Ramen, Vodka and Saké: Who Really Won the AI Summit
Three summits. Sixty governments. Zero binding commitments on labour. The Meridian examines who really won the AI governance race, and what the settlement means for workers everywhere.
The question they did not put on the agenda was the only one that mattered.
Three AI summits in three years. Bletchley Park in October 2023, staged in a Victorian mansion in Buckinghamshire with the gravity of a first act. Seoul in May 2024, more procedural, better attended, less watched. Paris in February 2025, where the AI Action Summit produced a declaration signed by sixty governments, celebrated for its ambition and noted, in quieter corridors, for its absence of binding commitments.
Each produced communiqués. Each produced language about frontier models, existential risk, responsible development, and international cooperation. Each brought together governments, civil society, and the carefully selected academics who had agreed that artificial intelligence was both enormously promising and enormously dangerous and that something, broadly, should be done.
What none of them put on the agenda was the only question that maps onto the actual stakes: when capital finally owns the capacity for cognitive labour, what does the worker have left to sell?
That question was absent because it does not produce communiqués. It produces redistributive politics. And redistributive politics is not what AI summits are convened to do.
Three delegations arrived at every summit with incompatible interests and identical language.
Japan brought ramen: scalable, practical, designed for mass consumption, optimised for delivery. Japan’s domestic case for AI is not ideological. It is demographic. The country’s population is declining. Its workforce is shrinking. Its old-age dependency ratio (the number of retirees supported by each working-age adult) is among the worst in the OECD and deteriorating further. For Japan, AI automation is not a threat to labour. It is a solution to the absence of it. Tokyo’s negotiating position at every summit was consistent: we need AI to function, we need it safe enough not to alarm the public, and we need international standards that do not obstruct domestic deployment. Japan got what it came for.
Russia brought vodka: potent, unsubtle, and largely indifferent to the conversation everyone else was having. Moscow’s interests in AI governance forums are what its interests in international forums have always been: veto leverage, legitimacy maintenance, and the domestic right to do as it pleases. Russia signed what it needed to sign. It departed when it was convenient to depart. Its AI development proceeds under state direction, with military and surveillance applications as the primary horizon. The governance debate, from Moscow’s vantage point, is theatre. It watches with professional attention and without anxiety.
The saké arrived from Silicon Valley, though it was poured with considerable ceremony in Washington, London, Paris, and Tokyo. Saké is subtle. It presents as refined. It is consumed at the correct temperature, in proper vessels, with appropriate ritual. The ritual, in this case, was safety. The proper vessel was the summit itself. The temperature was controlled by the companies whose models were nominally under discussion: companies that had funded the research institutes whose academics sat on the advisory panels, contributed to the think-tanks that drafted the background papers, and seconded the technical staff who briefed the government delegations. Those same companies, after every summit, returned to building the systems the summits were convened to govern.
The saké tasted like restraint. It was not restraint.
The economists have been arguing about automation and employment for two centuries. The first wave was mechanical: the loom replaced the hand-weaver, the threshing machine replaced the farm labourer, the steam engine replaced the horse. Labour absorbed the displacement, eventually, by moving into new sectors that the machines created. The transition was not gentle. The handloom weavers of England did not transition smoothly into factory supervisors. They starved, in many cases, over a generation. The macroeconomic outcome was eventually positive. The distributional outcome, in the interim, was catastrophic for those who lived through it.
The second wave was computational: the spreadsheet replaced the accounts clerk, the word processor replaced the typist pool, the automated assembly line replaced the production worker. Again, labour moved into IT support, logistics coordination, and the expanding service sector. Again, the transition imposed costs that the macroeconomic models aggregated away.
The third wave is different in one structural respect, and that difference is what the AI summits did not discuss.
The first two waves automated physical and routine cognitive tasks. They left untouched the class of work that economists call non-routine cognitive: analysis, synthesis, creative production, complex judgement, interpersonal communication. The implicit assumption of every labour economist writing about automation in the twentieth century was that non-routine cognitive work was the refuge, the sector into which displaced labour would perpetually migrate because machines could not do it.
That assumption is now being tested. It may not survive.
