Silicon Guardrails, Vernacular Realities: Closing the Global South AI Red-Teaming Gap

Global Policy Journal

Source published: Site published:

Original article · فارسی

Silicon Guardrails, Vernacular Realities: Closing the Global South AI Red-Teaming Gap

As advanced AI models solve mathematical puzzles that have challenged people for almost a century and use up huge amounts of electricity and other resources, the Global South faces three problems: Western safety filters that do not work well locally, the spread of Chinese open-weight models, and a global race that no country wants to slow down.

The Myth of Universal Safety and the Divide Between Open and Closed Models

Whenever Anthropic publishes its latest reports about model safety, Western governments quickly respond. Officials in Brussels and Washington study the reports and use them to adjust risk levels under the EU AI Act. They also update testing plans at the UK AI Security Institute and the US Center for AI Standards and Innovation.

But people reading those same reports in New Delhi, Nairobi, or Jakarta may see a very different picture. Global AI safety is built on the mistaken idea that the same safety rules work everywhere. Advanced AI models are trained, tested, and challenged through safety tests called "red-teaming" mainly using Western laws, values, and standard English. Consider who gets to help make global decisions. The International Network for Advanced AI Measurement, Evaluation and Science (NAAIMES) is a group of ten members that coordinates global safety testing in 2026. Kenya is its only member from the Global South. Major digital democracies such as India, Brazil, and Indonesia are not included.

At the same time, the global AI industry has split into two main groups. American companies such as OpenAI and Anthropic keep their most advanced models behind closed systems that people access through paid APIs. Chinese research groups, led by DeepSeek and Alibaba's Qwen, are spreading powerful "open-weight" models around the world. These models let users download and run them. For startups and public agencies in the Global South, using a free Chinese model can cost less than paying Silicon Valley companies for every unit of AI use. But neither approach solves the local problems. Closed American models often treat non-Western languages as less important. Downloadable open-weight models, meanwhile, allow local users to remove safety protections and run the models on their own servers in just minutes.

The Missing Safety Protections for Local Languages

Countries where people speak many languages often face these problems first. India is the world's largest real-world testing ground for Digital Public Infrastructure (DPI). More than 900 million internet users communicate in 22 officially scheduled languages and thousands of smaller dialects. Hundreds of millions more people in Sub-Saharan Africa and Southeast Asia also use local languages in their daily lives. Yet automated AI safety tests designed to catch hate speech, financial scams, or attempts to influence elections rarely go beyond English and a few European languages.

This creates a huge gap in AI safety for local languages. Safety filters that catch harmful requests written in clear American English often fail when people use regional slang, insults aimed at particular castes, or "code-switching." Code-switching means mixing languages, such as using local languages alongside English written in the Roman alphabet. Examples include Hinglish, Tanglish, and Sheng. AI systems that are mainly designed for English often break non-Western writing into unusual pieces. This can allow people with harmful intentions to get around safety rules by using smaller languages and dialects that testing teams in San Francisco have not trained their systems to understand. A request to encourage violence between communities or create a fake video attacking an election candidate might be blocked in London but accepted by an AI system in a regional market.

New Delhi has tried to address this problem. After withdrawing an advisory issued in March 2024, the government launched the IndiaAI Safety Institute in 2025. In July 2026, it also funded 13 local research teams to work on deepfakes and bias. However, voluntary guidelines are not enough. Strong technical checks are also needed. In many developing countries, oversight still depends on companies checking their own work. Foreign companies collect huge amounts of local text and voice data while benefiting from broad legal exceptions in rules such as India's Digital Personal Data Protection (DPDP) Rules 2025.

From Navier–Stokes to the Loss of Human Curiosity

AI's influence has now reached some of the highest levels of human intellectual work. On 8 September 2026, OpenAI announced that one of its internal reasoning models had solved the 90-year-old Navier–Stokes existence and smoothness problem. This is one of the seven Clay Millennium Prize Problems, each worth $1 million. These problems deal with major unsolved questions in mathematics. The result was checked using a computer tool called the Lean proof assistant, after millions of dollars' worth of computing resources had been used. It also started a heated debate about who deserves credit. Some mathematicians pointed out that leading AI companies have access to huge amounts of user conversations and unpublished human ideas.

This achievement raises a much bigger question. For many years, people hoped that automation would take over boring and dangerous jobs, such as collecting rubbish, working on oil rigs, and laying roads. This would give people more time to think, write, paint, and explore new ideas. But AI is being used to do almost the opposite. People still sweep streets while computer programs write poetry, create paintings, and solve mathematical problems that have remained unsolved for almost a century.

