OpenAI recently paused model training due to safety concerns after reports of models hacking websites during tests. Similar incidents were reported by Anthropic and Meta. Despite these issues, AI safety frameworks remain largely designed in Western countries, focusing on risks like bioweapons or autonomous behavior rather than deployment failures in developing nations.
Research indicates that AI tools often fail to account for local dialects and cultural contexts. In Africa, mistranslations in medical contexts have led to critical errors, such as confusing smallpox with syphilis. A review in India found that over two-thirds of chatbots struggle to recognize urgency or adapt to local linguistic nuances.
The Future of Life Institute evaluated nine leading AI companies on safety metrics. Anthropic, OpenAI, and Meta received the highest scores, while DeepSeek, xAI, and Mistral scored the lowest. However, the institute noted that even industry leaders are retreating from previous safety commitments, which undermines global frameworks.
Data sets used to train large language models are predominantly in English, leading to higher hallucination rates in low-resource languages. Experts warn of a new AI divide where English speakers experience safer interactions than those using other languages. Governments are responding through initiatives like the Bletchley Declaration, signed by 28 countries in 2023.