The large language models that underpinned the AI summit discussions can draft legal contracts, write financial analysis, produce engineering specifications, generate software code, synthesise medical literature, and structure a political economy argument. They are not doing any of these things perfectly. But perfection is not the threshold that matters economically. The threshold that matters is whether the AI-assisted worker produces at lower cost than the unassisted worker. At that threshold, across a widening range of non-routine cognitive tasks, the answer is already yes.
“The saké tasted like restraint. It was not restraint. The safety narrative was the vessel. The settlement was what the vessel contained.”
When a machine makes a worker more productive, the question of who captures the resulting surplus is a political question, not an economic one. It is settled not by markets but by the balance of power between capital and labour at the moment of settlement.
In the first industrial revolution, the settlement was extremely unfavourable to labour. Capital captured the overwhelming share of productivity gains through the first half of the nineteenth century. The labour share of income did not begin recovering until trade union organising, factory legislation, and political enfranchisement shifted the balance. That shift took roughly seventy years.
In the information technology revolution, the settlement was mixed. Some workers (particularly those with the technical skills to work with the new tools) saw significant wage gains. The broader labour share, however, declined across most OECD economies from the 1980s onward, a trend that IMF research has attributed, in part, to automation’s effect on the bargaining power of mid-skill workers.
The AI settlement is being negotiated now. The AI summits were, in one precise sense, part of that negotiation: an exercise in determining which institutions would govern AI development, which standards would apply, and therefore which interests would be served. The outcome, across three summits, was a governance architecture built around voluntary commitments, company-led safety evaluations, and national regulatory discretion. There are no binding international instruments on AI-driven labour displacement. There is no multilateral framework for taxing AI-generated productivity gains and redistributing them. The question of who captures the surplus from cognitive automation was not, formally, on any summit’s agenda.
This was not an oversight. It was a settlement.
| Sector | AI Exposure | Global South Dependence | Surplus Capture |
|---|---|---|---|
| BPO / Customer Service | Very High | Very High (Philippines, India, Kenya) | Capital |
| Software Development | High | High (India, Eastern Europe, Nigeria) | Capital |
| Legal Services | High | Low | Capital |
| Financial Analysis | High | Medium (India, South Africa) | Capital |
| Content / Media Production | High | Medium (Global) | Capital |
| Administrative & Clerical | Medium-High | High (Global South urban) | Mixed |
| Manufacturing (Advanced) | Medium | High (China, Vietnam, Bangladesh) | Mixed |
| Healthcare (Diagnostic) | Medium | Medium | Mixed |
| Agriculture | Low | Very High (Sub-Saharan Africa, South Asia) | Unclear |
The AI summit process has been, in its composition, a G7 affair with invited guests.
Bletchley Park was convened by the United Kingdom and attended principally by the United States, the European Union, and their partners. China participated, notably and carefully. The Global South was present in the margins, in the form of a handful of invited government representatives who could observe the terms being set but were not positioned to shape them.
This matters for two reasons.
The first is that the countries of the Global South will bear a disproportionate share of the labour displacement costs. The sectors most exposed to AI automation in the near term (business process outsourcing, software development, legal services, financial analysis, content production) are precisely those that have driven formal employment growth across South and Southeast Asia and parts of Africa over the past twenty years. India’s IT services industry, the Philippines’ BPO sector, Kenya’s growing tech workforce: these are built on the proposition that cognitive labour is transferable across borders and that lower-cost providers can compete on quality. AI automation challenges that proposition at its foundation.
The second reason is that the countries of the Global South will have no meaningful say in whether or how the productivity surplus from AI is redistributed. The governance architecture being built in London, Washington, Brussels, and Tokyo will determine those terms. Developing economies, as in previous waves of technological change, will receive the technology after the terms of its deployment have been settled elsewhere, and will adapt to those terms rather than shape them.
The EU AI Act, passed in 2024, represents the most substantive legislative response to date. It is a European instrument. Its risk-classification framework, its requirements for high-risk system transparency, its prohibitions on certain applications: these reflect European political priorities and European institutional capacity. They will, over time, acquire extraterritorial effect through the Brussels Effect, as companies seeking EU market access build to European standards. This is better than nothing. It is not a global settlement.
