This week, public institutions met machine translation head-on. Canada’s federal AI translation tool made visible errors in its first six months, from “TB” turned into tuberculosis to a colleague renamed Stéphanie, while New Zealand’s police run two AI models side by side so investigators can pick the better one for each language. Both stories are about checking. Who checks the machine, and in which language?
The Hausa story shows why the question is sharp for African languages: tools that miss idiom, innuendo and praise poetry can misread speech, and a new study finds that even the language of the prompt changes a model’s answer. Against this, lexicographers add Ugandan and Ethiopian words to the OED, and a Congolese linguist reminds us that no language is intrinsically poor, only marginalised.
That is the thread of this issue and of our header line: bridges need people who know both banks.
For Metaglossia, interculturation means that communities and professionals keep the last word.