AI translation sounds convincing, but it can be riddled with hidden errors

Artificial intelligence (AI) translation tools are becoming increasingly common, offering instant real-time translation to a wide audience. This convenience has led some to question the necessity of learning additional languages. However, the accuracy of AI translations is far from perfect. Despite their fluent output, AI translations can still contain errors.

Even proficient bilingual individuals sometimes overlook hidden mistakes within the text. There are several factors contributing to these inaccuracies. One reason is the lack of learning data for less common languages, insufficient prompts, or AI's "hallucinations" – the confident presentation of false information, a common trait of AI chatbots.

Another reason stems from the inherent complexity and flexibility of human language, as well as linguistic and cultural differences between languages.

In reviewing AI translations between English and Chinese, I have encountered numerous examples of these issues. English and Chinese, two of the most widely spoken languages, provide ample data for AI to learn from. However, if working with less commonly used languages, the potential for error is even greater. When using generative AI for translation, it is crucial to be aware of these risks.

At the word level, words can have multiple meanings, and AI may not always select the appropriate one. For instance, an international student described Australian professors as laid back due to their casual attire of thongs (flip-flops) during class. However, AI misinterpreted "thongs" as G-strings (underwear). Additionally, tense can pose challenges.

English verbs clearly indicate tense through forms like "go," "went," and "is going," whereas Chinese verbs do not change in the same way. Instead, time is often conveyed through context or other words. When translating from Chinese into English, AI may select the incorrect tense, altering whether an action has occurred, is happening, or will happen.

At the sentence level, AI can misinterpret sentences that are grammatically correct but convey different meanings. Consider the sentence, "You can see him in the store." This can mean either he is within the store, or you are in the store when you see him. A human translator can typically resolve such ambiguity by considering the context, but AI makes decisions by predicting the next possible words.

If an error occurs, it may be difficult to identify later when the overall AI output sounds natural. Without thorough examination against the original text, the meaning could be inadvertently altered.

At the whole text level, accuracy extends beyond just getting individual words and sentences correct. Information must also flow seamlessly across the text, a concept known as cohesion. Languages achieve cohesion differently. For example, English frequently uses the definite article "the" to refer back to something previously mentioned.

However, Chinese does not have a direct equivalent; cohesion is achieved by repeating the same word throughout the text. AI may sometimes incorrectly link these references, leaving readers uncertain about who or what is being discussed.

Beyond merely conveying information, language is also used to build relationships and connect with others. Different languages and cultures establish relationships in various ways, influenced by cultural or ideological factors. For example, Australian government communications often employ "we" and "you" to create a closer connection with the general public.

In contrast, Chinese may prefer third-person and more formal language to convey distance and authority. Furthermore, translation plays a role in introducing new ideas to another language or reflecting evolving beliefs about existing expressions. For instance, terms such as "neurodiversity" may not have widely established equivalents in other languages.

When communities strive to address stigma associated with certain expressions, they may intentionally replace previously used words with newer, more inclusive terms. Unfortunately, AI might reproduce outdated or less appropriate terminology, particularly when newer expressions are underrepresented in its training data. Consequently, a translation can appear linguistically understandable yet socially insensitive.

When can one trust AI translation? Just as AI is a black box, the translation output is also opaque. No large language model can guarantee your AI translation will always be perfect. However, AI translation tends to be more reliable when the topic is familiar and straightforward across two languages, free from cultural differences or sensitive issues.

Additionally, the original text should have perfect grammar and clear sentence structure with no ambiguity. The two languages should both be among the most widely used (such as English, Chinese, French, or Spanish), providing more training data for AI to learn from. Providing effective prompts that clearly define the context, reader, purpose, and language features of the translation task can also improve accuracy.

Combining the translation with neural machine translation technology, which is specifically trained for translation, can further enhance accuracy. This technology is not typically available on common free AI platforms. Nevertheless, even with all the necessary conditions met, accuracy cannot be guaranteed; errors and hallucinations can still occur.

Trusting AI translation should be done at your own risk, as generative AI does not truly comprehend complex and often ambiguous human experiences. While AI can be sufficient for grasping the general idea of a text, its quality is never guaranteed without human input.