A global settlement on AI and labour would look something like this: an international agreement that ties AI productivity gains to redistributive obligations, that requires companies above a certain scale to contribute to retraining funds indexed to displacement metrics, that gives governments in lower-income countries enforceable rights in the data and intellectual property generated within their borders. No such agreement exists. No such agreement is under negotiation. The summits did not discuss it.
Robert Frost’s farmer insisted on the fence without being entirely certain why. His father had built it. His father’s father before him. Good fences make good neighbours.
The AI summits built fences too. Around liability. Around competitive advantage. Around the question of what governments may require companies to disclose, and what companies may decline to disclose in the name of proprietary protection. Around the labour question itself, which was kept, carefully and consistently, outside the perimeter.
Who won the AI Summit?
Capital won. Specifically, the capital concentrated in the small number of companies that control the frontier models, the data centres, the semiconductor supply chains, and the talent pipelines that constitute the actual infrastructure of the AI economy. Those companies participated in the safety conversation, contributed to the safety discourse, helped design the safety architecture, and emerged from the safety process with their competitive positions intact, their models unencumbered by binding regulatory constraints, and the labour question still unanswered.
Japan won a productivity solution to a demographic problem. Russia won the right to continue doing what it was already doing. The European Union won a regulatory framework it can apply domestically and that its companies will find burdensome and its competitors will route around.
The workers did not win. The workers were not at the table.
The question of what labour has left to sell when capital owns the mind is not one the summits were designed to answer. It will be answered anyway, in productivity statistics and wage data and employment figures and trade union membership rates and the quiet arithmetic of households whose income does not keep pace with the economy growing around them.
The ramen was filling. The vodka was strong. The saké was excellent. The bill, as usual, will be paid by someone who was not invited to dinner.
- AI Safety Summit, Bletchley Park, 1–2 November 2023. Bletchley Declaration signed by 28 countries and the EU. Available at gov.uk. The declaration acknowledged risks from frontier AI models but contained no binding enforcement mechanism.
- AI Seoul Summit, 21–22 May 2024. Produced the Seoul Declaration and the Frontier AI Safety Commitments, signed by sixteen leading AI companies. Commitments were voluntary. Monitoring and verification mechanisms remained company-led.
- AI Action Summit, Paris, 10–11 February 2025. Joint Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet, signed by approximately sixty governments. The United States and the United Kingdom did not sign the joint statement. Labour displacement was not addressed in the final text.
- International Monetary Fund, World Economic Outlook, April 2024. Chapter on the macroeconomics of artificial intelligence. IMF estimated approximately 60 per cent of jobs in advanced economies are exposed to AI, with roughly half facing potential displacement and the other half facing augmentation. For emerging market economies, the exposure figure is approximately 40 per cent, with less capacity to capture augmentation gains.
- Daron Acemoglu and Pascual Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor,” Journal of Economic Perspectives, 33(2), Spring 2019. The paper’s framework distinguishes between automation (which displaces labour) and new task creation (which reinstates it). The authors argue the balance has shifted unfavourably for labour since the 1980s.
- International Labour Organization, The Impact of Artificial Intelligence on the Quantity and Quality of Jobs, ILO Brief, 2023. Estimated that generative AI could expose 300 million full-time jobs globally to automation or significant augmentation. Clerical work identified as the most immediately exposed occupational category.
- European Parliament and Council of the European Union, Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (AI Act). Establishes a risk-based classification framework with prohibited AI practices, high-risk system requirements, and transparency obligations. Does not address labour displacement or the redistribution of AI-generated productivity gains.
- McKinsey Global Institute, A New Future of Work: The Race Between AI and Workers, 2023. Projected that generative AI could automate tasks accounting for up to 70 per cent of workers’ time by 2030. BPO, software development, and knowledge work identified as highest-exposure sectors.
- World Bank, World Development Report 2019: The Changing Nature of Work. Examined automation’s differential impact on emerging market versus advanced economies, noting that developing countries face a compressed transition window with less institutional capacity to manage displacement.
- OECD, OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. Found that 27 per cent of jobs in OECD countries are in occupations highly exposed to automation. Workers in exposed jobs reported lower job satisfaction and greater employment insecurity.
© 2026 The Meridian. All rights reserved. Published at themeridian.info. Reproduction, distribution or republication without written permission is prohibited. For licensing enquiries contact The Meridian editorial team.