Efforts to strengthen AI safety are accelerating, but safeguards don't work with the same effectiveness everywhere. In developing countries, language gaps and cultural differences can turn everyday AI mistakes into serious real-world consequences.
اضافة اعلان
AI companies devote enormous time and resources to thinking about what might happen as their models become more capable and advanced. OpenAI even took the unusual step last month of temporarily pausing training on one of its models over safety concerns, at a time when the industry faces risks ranging from autonomous model behavior to increasingly sophisticated cyber capabilities.
But there's another AI safety problem that's easy to overlook: the safeguards already built into these systems don't necessarily work with the same effectiveness for all users.
A new report from Rest of World points out that AI safety efforts remain heavily concentrated on the needs of wealthier, English-speaking countries.
This can mean users in parts of Asia, Africa, and other developing regions encounter basic but potentially dangerous problems, such as a chatbot's inability to properly understand their language.
AI Safety Looks Different Outside Silicon Valley
Many of the largest AI companies have extensive trust and safety teams, but the standards used to evaluate models are largely shaped in high-income countries.
Researchers say this creates blind spots when the same models are used in countries that differ in language, infrastructure, laws, and cultural expectations.
Healthcare offers a particularly troubling example. Research examining natural language processing AI systems used in Africa uncovered errors in translating medical terminology.
For instance, machine translation in Tigrinya confused smallpox with syphilis, and even translated the phrase "intravenous antibiotics" as "intravenous pesticides."
The problem isn't limited to translation alone.
Research has also found that AI models may produce fabricated or incorrect information at a higher rate when dealing with languages for which relatively limited training data is available.
Similarly, safeguard systems designed to efficiently detect harmful requests in English may be less effective, or easier to bypass, when used in other languages.
This issue is becoming increasingly important as more people rely on chatbots for health information and help with important decisions.
This effectively creates a new kind of AI divide, where two people might use the same product but receive vastly different levels of protection simply because they speak two different languages.
AI Adoption Is Outpacing the Development of Safety Measures
This issue is gaining greater importance as enthusiasm for using AI grows very rapidly in many of these markets.
China perhaps offers the clearest example of this. AI use there has extended far beyond tech companies, with schools and local governments adopting the technology.
AI has even reached dumpling shops in Beijing, where one business began giving customers AI-linked tokens and developed AI-based tools to manage waiting lines.
This enthusiasm isn't a problem in itself, but researchers warn that safety infrastructure needs to expand at the same pace. Governments have already begun moving to address this issue.
AI safety has featured in a number of international initiatives, including the Bletchley Declaration and the AI summit held in India this year, while China has proposed mechanisms for managing AI risks in developing countries.
But the bigger challenge lies in ensuring that the concept of "AI safety" doesn't become limited to preventing some future, advanced model from bypassing the restrictions placed on it or launching a cyberattack.
For the millions of people already using these tools, safety may be far simpler than that: a chatbot that properly understands their language, recognizes when a situation is an emergency, and doesn't confidently give them wrong information that could put them at risk.
As AI spreads into the details of daily life, these issues are no longer marginal or rare cases, but have become a core part of what it means to make AI safer.
