Don’t wait for a consensus

Why in the News

Leading US artificial intelligence (AI) companies have signed a voluntary Accord on Super Intelligence. Separately, 28 countries have endorsed Finland and Norway’s “A Call for Control of Frontier AI Models”. Neither the US nor China joined the call, so others must decide whether to act without them.

What is frontier AI governance, and why is it urgent?

1. What it is: Frontier AI means the most capable models, built by a handful of companies. Governing it means rules on testing and release, like the clinical trials a new drug must pass.
2. Industry’s preferred pace: Anthropic, OpenAI and Google DeepMind back “pacing the frontier”, meaning an internationally coordinated adjustment of how fast AI advances, so risks can be managed.
3. Civil society demand: Civil society groups want a global moratorium on frontier models until binding safeguards exist.
4. Trigger incidents: Unauthorised and deceptive behaviour by frontier AI agents (systems acting on their own) and debate at the UN General Assembly have raised the stakes.
5. The takeaway: Calls for governance are louder, but actors want very different things, from self-policing to a full pause.

How does the White House accord differ from the Finland-Norway call?

1. Accord on Super Intelligence: The “morally binding” accord promises internal controls, independent external evaluation and an independent board committee, but names no standards or enforcement.
2. Three-step agenda: The Finland-Norway call seeks:
   • mandatory pre-deployment testing and independent evaluation;
   • common incident reporting standards;
   • an institutional mechanism for standard-setting and verification.
3. Unclear path: The call sets out significant proposals but no route for putting them into practice.

Where do the US and China stand?

1. US position: Washington prefers permissionless innovation (building first, without prior approval) and rejects any “globalist scheme of control for superintelligence”.
2. China’s position: Beijing treats the UN as the main channel for AI governance but distrusts Western-centric agendas.
3. Bilateral opening: The two recently opened a bilateral dialogue on advanced AI, but broader consensus remains elusive.

Can China’s WAICO offer an alternative path?

1. World Artificial Intelligence Cooperation Organisation (WAICO): Launched by Beijing in July with 29 founding members, it is headquartered in Shanghai.
2. Functions: It will promote “supply-demand matching” (matching AI supply with demand across countries), standard setting and convergence on AI governance.
3. Open membership: Any country may join. The Global Partnership on Artificial Intelligence (GPAI) and the G7 Hiroshima Process instead assume members share liberal-democratic values.
4. Formal forum, limited reach: As an intergovernmental organisation, it can build formal consensus beyond soft law (non-binding guidelines). Its China-led image may confine it to the Global South.

Should the world wait for a consensus?

1. No excuse for inaction: Missing US-China consensus cannot justify inaction by others, or let the tech industry define governance as the White House accord does.
2. India’s stake: India is not at the frontier but sees extensive deployment of frontier models, so it needs accountability safeguards.
3. Claim to a voice: India’s stand on digital sovereignty supports a greater say in the rules, a rational basis to back the Finland-Norway proposal.

Challenges

1. Toothless pledges: Voluntary company pledges carry no penalty for breach.
2. Access for evaluators: Independent testers need model access and compute that a few firms control.
3. Forum fragmentation: GPAI, the Hiroshima Process and WAICO may set incompatible standards.

Way Forward

1. Safeguards for deployed AI: India should build legal and institutional safeguards for AI already in everyday life, not only for future artificial general intelligence.
2. Risk-based classification: Regulators should impose stricter duties on high-risk uses such as medical diagnosis.
3. Testing capacity: India should equip its AI Safety Institute to run independent pre-deployment tests.

Conclusion

Without a US-China consensus, frontier AI governance is split between company self-policing and middle powers seeking binding checks. What to watch is whether India formally backs the Finland-Norway call and builds domestic safeguards.

About AI Regulation

1. EU AI Act (2024): The European Union sorts AI into four risk tiers, from unacceptable (banned) to minimal.
2. China’s Generative AI Regulations (2023): Mandate security assessments and algorithm registration.
3. India’s light-touch model: India has no dedicated AI law and relies on the Information Technology Act, 2000 and the Digital Personal Data Protection Act, 2023.

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”