The Most Significant AI Summit You've Never Heard Of
In New Delhi last week, something genuinely remarkable happened. Modi, Macron, Guterres, and Sam Altman were in the same room. 88 nations agreed to sign a declaration on AI. The global community appeared, for a moment, to be moving in the same direction on the most consequential technology of our era.
Then you read the fine print.
The New Delhi Declaration on AI is a meaningful diplomatic achievement and an almost completely toothless governance document. Understanding both of these things simultaneously is the only way to think clearly about what it means.
What the Declaration Actually Says
What It Promises
- "Safe, trustworthy, and robust" AI
- Democratizing AI access for emerging economies
- AI integration in health, education, and science
- Inclusive AI development that benefits all nations
- Cooperation on AI research and governance
What It Doesn't Say
- No enforceable measures
- No binding commitments for any signatory
- No real oversight mechanism
- No defined penalties for violations
- No timeline for implementation
The language of the declaration is aspirational by design. Diplomatic documents that require consensus among 88 nations can't be specific — specificity creates objections, and objections prevent signatures. The result is a document that everyone can agree to precisely because it commits no one to anything concrete.
The Summit's Central Tension
The most revealing moment of the New Delhi summit wasn't in the declaration itself. It was the clash between two visions of AI governance that played out in the corridors and in the public statements.
UN Secretary-General Antonio Guterres proposed an international commission with real authority — a body that could ensure meaningful human control over advanced AI systems, similar to how international bodies govern nuclear technology or civil aviation. The idea had a compelling precedent: global coordination on dangerous technology can work when there's sufficient political will.
Washington said no.
The United States' rejection of any binding global AI governance framework wasn't surprising — it's consistent with longstanding US policy on technology sovereignty — but it was revealing. It exposed a fundamental tension that no amount of diplomatic language can paper over: the countries most committed to developing advanced AI are the least willing to subject that development to international oversight.
The 5-Company Problem
5 American companies still control 90% of the world's foundational AI models. While 88 nations sign declarations, a handful of corporate boardrooms make the decisions that actually shape how AI develops.
The concentration of AI capability in a small number of private American companies creates a structural problem for global governance. Unlike nuclear weapons — which are controlled by sovereign states and therefore subject to interstate agreements — foundational AI models are controlled by corporations that are primarily accountable to shareholders, not to democratic publics or international bodies.
This doesn't mean governance is impossible. But it means that traditional treaty-based approaches, designed for state actors, are poorly suited to the actual structure of the AI industry. Effective AI governance needs to engage not just governments but the companies that actually build and deploy these systems.
AI Speed vs. Governance Speed
This speed mismatch is the central challenge of AI governance, and the New Delhi summit illustrated it vividly. The declaration being signed today addresses AI capabilities that were cutting-edge in 2024. By the time any binding commitments could theoretically be negotiated, implemented, and enforced, the technology will have moved two or three capability generations forward.
Governance frameworks designed for static technologies don't work for exponentially improving ones. This isn't a reason to abandon governance — it's a reason to fundamentally rethink how governance gets designed.
What Would Actually Help
Adaptive Regulatory Frameworks
Rather than treaties that take years to ratify and become obsolete before implementation, effective AI governance needs frameworks that can update through regulatory action rather than legislative amendment. The EU's AI Act includes provisions for this; most other approaches don't.
Company-Level Commitments
Since the companies building AI are the locus of real decision-making, governance frameworks that engage companies directly — through voluntary commitments with real accountability mechanisms — may be more immediately effective than intergovernmental declarations.
Shared Technical Standards
Technical standards for AI safety evaluation, transparency reporting, and incident notification could be more valuable than political declarations. Standards can be adopted by companies independently of government action and can be updated more quickly than legislation.
Capacity Building for Emerging Economies
The New Delhi Declaration's focus on democratizing AI access is genuinely important. Countries without significant AI capability are subject to AI's consequences without meaningful influence over its development. Building AI capacity globally creates more stakeholders with a vested interest in responsible development.
The Next Summit: Geneva, 2027
The declaration closes with a commitment to reconvene in Geneva in 2027. That's 18 months away — approximately three AI capability generations at current pace. The question for Geneva will be whether the intervening period produces real governance progress or simply more declarations.
The optimistic reading: New Delhi begins a process that builds toward more substantive commitments. Each summit creates slightly more political will and slightly more institutional capacity for actual governance. It's slow, but it's moving.
The pessimistic reading: By 2027, the technology will have moved so far that any framework agreed in New Delhi will be largely irrelevant to the actual AI systems being deployed. Governance will perpetually lag capability by two or three generations.
The realistic reading: Both are probably true. Governance will lag, but it will also improve. The question is whether the gap stays manageable or becomes catastrophic.
Frequently Asked Questions
The US position reflects concerns about sovereignty, competitive advantage, and the practical difficulty of enforcing international AI agreements. American officials argue that national regulatory frameworks (like potential AI legislation) are sufficient and that international governance risks slowing US AI development relative to China. Critics argue this position prioritizes short-term competitive advantage over long-term safety.
The EU AI Act is the most comprehensive binding AI regulation currently in force, and it does provide some model elements: risk-based classification, mandatory safety requirements for high-risk applications, and transparency obligations. Its limitation as a global model is that it applies only to systems used in the EU, creating potential regulatory arbitrage for systems deployed elsewhere.
Meaningful governance would include: mandatory safety evaluations before deploying frontier AI models, binding incident reporting requirements, liability frameworks that create financial incentives for safety, international information sharing on AI risks, and capacity building that allows all countries to participate in governance discussions meaningfully. None of these require a single global treaty — they can be implemented through combinations of national regulation, industry standards, and bilateral agreements.
Conclusion
The New Delhi Declaration is simultaneously historic and insufficient. 88 nations agreeing on anything is remarkable. 88 nations agreeing on principles without enforcement is, at best, the beginning of a conversation.
AI moves at the speed of innovation. Governance moves at the speed of diplomacy. Until we find ways to close that gap — through adaptive frameworks, company-level accountability, and technical standards — declarations will continue to describe a world we want without the mechanisms to build it.
The question for every subsequent summit isn't "did we achieve consensus?" It's "did we build the capacity to actually enforce what we agree on?"
Related Articles:
