Regulation needs China on board

Why in the News

The global effort to govern artificial intelligence (AI) has split into rival camps. Twenty countries and the European Union (EU) called for keeping AI under human control, possibly through a global oversight body, but the US, China and India did not sign.

What models of AI governance now compete?

1. What it is: A global AI governance architecture is a shared set of rules on how powerful AI is built, tested and watched across borders, similar to the rules for nuclear energy.
2. Industry warnings: At the UN Security Council, the heads of Anthropic and OpenAI warned that badly managed AI could endanger humanity, a rare industry plea for regulation.
3. American doctrine: The US President’s science adviser rejected centralised international control. Under a White House voluntary accord, AI firms accept monitoring, auditors and board oversight as “morally binding” self-regulation, not law.
4. Four competing models: Each major actor governs AI differently:
   • the EU uses binding law, with stricter rules for riskier uses;
   • the US leaves it to the market and voluntary company pledges;
   • China keeps AI under state direction;
   • Organisation for Economic Co-operation and Development (OECD) principles and summit declarations add an international layer that binds no one.
5. The takeaway: No single model is enough, so the real task is combining them into one architecture.

What does cyber governance teach about AI rules?

1. UN Group of Governmental Experts (GGE): This UN panel of national experts first met in 2004. It spent a decade establishing that international law applies to cyberspace.
2. 2015 voluntary norms: Its report set 11 voluntary norms, endorsed by the UN General Assembly. Eg. States should not attack critical infrastructure and should report vulnerabilities.
3. Two rival tracks: The GGE deadlocked over self-defence in cyberspace. In 2018 the Assembly created a Russian-sponsored Open-Ended Working Group (OEWG) beside a US-backed GGE, both non-binding.
4. Value of soft norms: Even unenforced norms build habits of consultation and a common language.
5. Two lessons: Consensus norms need the principal adversaries at the table, and a decade-long process cannot keep pace with AI that shifts every few months.

What architecture would suit AI?

1. Layered design, not one treaty: AI needs several layers working together:
   • binding national law where frontier laboratories (firms building the most capable models) operate;
   • capability thresholds that trigger pre-deployment testing;
   • mandatory cross-border incident reporting;
   • a scientific body like the Intergovernmental Panel on Climate Change (IPCC) to establish shared facts;
   • a verification regime like the International Atomic Energy Agency’s (IAEA) nuclear inspections, based on compute monitoring (tracking the computing power used) for the most capable systems.
2. Closest existing proposal: The 20-nation call comes nearest to this design.

Why can no AI regime work without China?

1. Only other frontier power: China is the only country besides the US with genuine frontier AI capability.
2. Open-weight reach: Chinese open-weight models (free to download and run) power applications across Asia, Africa and Latin America, beyond any Western-only regime.
3. Beijing’s two-level approach: At home it uses algorithm registries and labelling of synthetic content. Abroad it presents AI as a development right and has proposed a world AI cooperation organisation.
4. Risk of rival blocs: Excluding Beijing creates a Western club and invites a parallel Chinese bloc of standards.
5. Minimum foundation: The US-China AI incident communication mechanism, agreed after the Trump-Xi summit, holds talks in November. It must widen into multilateral confidence-building (steps that reduce mistrust) open to both powers.

Challenges

1. Hard-to-verify compute: Chips and cloud capacity are spread across many firms, so compute monitoring is hard to enforce.
2. Irreversible open release: Once model weights are published, no regime can recall them. Eg. Meta’s Llama models.
3. Tech rivalry erodes trust: US export controls on advanced AI chips to China make Beijing wary of US-led rules.

Way Forward

1. Conditional participation: India should join any open framework, conditioning oversight on equitable access to compute and models.
2. Bridge role: India should use its hosting of the AI Impact Summit and service on cyber GGEs to link frontier powers with the Global South.
3. Stronger AI Safety Institute: India should strengthen its AI Safety Institute so Indian evaluators shape testing regimes.
4. Seat at incident reporting: India should seek a seat in any incident-reporting framework, since harms from abroad land in Indian markets.

Conclusion

The unresolved tension is that any workable AI regime needs Washington and Beijing, yet neither accepts rules the other writes. What to watch is whether their bilateral incident channel grows into wider talks with a seat for India.

Matching Previous Year Question

“[2026] Which of the following statements with regard to Large Language Models (LLMs) used in machine learning is/are correct? 1. LLMs assign probabilities to the next possible words and then pick the one with the highest probability. 2. LLMs process data through mathematical optimization to minimise prediction errors. 3. LLMs produce unbiased outputs. (a) 1 only (b) 1 and 2 only (c) 2 and 3 only (d) 1, 2 and 3 Answer: B”