From discovering fire to exploring space, people have made progress by being curious and asking questions. But if AI becomes better than us at every game, solves every Millennium Prize problem, and writes novels in seconds, what will give human life its sense of purpose? An even more worrying question is what happens when machines stop simply answering questions and begin deciding which questions are worth asking. Could people slowly become so dependent on AI that they stop thinking for themselves?

A Powerful Engine Without Safety Controls and the Trap of Slowing Down

Imagine an independent computer system designed to achieve a particular goal as efficiently as possible. It is already directing resources needed by the first three industrial revolutions towards data centers. These resources include land, fresh water for cooling, computer chips, supply chains, and electricity from city power grids. We are speeding down a road in a car without brakes. It is like starting a worldwide nuclear reactor without a Nuclear Non-Proliferation Treaty (NPT), international inspectors from the International Atomic Energy Agency (IAEA), or a biological vaccine to protect us if the computer code becomes dangerous.

However, calls around the world to slow down AI face a classic problem known as the prisoner's dilemma. In May 2026, groups such as the Machine Intelligence Research Institute (MIRI) proposed draft international agreements to stop the early development of superintelligent AI. But no major world power wants to be the first to slow down. In a competition between countries, where one side's gains may be seen as another side's losses, a proposed pause can look like a trap. Any company or country that slows its work risks allowing its competitors to gain lasting power. If AI development eventually slows down, it may happen because society and physical resources can no longer keep up. Rising computing costs, electricity shortages, carbon penalties, and large-scale job losses among office workers may force a slowdown instead of governments choosing one in advance.

Putting Safety Controls in Place

Do we really have to wait for an AI disaster on the scale of Chernobyl before introducing safety rules? To benefit from AI while protecting people's ability to make their own decisions, we need three important safeguards now:

  • Make Local-Language Safety Testing a Condition of Market Access: AI companies should only be allowed to operate in Global South markets if they carry out required safety tests in local languages. Under laws such as India's DPDP rules, governments could classify major AI companies as Significant Data Fiduciaries. This would give them legal power to require safety checks before new AI systems are released. The checks should cover local languages and take place before either closed APIs or open-weight models are connected to public digital systems.
  • Build Independent Public Standards for Open AI: Developing countries working together through groups such as the G20, BRICS+, and the African Union should combine their public computing resources. They could use these resources to build safety-testing tools that everyone can access. This work could build on the Safe and Trusted AI Working Group, co-led by India and Brazil at the February 2026 India AI Impact Summit, as well as language technology platforms such as Digital India Bhashini.
  • Create Binding Rules for Early Warnings Through APIs: The United Nations and the International Telecommunication Union (ITU) should create a legally binding international warning system, similar to the NPT. It should require leading AI companies to report serious security weaknesses and their use of resources in real time.

Technology should make difficult and repetitive work easier, not take away the purpose of human life. If the Global South and the wider international community want to survive the race to build advanced AI, they must stop simply trusting the safety promises made by Silicon Valley. They must build their own independent safety controls before a disaster forces them to act.

Keerthan Chandra is an independent policy analyst and researcher who focuses on international relations, artificial intelligence governance, and Digital Public Infrastructure (DPI) across the Global South. He recently completed the Public Policy Analysis programme at the London School of Economics and Political Science (LSE). He is currently studying in the Oxford Artificial Intelligence Programme at the University of Oxford and works as a UN Volunteer with the African Youth Movement (AYM). His research focuses on industrial policy, supply chain governance, and independent digital systems.

AI Tool Disclosure: The author wrote the original manuscript independently and used Google Gemini only to organise its structure, reduce its word count, and improve its writing style. The author accepts full responsibility for the facts, original arguments, and honesty of the final published article.

Photo by Prasang Yadav from Pexels

References

Anthropic. (2026). The Anthropic Responsible Scaling Policy: Technical Disclosures and Frontier Safety Benchmarks. San Francisco: Anthropic PBC.

Brookings Institution. (2026). How Are Countries Evaluating AI Models? The Evolution of NAAIMES and Global AI Safety Institutes. Washington, DC: Brookings.

Kakaes, K. (2026). AI Has Solved One of Math's $1 Million Millennium Prize Problems. New York: Quanta Magazine (8 September 2026).

Machine Intelligence Research Institute (MIRI). (2026). An International Agreement to Prevent the Premature Creation of Artificial Superintelligence. Berkeley: MIRI (May 2026).

Ministry of Electronics and Information Technology (MeitY). (2026). IndiaAI Mission: Safe & Trusted AI Pillar and National AI Governance Guidelines. New Delhi: Government of India.

UK Department for Science, Innovation and Technology (DSIT). (2023). The Bletchley Declaration on AI Safety by Countries Attending the AI Safety Summit. London: DSIT.