The UN and AI
The Body That Banned Chemical Weapons Is Now Trying to Govern ChatGPT
The institution that took 30 years to build a global consensus on landmines is now being asked to regulate a technology that changes every six months.
The Idea
In 2024, the United Nations General Assembly passed its first resolution on artificial intelligence — unanimously, which is itself a minor miracle for a body where members routinely disagree on the time of day. The resolution called for AI to be 'safe, secure, and trustworthy' and for its benefits to be shared across all nations. It was also almost entirely non-binding, which tells you everything about the genuine difficulty of the problem. The core tension is structural. The UN was designed to manage relationships between nation-states, but AI governance requires managing relationships between nation-states, corporations, research labs, and individual developers — some of whom have more compute than most governments. The big AI companies are headquartered in two countries, the US and China, which happen to be in active geopolitical competition. Asking the UN to referee that dynamic is a bit like asking a homeowners' association to mediate a war. There are also serious technical challenges beneath the diplomatic ones. Unlike nuclear weapons, which require rare materials and enormous infrastructure, capable AI models can be trained and deployed with surprising speed by increasingly small teams. There is no uranium to track. Verification — the backbone of every serious arms control treaty — becomes nearly impossible when the dangerous thing is a set of mathematical weights on a hard drive. What the UN can do, and what it is genuinely trying to do, is build shared language, shared norms, and shared infrastructure — particularly for the countries that are not America or China and currently have very little say in how this technology develops.
In the World
In September 2024, the UN Secretary-General António Guterres convened a 'Summit of the Future' in New York, where AI governance was one of the central agenda items. Alongside it, the UN established a new advisory body — the International Scientific Panel on AI — modelled loosely on the Intergovernmental Panel on Climate Change, the IPCC. The idea: get the science agreed upon first, then use it as a foundation for policy, the same way climate scientists built the evidentiary base that eventually produced the Paris Agreement. The comparison is instructive and a little sobering. The IPCC was founded in 1988. The Paris Agreement was signed in 2015. That is 27 years from panel to treaty — and the treaty still has no enforcement mechanism. AI researchers are not known for their patience with 27-year timelines. Meanwhile, the countries with the most to gain from international AI governance are not the US or China — they are the 130-odd nations in the Global South that currently have almost no domestic AI industry, no seat at the table where safety standards are being written, and no ability to shape the technology that will increasingly shape their economies. For those countries, the UN process is not slow and frustrating — it is the only process that includes them at all. That is the quiet argument for why this ungainly, non-binding, diplomatically tortured effort might still matter more than it looks.
Why It Matters
Most coverage of AI governance focuses on the drama between Washington and Beijing, or the lobbying battles in Brussels and Westminster. The UN angle tends to get filed under 'well-meaning but toothless.' That framing is worth questioning. Every major international norm we now take for granted — on biological weapons, on nuclear non-proliferation, on the law of the sea — started as a non-binding resolution that serious people dismissed as wishful thinking. The UN does not enforce things; it legitimises them, which turns out to be a necessary precursor to enforcement. For anyone thinking about who actually controls AI, the UN story is a useful corrective to two lazy assumptions: that governance is purely a technical problem (solvable if the engineers get it right), and that it is purely a geopolitical problem (solvable if the superpowers cut a deal). The messier truth is that it is a coordination problem involving actors with radically different timescales, incentives, and levels of power — and the UN, for all its dysfunction, is the only room where all of them are technically present. The question worth carrying is not whether the UN will succeed. It is what 'success' could even look like when the thing being governed refuses to sit still.
A Question to Ponder
If the countries building the most powerful AI systems have the least incentive to accept outside constraints on them, what would actually have to change — technologically, politically, or economically — for meaningful international AI governance to become possible?
Get a new one of these every morning.
Start learning with Thinkable