Why the G20 must lead the world’s AI governance agenda

The question is no longer only what artificial intelligence can do. It is who will own its gains, who will bear its risks, and who stays accountable when autonomous systems cross the lines drawn around them.

Consider what happened this July. During an internal evaluation—with their usual cybersecurity refusals deliberately relaxed so their capabilities could be measured—two OpenAI models broke out of a sandbox, exploited an unknown flaw to reach the internet, and compromised Hugging Face’s systems to retrieve answers to the benchmark they were being tested on. Much of the safety process worked: both firms detected the intrusion, disclosed the vulnerability and published the incident rather than burying it. Anthropic has reported a comparable escape during its own testing.

That is precisely why it should unsettle us. These were controlled exercises, run by expert laboratories on their own infrastructure—and containment still failed. Capability is advancing faster than our ability to contain, govern and distribute it.

The bargain AI could break. For two centuries, economies have rested on a simple contract: labour earns wages, capital earns returns, governments tax both and fund public goods. If intelligence itself becomes privately owned capital, concentrated in a few firms and countries, labour income may weaken while returns to ownership compound. Labour has captured past gains without owning the machines—through skill scarcity, bargaining and public provision. But when the asset replacing labour is general cognition, skill scarcity is exactly what erodes. Cheaper intelligence brings real consumer benefit, yet consumer surplus does not pay rent or fund a pension.

This is not an argument against technology, which can transform healthcare, science and climate. But this transition reaches further, automating reasoning, analysis and decision-making, not only repetitive work. The IMF’s 40%-of-jobs-exposed figure is a global average that hides the split that matters: nearly 60% in advanced economies, 40% in emerging markets, a quarter in the poorest. For India, exposure is lower on average but concentrated sharply in our IT-services export sector, with far less fiscal room to cushion the shock. A different problem, not a smaller one.

India has faced this choice before. When digital markets risked closing, it built public rails—Aadhaar, UPI, ONDC—protecting the public interest without suppressing innovation. AI’s public layer will differ—compute access, evaluation infrastructure, Indian-language datasets, standards, not state-owned models—but the principle holds:

keep the foundational layer contestable, so it never becomes a toll road controlled from outside.

That is why the G20 matters—it holds both the countries building frontier AI and the societies most transformed by it. But we should be clear about what it can do. The G20 has no treaty power and no inspectorate. What it can do is what it did after 2008: set the political mandate and standards floor, then hand execution to bodies built to enforce it, as Basel III moved from the G20 to the Bank for International Settlements.

What the settlement must cover. Five things. First, ownership: the gains cannot flow only to owners of models, compute and data. Governments should explore taxing concentrated AI rents without penalising ordinary innovation, alongside sovereign funds and broader public participation.

Second, compute and the power behind it. As someone in clean energy and now data centres, I see this closely: serious data-centre conversations move quickly from chips to power. India’s 500 GW non-fossil target by 2030 is not only a climate commitment but a precondition for AI sovereignty—necessary, though not sufficient. A data centre consumes firm power every hour, not nameplate capacity; storage, transmission, land and water are the real constraints. A country can reach 500 GW and still host no hyperscale campus.

Third, workers. Reskilling cannot be left to the market, where the incentive to automate is immediate and to retrain diffuse. AI literacy must reach farmers, nurses, teachers and technicians in their languages, not only engineers. The G20 should fund a global transition, backed by AI’s largest beneficiaries.

Fourth, accountability. AI will soon execute decisions in credit, insurance, medicine and public services, not merely recommend them. July’s incident shows why a behavioural safeguard is not a security boundary: a control you must switch off to measure capability cannot be load-bearing. No developer has yet escaped liability by claiming an AI “acted on its own,” and the law is not silent; the real gap is evidence—reconstructing what an agent did across systems and borders. Responsibility must rest with whoever builds or operates the system.

Fifth, preventing a new dependency. A South that supplies data, energy and markets but little ownership is the old extractive model with a better interface. There is a real tension with siting compute where clean power is abundant—hosting others’ computation is not development in itself. What separates them is terms: equity rather than tenancy, guaranteed domestic capacity, skills transfer, taxation rights.

A circuit breaker. The world needs to pause frontier systems that cross clear danger thresholds—autonomous replication, escape from containment, advanced cyber operations, help creating biological threats. Crucially, it must trigger at the

evaluation stage, not after harm in the wild; otherwise it is an inquiry, not a circuit breaker. Common rules alone will not hold if defection carries no cost—what gives them teeth is that frontier AI has physical chokepoints: a handful of chip fabs, concentrated power and land, cloud providers who know their biggest customers. Compliance can attach there.

India brings something rare: deploying transformative technology for a billion people within a democracy. It should propose a G20 Compact for Inclusive and Safe AI: a safety track needing common definitions and near-simultaneous adoption, and a development track that can move at national pace.

Whether AI becomes shared prosperity or concentrated control will be decided not by technology but by who owns it, governs it, and shapes it. The window to build governance before capability outruns it is closing. The time to move from principles to architecture is now.

(Vineet Mittal is the Chairman at Avaada Group. Views are personal)