The Asymmetry of Pace

There is a sentence I keep coming back to. I read it in a paper by an Estonian diplomat a few months ago, and I have not been able to shake it. The sentence was something like this. The most serious problem in AI governance is not ideological. It is temporal. I think that is the most honest description of where we are that I have read anywhere. We are not actually arguing about whether AI should be governed. Almost everyone agrees it should be. We are not really arguing about who should govern it, although the surface noise suggests otherwise. The deeper, harder, less photogenic problem is that the technology operates on one clock and the institutions that are supposed to govern it operate on another, and the gap between the two clocks is widening every year. The numbers A frontier model in 2020 had something like 175 billion parameters. By 2023 the largest models were roughly an order of magnitude larger. By 2025 the largest models, depending on how you count, were larger by another factor of three to ten. Capability benchmarks that researchers thought would take a decade to crack were being broken in eighteen months. The MMLU benchmark, which tests broad knowledge across 57 academic subjects, went from below random in 2019 to above human expert level by 2024 (Hendrycks et al., 2021; subsequent evaluations). Meanwhile, the average United Nations treaty negotiation, from initial proposal to entry into force, takes between seven and twelve years. The Convention on Biological Diversity took five years to negotiate and another two to come into force. The Paris Agreement was unusually fast by UN standards and still took roughly four years from initial mandate to entry into force. The Pact for the Future, which most observers consider a process success, took two full years of preparatory work and is now sitting on a 2027 review cycle for one of its key components, the Global Digital Compact. If you stack those two timelines on top of each other, the picture is uncomfortable. By the time a binding multilateral AI instrument finishes negotiation, the technology it was meant to govern will have been superseded twice over. This is the asymmetry of pace. Why this is harder than it looks It is tempting to say the solution is just to make the UN faster. Everyone who works near a multilateral institution has thought this at some point. The problem is that "the UN" is not slow because it is poorly run. It is slow because consensus is slow, and consensus is the source of its authority. A faster UN, without consensus, would not be the UN. It would be the G7 with a flag. So we cannot simply accelerate the institutions. What we can do, and what is quietly happening, is build new institutional shapes that move at different speeds in parallel. A slow layer for binding principles. A medium layer for thematic dialogues and reviews. A fast layer for scientific assessment and technical updates. This is roughly what the architecture coming out of the Pact for the Future is trying to do. Three mechanisms now being tested The first is the Independent International Scientific Panel on Artificial Intelligence. Forty members, appointed by the General Assembly in February 2026, serving in personal capacity, selected from 2,600 applicants across 140 countries (UN Secretary-General, 12 February 2026). The panel was deliberately modelled on the IPCC. Its job is not to make policy. Its job is to produce annual evidence-based assessments that policy bodies can then react to. The panel's first report goes to the Geneva dialogue in July 2026. Annual assessments are still slower than frontier model releases. But they are roughly an order of magnitude faster than treaty cycles. That is not nothing. The second mechanism is the Global Dialogue on AI Governance itself. The Dialogue is interesting because it deliberately does not produce binding outcomes. Each session ends in a "co-chair summary" rather than a negotiated text (UN Global Dialogue FAQ, 2026). This was a design choice. The architects of the Dialogue recognised that if they tried to produce binding text on a two-year cycle, they would fail. So they built something more modest. A standing conversation. A space where governments and stakeholders show up regularly, share evidence, build relationships, and slowly accumulate shared understanding. This sounds weak. It is actually how international law often begins. The customary norms of diplomatic protection, of state responsibility, of humanitarian intervention all started as patterns of state behaviour that eventually crystallised into rules. The Geneva dialogue is trying to compress that crystallisation process from centuries into years. The third mechanism is bilateral and plurilateral. The US-UAE AI Acceleration Partnership, announced in May 2025, is one example. The OECD AI Principles, which were adopted in 2019 and have already been revised once, are another. These instruments move much faster than the UN because they have fewer