist technical terminology, slang, dialect words, or historical usages.

Problems with creative works

While AIs can be given instructions to mimic a particular style, they may fail to fully convey the voice or flavour of creative works. Expert translation of creative works will sometimes depart from an exact translation to better reflect the underlying mood and intention of a piece for another language or culture. AIs have no understanding as such of what they are translating or writing, and so tend to be more literal, often at the cost of voice, feeling or meaning.

For example, here is Google Bard's response to the following prompt:

The human translation reads much more naturally, because it maintains the sense and intent of the words precisely by not being absolutely literal (and because humans can better judge when the epic style may or may not be appropriate).

Lack of cultural background

There is often more to understanding a work than simply possessing a translation of the words: there might well be unique cultural issues and expressions you need to comprehend to properly understand the source, and the AI will not automatically provide these. You can use the AI to search for more information on these topics: but you can only do that if you are aware of an issue in the first place. Professionally translated and edited editions of works, when they are available, will often provide notes to help you understand the cultural background.

Difficulties in spotting errors

A problem with using sources in a language you do not understand is that it is hard to confirm if a translation is accurate: if you don't know roughly what a source says in the first place, how can you be really sure the AI's translation is correct, or that it's fully reflected every nuance of the source? How can you tell the AI what style to use for the translation when you have no idea what the source is about?

All of these warnings are doubled if you are using AIs to produce content in a language you don't understand: if you can't detect possible errors in the translation, you can end up making some very embarrassing mistakes!

Translations that have been professionally edited and published may provide a more reliable alternative when they are available, as they will at least have undergone expert scrutiny. Getting assistance from someone who does understand the language can also provide a helpful safety net. ;And if you are regularly using sources in another language, it may be worth learning that language in the long run!

Importing texts into the AI

For the AI to translate a text, you need to import it into the AI by some means. Usually this will be very straightforward: but occasionally, particularly for longer texts, there may be difficulties copying or importing the text into the AI.

References

Beowulf. (2000). (S. Heaney, Trans.). Faber & Faber. (Original work published circa 1000).

Google LLC. (2023).Google Bard [Large Language Model].