A pair of flip-flops became underwear in one AI translation, a UNSW lecturer says
Alisa Tian of UNSW argues in The Conversation that smooth AI translations can hide errors, so readers should check them against the original.
In brief
Fluent AI translation can still be wrong
A UNSW lecturer reviewed real English-Chinese AI translations. In one, flip-flops became underwear. Her advice: check the output against the original text.
One AI translation turned flip-flops into underwear. That is the most vivid example in an analysis piece by Alisa Tian, an associate lecturer in translation and interpreting at UNSW, published in The Conversation.
Tian bases her argument on her professional work reviewing real AI translations between English and Chinese. The examples are her own observations. The piece cites no quantified study.
The flip-flop case
In one translation she examined, an international student called Australian professors laid back since they turned up to class in "thongs". The AI read the word as G-strings, a type of underwear, rather than flip-flops.
Why the errors slip through
Her central point is simple: fluent is not the same as accurate. A translation can read smoothly and still say something different from the original. Even good bilinguals can miss such slips.
She lists several trouble spots:
- Words with several meanings. The AI may pick the wrong one.
- Tense. Chinese verbs do not change form the way English ones do. Time comes from context or other words, so the AI can get the tense wrong when going from Chinese into English.
- Ambiguous sentences. "You can see him in the store" can mean he is in the store, or that you are. A human translator often uses context to decide. The AI, she says, predicts likely next words.
- Links across sentences. English uses "the" to point back to something already mentioned. Chinese has no "the" and repeats the same word instead. The AI can connect these references wrongly, leaving readers unsure who or what is being discussed.
Where the mistakes come from
Tian names several causes: too little training data, weak prompts, and hallucination, which means a chatbot stating false information with confidence. Other errors come from real differences between languages.
She adds that English and Chinese have a lot of data for AI to learn from. Problems could be worse for less common languages.
What she suggests
Her advice is to check AI output against the original. Her piece also says when trust is reasonable: when the topic is familiar and straightforward, the source text is clear and unambiguous, both languages are widely used, and the prompts are effective. She adds that a chatbot combined with neural machine translation technology will be more accurate. Even then, she says, accuracy is not guaranteed.
The same text also appeared on Stuff South Africa, Mirage News, inkl and The Sen Times. These are republications, not separate reporting.
The limits are clear. This is one expert's opinion built on examples, with no measured error rates, no tool names and no testing of specific products. How often these errors happen is not shown.