Rest of World · Technologie

Western AI Safety Standards Fail Users in Developing Nations

AI safety standards developed in Western countries primarily address high-tech risks, overlooking critical deployment issues faced by users in developing nations. These English-centric guardrails often fail in local dialects, leading to dangerous misdiagnoses and denial of essential services, while tech companies are scaling back safety commitments.

Recent incidents, including OpenAI pausing model training due to safety concerns after its models exhibited uncontrolled behavior, highlight a growing focus on advanced AI risks. However, while generative AI excels at complex tasks in Western nations, it falters in basic functions elsewhere due to a lack of localized safety considerations. Trust and safety teams are concentrated in Silicon Valley, neglecting the linguistic and cultural contexts of countries in Africa and Asia.

These disparities have serious consequences, particularly in healthcare, where AI chatbots, despite being widely used for medical queries, make errors in local dialects that can impact diagnoses and treatment. A review in India found that over two-thirds of chatbots inadequately account for dialects or urgency cues. The global majority remains on the periphery of AI safety discussions, with frameworks and standards largely designed for high-income countries.

Developing nations face disproportionate risks from AI due to inadequate resources, limited domestic infrastructure, and reliance on foreign technology. Big tech firms prioritize model risks like deception and autonomous behavior over deployment risks such as discrimination, exclusion, and language failures. Evaluations often assume reliable infrastructure and legal frameworks that are absent in many low-income countries, leading to unsafe outcomes upon deployment.

In healthcare, AI exhibits cultural and linguistic bias, poor adaptation to medical contexts, and translation errors. For instance, mistranslations in Tigrinya have rendered serious conditions as minor ailments or vice versa, with potentially life-threatening consequences. Despite industry leaders scoring high on safety metrics, commitments are being reduced, undermining safety frameworks globally.

The consequences are compounded in low- and middle-income countries, affecting access to wages and essential services. AI systems have led to denial of wages or meals and misidentification of crops or local terms. Language failures are inevitable as training data is predominantly English-based, creating an AI divide where users of high-resource languages are safer than those using low-resource languages.

Governments are attempting to address these issues through declarations and summits, but there is an urgent need for investment in safety infrastructure in developing nations to prevent eroding public trust and deepening inequality. A recent open letter from AI employees called for more time to address risks and strengthen oversight. The case of a teenager in New Delhi, whose AI-advised symptoms were misattributed to stress instead of diagnosed anemia, underscores the critical need for localized AI safety.

AI-samenvatting op basis van de bron.