A patient reading that they have been given intravenous insecticides has no way to tell whether the mistake came from the clinic or the translation layer. In Tigrinya, spoken by about 9 million people in Eritrea and northern Ethiopia [17], machine translation produced that sentence in place of "you have been given intravenous antibiotics," turned smallpox into syphilis and gonorrhea into diabetes [18]. Elizabeth Orembo, a fellow at Research ICT Africa, told Rest of World that mistranslations like these can be life-threatening [19]. Researchers examining natural language processing in healthcare in Africa found linguistic and cultural bias alongside poor adaptation to medical contexts and outright translation errors [16].
Orembo's account of why that survives review is the part worth borrowing. Large AI firms concentrate on model risks: deception, autonomous behavior, cyber capabilities, aiding bioweapons [12]. Deployment risks get much less attention, among them discrimination, exclusion, surveillance, language failures, and the inability of affected communities to get anything remediated [13]. A risk taxonomy is also a test plan. What sits outside it will not be found by the people paid to look.
New-market teams often assume that a multilingual model has the language covered, but users write the way they speak and carry the urgency in the phrasing, not in the vocabulary a benchmark tests. A review in India found that more than two-thirds of chatbots did not adequately account for dialects or recognize urgency cues [6], which leaves under a third that did [7]. That is a depth-of-use number rather than a volume number, and it is the metric that belongs in a launch review, measuring how often the assistant missed the signal that a question was an emergency rather than how many sessions the pilot generated.
The second failure mode is environmental. Trust and safety evaluations assume reliable electricity and internet connectivity, functioning courts, robust data protection laws, formal labor markets, and a press and civil society that report failures [14]. Where several of those are absent, the support thread that would tell you the product is hurting someone never gets written, which is how, in Orembo's words, "a model can pass every frontier safety evaluation and still produce unsafe outcomes when deployed" [15]. The UN has reported that risks fall disproportionately on developing nations, which adopt AI more slowly but have inadequate resources, limited domestic AI infrastructure and more dependence on foreign technology [10].
There is no universal trust and safety standard to procure against - each company writes its own framework and its own enforcement methods [9], so procurement alone cannot settle this. Urvashi Aneja of Digital Futures Lab, who is preparing a report for the UN on AI safety in developing nations, told Rest of World that the evaluation frameworks, the standards and the overseeing institutions have largely been designed in and for a small set of high-income countries [8].
So this is for the person shipping a health-adjacent assistant into six locales next quarter and answering for it when a clinic complains. Two questions per locale, each answered with a name rather than a percentage. Has someone tested the language as users actually write it, dialect and urgency phrasing included, on questions where a wrong answer changes a treatment decision? And when it fails, does the user reach a human, and does the complaint land somewhere a person reads it? A locale that cannot answer either question with a name is unsupported, whatever the launch tier calls it, and the localization work that would change the answer sits in the safety budget rather than the growth one.
