Generative artificial intelligence (AI) tools are more widespread than ever, and one of the functions they offer is fast, real-time translation. Some people are even questioning the need to ever learn another language if near-instant translation is so widely available to all of us, often for free.
However, those translations are far from perfect. While AI translations are very fluent, fluency doesn’t equal accuracy. Even proficient bilinguals can overlook errors hidden under the surface.
Some of the mistakes in AI translation come from a lack of learning data, insufficient prompts, or even hallucations – confident presentation of false information, a hallmark of AI chatbots.
Other errors reflect something more fundamental: the inherent complexity and flexibility of human language, and the linguistic or cultural differences between two languages.
In my professional work reviewing real AI translations between English and Chinese, I have found some revealing examples of this problem. English and Chinese are two of the most used languages, with a large amount of data for AI to learn from; the issues could be more serious if you’re working on less common languages.
For those using generative AI to translate, it’s important to keep these risks in mind.
Where can AI translation go wrong?
At word level
Words can have multiple meanings and AI may not always choose the right one.
In one translation I reviewed, an international student described Australian professors as laid-back because they wore thongs to class. AI interpreted “thongs” as G-strings (underwear) rather than flip-flops (footwear).
Tense can be problematic too. English verbs clearly indicate tense through forms such as “go”, “went” and “is going”. Chinese verbs don’t change in the same way. Instead, time is often conveyed through context or other words. When translating Chinese into English, AI can choose the wrong tense, changing whether something has happened, is happening or will happen.
At sentence level
A sentence can have different interpretations that are grammatically correct but different in meaning.
For example, “You can see him in the store” can mean he is in the store – or you’re in the store when you see him. A human translator can often resolve such ambiguity by considering the context, but AI makes a decision by predicting the next possible words.
And if a mistake is made, it can be difficult to spot later when the overall AI output sounds completely natural. Without careful checking against the original, the meaning can be changed without you noticing.
At whole text level
Accuracy involves more than getting individual words and sentences right. Information must also flow across a text, which is known as cohesion. And languages often achieve cohesion differently.
For example, English frequently uses definite article “the” to refer back to something previously mentioned. “The” doesn’t exist in Chinese; cohesion is achieved by repeating the same word over and over throughout the text.
AI can sometimes connect these references incorrectly, leaving readers confused about who or what is being discussed.
Beyond communicating information
Language doesn’t just communicate information. People also use it to establish relationships and connect with others.
Different languages and cultures can build similar relationships in different ways, depending on cultural or ideological differences. For example, government communications in Australia often use “we” and “you” to build a closer relationship with the general public. By contrast, another language – such as Chinese – may prefer third person and more formal words for distance and authority.
Translation is also about introducing new ideas to another language or reflecting evolving beliefs about existing expressions.
For example, terms such as “neurodiversity” may not have widely established equivalents in another language.
And when communities want to address stigma inherent to certain expressions, they might intentionally replace previously used words with new and emerging ones. For example, Chinese terminology referring to transgender people has evolved – more inclusive expressions are gaining recognition.
AI may reproduce older or less appropriate terminology, particularly where newer expressions are underrepresented in its training data. A translation can therefore be linguistically understandable yet socially insensitive.
So when can you trust AI translation?
Just as AI is a black box, your AI translation is also a black box. No large language model can guarantee your AI translation always comes out perfectly. But overall, AI translation tends to be more reliable when:
- the topic is familiar and straightforward across two languages, not involving cultural difference or sensitive issues
- the original text is written in perfect grammar and clear sentence structure with no ambiguity
- the two languages are both among most widely used languages (such as English, Chinese, French or Spanish), which means more training data is available for AI to learn from
- you provide effective prompts: clearly define the context, the reader, the purpose, and the language features of the translation task.
The translation provided by a chatbot will also be more accurate if it’s combined with neural machine translation technology – this is AI that’s dedicated to translation specifically and trained accordingly. It’s not a feature usually available on the common free AI platforms.
However, even when all the boxes are ticked, accuracy can’t be guaranteed; errors can crop up, hallucinations can happen. You trust it at your own risk.
Remember, generative AI doesn’t truly understand the complex and often ambiguous human experiences in a text. It may be good enough for understanding the general idea of a text, but its quality is never guaranteed without human input.
- is an Associate Lecturer in Translation and Interpreting, School of Humanities & Languages, UNSW
- This article first appeared in The Conversation
