Generative artificial intelligence (AI) - a system that creates text or translation based on a trained model - is increasingly used for instant translation. Many wonder: why learn a foreign language if a translator is at hand and often free? However, experts warn: fluent translation does not mean accurate.
The author professionally checks real translations between English and Chinese - the two most widely spoken languages. Even here there are telling errors. For example, a student called Australian professors "laid-back" because they came to class in flip-flops. AI translated the Australian word "thongs" as "thong underwear" rather than flip-flops. In less widely spoken languages there may be even more problems.
Errors arise for various reasons: a lack of training data, unsuccessful prompts (instructions for AI) or hallucinations - the confident presentation of false information typical of chatbots. But there is also a more fundamental reason - the complexity and flexibility of human language, as well as cultural differences.
At the word level, AI may choose the wrong meaning of a polysemous word. Verb tenses also create difficulties: in English the verb form explicitly indicates time, while in Chinese time is often clear from context. When translating from Chinese, AI may err and change whether an action happened, is happening or will happen.
At the sentence level, the phrase "You can see him in the store" may mean that he is in the store, or that you are in the store when you see him. A human usually resolves such ambiguity from context, while AI predicts the next words. If an error slips through, it is hard to notice: the text sounds natural, and without checking against the original the meaning may change imperceptibly.
At the level of a whole text, cohesion - the logical connections between parts - matters. English often uses the definite article "the" to refer to something already mentioned, while Chinese has no article - the connection is achieved through repetition. AI sometimes links references incorrectly, and the reader is confused about who or what is being discussed.
Language not only conveys information but also builds relationships. Australian government messages often use "we" and "you" for closeness with the audience, while Chinese may prefer the third person and formal words for distance and authority. AI may reproduce outdated or less appropriate vocabulary if new expressions rarely appear in its training data. The translation may be understandable but socially insensitive.
When can AI translation be trusted? It is more reliable if the topic is familiar and simple, does not involve cultural differences and sensitive issues; the original is written correctly and unambiguously; both languages are widely used (English, Chinese, French, Spanish), meaning AI has more data for training; you have given an effective prompt: clearly defined the context, reader, purpose and language features. Accuracy improves if the chatbot is combined with neural machine translation - a technology specially trained for translation (usually unavailable on free platforms).
But even under all conditions there are no guarantees: errors and hallucinations are possible. Generative AI does not understand the complex and ambiguous human experience in a text. It may be enough for the general meaning, but without human involvement quality is not guaranteed.