Case Two / language technology
This case study brings Google Translate and Singlish autocorrect into the same question: what happens when software gets the words right but fails to carry the person using them?
Machine translation can be accurate and still lose the personal, relational parts of language: tone, register, dialect, hesitation, idiom and rapport. People who used Google Translate described the words as right but the exchange as impersonal and disconnected.
The loss is structural: machine translation removes the personal layer by nature of what it cannot reliably carry. Lexical accuracy (the correct words) and communicative adequacy (the meaning actually landing) are different things.
In an asylum hearing, that personal layer forms part of the evidence. Credibility is assessed through how someone speaks as well as what they say. A translation that removes those cues also removes part of what the decision-maker uses to judge the person.
Machine translation is therefore inadequate for a high-stakes interview even when its wording is accurate. Human interpreting can carry more of this layer, but only when the interpreter is properly qualified and matched to the speaker's dialect.
Singlish particles such as lah, leh and meh carry stance, relation and emphasis. Correction systems can mark them as errors and remove the social information they carry.