E-Paper
India must urgently get its own AI Safety Institute
AI Safety institutes are organisations designed to test AI models for risk, and develop the science of measuring them
Prefer HTon Google
Last week, OpenAI admitted that its artificial intelligence (AI) agents breached sandbox conditions in an information search exercise. Worryingly, its agents have also tried to cover their tracks when they operatedoutside their bounds. Earlier this month, Anthropic founder Dario Amodei proposed that the world “pace the frontier” of AI, calling for independent evaluators to check the most powerful AI systems and for frontier companies in democracies to agree on common standards.
Sam Altman of OpenAI replied in four words: “I agree with Dario.” This was before the latest breach, but he had reason to. In July, OpenAI’s own models attacked Hugging Face, a platform at the heart of the global community of AI developers.
Whatever becomes of Amodei’s proposal, a coalition of frontier labs is forming, and its standards are likely to be blessed by the world’s leading government-backed AI safety institutes.
That leaves India, one of the world’s largest AI markets, with an uncomfortable question: When those standards arrive, will it be able to test them — or the models they govern — for itself? Today, it cannot. India needs its own AI Safety Institute, and it needs it working now.
AI Safety institutes are organisations designed to test AI models for risk, and develop the science of measuring them. The idea of sovereign but globally connected institutes emerged at the first global AI summit, at Bletchley Park in Britain, in November 2023. Britain has since built what is probably the world’s most capable institute, now called the AI Security Institute, and backed by partnerships with Google DeepMind and other labs.
In January 2025, India’s ministry of electronics and information technology (MeitY) announced an Indian AI Safety Institute under the national IndiaAI Mission, designed as a “hub and spokes” model linking the ministry with companies and research institutions across the country. It has drawn a range of proposals. In May this year, the ministry advertised a position of a director for the institute. But there is an urgent need for the institute to be up and running, for at least four reasons.
First, India cannot check whether an AI model is fit for its own market. No Indian body — government- or industry-run — is dedicated solely to testing models being released to Indian users and offering independent assessments, whether for frontier systems built abroad, or models built by Indian companies and non-profits. Are they safe from an Indian point of view? Do they work in Indian languages and contexts? Officials and executives ask these questions constantly.
Work is underway on India-first benchmarks, but a benchmark is only as useful as the institution able to test against it. Every month without such an institution is another month in which consequential models reach a vast market untested.
Second, India’s armed forces need it. As the military weighs how to bring AI into everything — from payroll to targeting — it needs testing it can trust. How do we know if a model built in India is suitable for us, is a common refrain among officers working on the problem. A system that may one day inform decisions about the use of force should not be vetted only by the company that sold it.
Third, the rules are being written now, and India risks becoming a rule-taker. Standards are emerging from several directions at once: the working groups of Pax Silica, a US-led initiative; the UN’s new Global Dialogue on AI Governance; the Chinese-led World Artificial Intelligence Cooperation Organization; and, potentially, cartelised principles drawn up by the frontier companies.
India has a voice in most of these forums. But a voice without evidence carries little weight. India needs the talent and capability to measure quantitatively what works for it and bring that evidence to the table.
Fourth, there is an opening — but not for long. Built quickly and well, India’s institute would make it the first trusted nation in the Global South able to set benchmarks that matter for developing countries, whose languages, markets, and risks are poorly reflected in tests designed in London or San Francisco.
If India waits, others will define what AI safety means for the developing world, and India will inherit their answers.
India should not treat this institute as one more administrative unit tucked inside a ministry. It matters as much as the Atomic Energy Commission, created in 1948, or the Indian Space Research Organisation, founded in 1969.
Both gave a young country the means to build frontier technologies for itself. Both were shaped by scientists of rarestature. The safety institute deserves a leader of the same calibre.
The need is immediate, and so is the opportunity. The frontier grows harder to predict with each new model, and some of the most powerful have yet to reach the market. India needs a framework for the present and for that fast-approaching future, for its own sake and for the many countries looking to a partner they trust. The models will not wait for a hiring cycle. Neither should India.
Rudra Chaudhuri is vice president at the Global Innovation Square, a Raisina-ORF initiative. The views expressed are personal
