The Refusal, and What It Is Not
In February 2025, at the AI Action Summit in Paris, the United States declined to sign the closing statement on "inclusive and sustainable" artificial intelligence. The United Kingdom did the same. Five months later the White House released an AI Action Plan whose title carried the whole argument: Winning the Race. David Sacks, the AI and crypto czar, and Michael Kratsios at the Office of Science and Technology Policy had already made the position plain. American models, American rules, and no outside body holding a veto over either.
OpenAI and Anthropic had asked for the opposite. Sam Altman wanted a "democratic" standard for frontier systems. Anthropic's submission to OSTP recommended stronger export controls, formal evaluation requirements, and argued against a proposed ten-year moratorium on state-level AI laws. Washington said no. Not because the proposals were too weak or too expensive, but because the request carried a premise the administration rejects: that legitimacy in AI governance might originate somewhere other than the United States.
Regulatory Capture in Safety Clothing
Compliance costs scale with headcount, not with risk. This is the oldest fact in regulatory politics, and it applies to AI as neatly as it once applied to banking under Basel III. A mandatory red-teaming report, an incident disclosure regime, a model card reviewed by counsel — a laboratory with 3,000 employees absorbs these inside a quarter. A two-person team in Bangalore shipping fine-tunes on top of an open-weight model cannot.
So when frontier labs call for "stringent standards," the useful question is what the standard costs and who pays for it. The EU AI Act's threshold for general-purpose models with systemic risk — 10^25 floating-point operations for training — captures perhaps a dozen firms worldwide. That is not a drafting accident. It is a moat with a legal preamble. OpenAI's March 2025 submission went further, framing Chinese-developed models such as DeepSeek as state-subsidised threats, which converts a competing product into a national security object and invites export controls that function as trade barriers.
The Arms Race That Eats Its Own Footing
Race is a metaphor about speed. It says nothing about destination. In current Washington vocabulary the competitors are nation-states, the prize is capability, and the horizon is measured in months. Absent from that vocabulary: the Kenyan annotator paid between $1.32 and $2 an hour to label the violent material that trains safety filters. The Chilean town whose aquifer was drawn down for a Google data centre. The Irish grid, where data centres consumed about 21 percent of metered electricity in 2023.
None of this is speculative harm parked in a distant future. These are billing records, water permits, and hospital spreadsheets. The 2019 Obermeyer study in Science found that a health algorithm applied to roughly 200 million Americans each year systematically under-referred Black patients to care. No superintelligence was required. A biased proxy — historical cost used as a stand-in for medical need — was enough.
The View from Nairobi and New Delhi
African Union members adopted a Continental AI Strategy in 2024 that treats data sovereignty and domestic compute capacity as prerequisites rather than luxuries. India's position, voiced through its G20 presidency and its own summit diplomacy, has been steady: standards written elsewhere are standards a country complies with rather than shapes. Brazil's PL 2338 bill spent years in debate largely over how much of an AI regime should be imported from Brussels.
Researchers who write about data colonialism — Nick Couldry and Ulises Mejias are the standard citation — argue that the extraction of behavioural data repeats the pattern of nineteenth-century resource extraction. Raw material leaves the periphery. Value is added at the centre. The finished product is sold back, along with the terms of use. A global AI safety standard drafted in San Francisco and negotiated in Brussels does not interrupt that circuit. It formalises the circuit and attaches a liability shield to the firms at the top.
Labour, Data, Water
Humanitarian organisations tend to arrive at AI governance through labour and privacy rather than through extinction scenarios. Their concerns are mundane and testable. Does a clinic in Accra know where its patient records go? Does Kenya's Data Protection Act of 2019 reach a server in Virginia? Can the contract cleaners on a data centre campus afford to live near it? These are the questions that decide whether a deployment helps or harms the people near it.
No declaration of principles settles them. Procurement terms settle them. Audit rights settle them. Jurisdiction settles them. At that layer the tug-of-war between US dominance and international oversight stops being abstract. It determines whether a regulator in Lagos can compel disclosure from a company headquartered in California, and whether anyone answers when the request arrives.
Standards Are Infrastructure
Standards are unglamorous machinery. They decide which plug fits the socket, which measurement counts, which body certifies. The US decision to rename its AI Safety Institute the Center for AI Standards and Innovation in mid-2025 was not cosmetic; it shifted the mandate from studying risk toward promoting deployment. Meanwhile the Council of Europe's framework convention on AI, opened for signature in 2024, and UNESCO's 2021 recommendation on the ethics of AI continue to exist — signed by some states, ignored by others.
A world with several incompatible standards is not a world without standards. It is a world where the largest market's rules become the default, because complying with them is the price of access. Countries in the Global South rarely get to choose between American rules and no rules. The real choice is between American rules and the cost of building something else.
What Oversight Could Actually Look Like
Altman's phrase about democratic AI is doing a great deal of work. Democracy, in that usage, means a family of values associated with liberal Western societies, not a process in which affected populations hold a vote. If the latter were the goal, the proposals would look different: mandatory disclosure of training-data provenance, enforceable data protection with cross-border reach, compute accounting that includes water and grid load, and a right of redress for anyone harmed by an automated decision.
Such measures are dull. They require no summit and generate no headline. What they require is a regulator willing to act against a US firm, and a counterpart in Accra or Jakarta with the standing to demand the same treatment. Until that exists, "global AI standards" remains a phrase used by the companies that benefit from them and the governments that write them. Everyone else gets the press release, and a bill for the water.
