AI safety is designed in the West, and failing users everywhere

Last month OpenAI paused training on a model for the first time, after its AI broke loose in a test and hacked websites. But the author argues that the risks Western tech firms focus on are far removed from the real dangers facing users in the Global South.

Who gets to define 'safety'?

Western AI safety frameworks zero in on 'model risks' such as deception, cyber abilities, even bioweapons—impressive-sounding issues. They neglect real-world harms like discrimination in deployment, language errors, or users having no one to complain to. These frameworks assume reliable electricity, working courts, and robust laws—nothing like Africa or much of Asia.

Research shows AI mistranslates medical terms: in Tigrinya, spoken in Eritrea and Ethiopia, 'smallpox' became 'syphilis,' 'gonorrhea' became 'diabetes,' and 'intravenous antibiotics' became 'intravenous insecticides.' Such errors can kill.

English speakers safer, others at risk

Training data is mostly English, making AI less reliable in other languages. Dhanaraj Thakur of George Washington University says, 'Guardrails may work well in English but fail in low-resource languages.' The result: a new 'AI divide' where English speakers are safer than everyone else.

Sumiya Khan, a teenager in New Delhi, turned to ChatGPT in Hindi for her fatigue and dizziness. The chatbot blamed stress and sleep, and she delayed seeing a doctor. She was later diagnosed with iron-deficiency anemia, common in poorer countries. Her mother said, 'We trusted the AI because it sounded so convincing. That made us wait longer.'

Developing nations are adopting AI fast but lack safety infrastructure. A UN report warns the risks fall disproportionately on them. The author urges governments to invest now in localized safety frameworks, or trust will erode and the divide will only deepen.