When Sam Altman told a US Senate committee that the future of artificial intelligence should not be decided by "labs in San Francisco alone," the line drew nods from both parties. Weeks later, the Trump administration declined to sign the Paris AI Action Summit's communiqué on inclusive AI, and Michael Kratsios, the White House science and technology chief, made the position plain: American rules would be American, calibrated to win a race. Two camps, one argument. Both assume the question is who governs — Washington or Silicon Valley. Neither asks who is governed, and whether anyone thought to invite them.
The Regulatory Capture Behind "Global Standards"
"Global standards" sounds neutral. Standards, in practice, are drafted by whoever can afford the flight. The G7 Hiroshima Process, the OECD AI Principles, the Bletchley and Seoul declarations, the working groups of the Global Partnership on AI — these are technical committees staffed largely by engineers from the same dozen firms. OpenAI and Anthropic have argued for licensing regimes and third-party evaluations; the EU AI Act's threshold of 10^25 FLOPS (floating-point operations per second) for models deemed to pose "systemic risk" shows how such numbers acquire legal force. Those numbers also function as barriers.
A licensing regime that requires audits costing millions, plus compliance counsel and documentation teams, is survivable for a company valued near $157 billion. It is fatal for a university lab in Nairobi or a civic-tech cooperative in Dhaka. Regulatory capture rarely requires corruption; it requires asymmetry. Open-source maintainers, public-health researchers, and small model developers have no lobbyists in Brussels or Washington, so the rules that emerge — ISO/IEC 42001, the NIST AI Risk Management Framework, the GPAI obligations under the EU act — reflect the business models of the firms writing them. Altman's complaint is not wrong. It simply skips the follow-up: the labs he worries might be excluded are the ones drafting the exclusions.
Catastrophe Talk and the Harms Already Here
Anthropic's Dario Amodei warns of "kill switches" and engineered pathogens. OpenAI's charter invokes existential risk. This framing has purchase in Washington because it is flattering. It casts AI governance as a Manhattan Project for the species, run by physicists in a desert. Meanwhile, harms that already exist are documented, unglamorous, and distributed unevenly across the world.
In 2019, researchers at UC Berkeley found that a health-care algorithm used on roughly 200 million Americans each year assigned Black patients lower risk scores than equally sick white patients, rationing care along a colour line drawn by a regression. Pulse oximeters, standard equipment from Lagos to Manila, overestimate blood oxygen in patients with darker skin, a bias traced to device calibration. Kenyan workers contracted through the outsourcing firm Sama to label toxic text for OpenAI earned about $2 an hour and later described trauma symptoms. Data centres in Ireland consumed 21 percent of the country's electricity in 2023; in Memphis, turbines powering xAI's supercomputer drew air-quality complaints from a predominantly Black neighbourhood. None of this fits the supervillain plot. All of it is on the record.
Obsessing over a hypothetical superintelligence postpones the governing of a real one. Aligning a future god and aligning a present-day triage tool are different problems, and treating the first as urgent makes the second look provincial.
Compute as Colonial Infrastructure
When policy is framed as a race the United States must win, the rest of the world stops being a participant and becomes a market. The substrate of AI — fabrication at TSMC, lithography at ASML, chip design at Nvidia, cloud capacity in a handful of American hyperscalers — sits in a few jurisdictions and moves under export controls. Countries with neither fabs nor data centres occupy a familiar position: suppliers of minerals and cheap labour, buyers of finished product.
Scholars such as Michael Kwet call this digital colonialism, extraction without occupation, and dependency theory offers an older vocabulary that still fits. The African Union adopted a Continental AI Strategy in July 2024; India's AI Mission subsidises GPU access for domestic researchers. Both are bids for sovereignty, and both have thin leverage. When Washington restricts H100 exports, it is not governing AI in the abstract. It is deciding which countries get to build rather than rent.
Seats at the Table, or a Different Room
A genuine alternative would begin with questions nobody in Paris or Washington was asked. Who audits a diagnostic model deployed in a Kenyan hospital? Who pays for the water a data centre drinks in drought-stricken Chile? What does consent mean for a Punjabi farmer whose land records train a credit-scoring model sold back to his own bank? Institutions exist that could host such work — the ITU, UNESCO, and the UN's High-level Advisory Body on AI, which proposed a scientific panel modelled on the IPCC. Proposals are cheap. Funding and political will are not.
One test is procedural: any instrument called an international standard should be negotiated under rules where a majority of voting members represent neither the wealthiest governments nor the firms those governments host. A second is material: publish compute and energy disclosures as a condition of market access rather than as voluntary ESG reporting, so that the cost of inference is visible where the grid strain lands. A third is financial: build public compute in low- and middle-income countries instead of asking them to adopt safety frameworks designed elsewhere and priced for someone else.
Altman is right that AI should not be decided by San Francisco labs alone. Kratsios is right that a sovereign state may set its own rules. Both statements dodge the harder claim: a standard that excludes four-fifths of humanity is not a global standard. It is a licence, granted to the few, and the queue for renewal runs through Washington.