parties. The trade-off is obvious. They have less legitimacy. But they can do things the UN cannot, and they can do them in months rather than years. What is missing None of these three mechanisms address the deepest problem, which is that the people who will live with the technology longest, by which I mean the generation now in school, have almost no formal voice in any of these processes. This is not a sentimental point. It is a structural one. If AI policy is being set on a twenty year horizon, then the people whose lives will be most shaped by it are not the diplomats negotiating now. They are the students sitting in lecture halls in Singapore and Lagos and Sao Paulo and yes, Abu Dhabi, who will inherit whatever is decided. And the existing mechanisms for getting their input are, to put it generously, ad hoc. UN DESA's Youth Office runs consultations. Some countries include youth delegates in their UN missions, currently 40 out of 193 member states as of the 79th General Assembly (UN DESA, 2024). UNESCO runs occasional youth dialogues on digital governance. But there is nothing systematic. Nothing that ensures that when the Scientific Panel publishes its annual report, there is a parallel youth scientific advisory that has actually engaged with the same evidence. Nothing that ensures the Geneva dialogue's working groups have intergenerational representation built into their composition. This is the gap I would close, if I were designing the next iteration of this architecture. A modest proposal I am wary of proposing institutional innovations because most of them are bad. But I want to gesture at one shape that seems to me worth thinking about. What if each of the Scientific Panel's annual assessments were accompanied by a parallel process, run by a small standing youth scientific advisory, that produced its own assessment of the same evidence? Not a separate document with separate findings. A commentary, embedded in the official report, written by people in their twenties who will live with the consequences for fifty years. The cost would be small. The Scientific Panel already has a secretariat. Adding a parallel youth track would require perhaps another six positions, funded by member states that already fund youth programmes. Norway, Finland, the Netherlands, the UAE, Australia, all have the budgets and the appetite. The benefit would be larger than it looks. It would not just produce better reports. It would create, for the first time, a structural pathway from being a national youth delegate at a Third Committee debate to being part of the global scientific assessment process. That pathway is what is missing right now. We bring young people to the UN, we give them a few minutes at a microphone, and then we send them home. We do not give them anywhere to go next. This is a fixable problem. Why this matters now The 2027 review of the Global Digital Compact is going to happen whether we are ready or not. The Geneva dialogue's second session is scheduled for New York in May 2027. The Scientific Panel will produce its second annual report ahead of that meeting. The window for designing the next layer of architecture is now, in the two years between Geneva 2026 and New York 2027, when the institutions are new enough to be reshaped but old enough to know what they need. I do not know how this will go. I am genuinely uncertain whether the mechanisms now being built will close the asymmetry of pace or whether they will end up running too far behind to matter. What I am certain of is that the work is not finished. The architecture as it stands is a first draft. It is open to amendment. And the people best placed to amend it are the ones who can see the gap most clearly. That is the generation currently treated as the audience of these conversations. It should, soon, be treated as a participant. References Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2021). Measuring Massive Multitask Language Understanding. International Conference on Learning Representations. UN Secretary-General. (2026, February 12). Statement upon the appointment of the members of the Independent International Scientific Panel on AI. Retrieved from https://www.un.org/sg/en/content/sg/statements/2026-02-12/ United Nations. (2026). Global Dialogue on AI Governance: FAQ. Retrieved from https://www.un.org/global-dialogue-ai-governance/en/faq United Nations General Assembly. (2025, August 26). Resolution A/RES/79/325 establishing the Global Dialogue on AI Governance and the Independent International Scientific Panel. UN DESA. (2024). 79th General Assembly Youth Delegate Programme Update. Division for Inclusive Social Development. United Nations. (2024, September 22). Pact for the Future, including the Global Digital Compact and the Declaration on Future Generations. A/RES/79/1. Ada Lovelace Institute. (2026, May). The case for a global AI governance floor. Retrieved from https://www.adalovelaceinstitute.org/blog/the-case-for-a-global-ai-governance-floor/