The San Francisco–Washington Binary Is a Power Map, Not a Debate
In 2024, Sam Altman of OpenAI asked for global AI standards. Anthropic executives echoed him. They spoke of democracy, safety, and the collective good. The White House under Donald Trump, with Michael Kratsios shaping tech policy, rejected the plea. America first, they said. American competitive advantage. The press framed this as San Francisco versus Washington. Two power centers tugging at a rope. That frame is a map of who holds the rope, not of who gets hurt when it snaps.
Altman and Anthropic are not neutral referees. They build the systems. Their calls for global rules arrive after they have trained models on data scraped from the open web, after they have hired thousands of low-wage annotators, after they have signed computing deals that lock in Nvidia GPUs for years. When an architect asks for building codes, ask who writes the codes and who can afford the permits. A global standard that requires safety audits, red teams, and compute thresholds is a barrier. OpenAI can pay. A startup in Nairobi cannot. Regulatory capture does not need a bribe. It needs a compliance form.
Regulatory Capture Wears a Human Face
Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework both promise caution. They classify models by dangerous capabilities. They talk about CBRN risks, cyber offense, autonomous replication. The language is technical and sober. It also functions as a moat. Only labs with legal teams, evaluation teams, and spare compute can meet the tests. Smaller developers in India, Brazil, or Nigeria are locked out of the market before they ship a product. The EU AI Act's general-purpose AI rules already show this pattern. Companies with lobbyists in Brussels shaped the thresholds. Companies without them must comply or leave.
Altman's plea for democracy deserves scrutiny. No one elected him. His 'collective good' is defined by boardrooms in San Francisco and London. The US rejection is not a rejection of corporate power. It is a demand that corporate power remain American. Kratsios and Trump want Nvidia chips to stay ahead of Huawei. They want American model weights to dominate. The fight is not regulation versus competition. It is one set of monopolists versus another. The Global South is the audience, not the negotiator.
The Humanitarian Gap: Bias, Labor, Carbon
While executives debate existential risk, concrete harms accumulate. In healthcare, algorithmic bias is not hypothetical. Pulse oximeters overestimate oxygen saturation in darker skin, and AI triage tools inherit that error. The eGFR equation for kidney function once included a race coefficient that delayed transplants for Black patients in the US. Similar tools are exported to hospitals in Kenya, India, and Brazil without local validation. Skin cancer datasets remain overwhelmingly white. An LLM that summarizes clinical notes may hallucinate a dosage. The patient in a rural clinic has no redress.
Labor displacement is already here. Business process outsourcing in the Philippines and India employs millions. Generative AI now drafts insurance claims, customer emails, and legal discovery. The workers who lose those jobs do not appear in AI safety papers. Content moderators in Kenya, paid roughly two dollars an hour, watch beheadings and child abuse so that ChatGPT and Claude can appear safe. They suffer PTSD, anxiety, and retaliation for unionizing. Anthropic's executives worry about a future where AI kills humanity. Those moderators worry about a present where AI already injures them.
The environmental cost is measurable. Training a model like GPT-4 consumed tens of gigawatt-hours. Inference adds more. Data centers in Arizona, Chile, and South Africa draw water for cooling in drought-prone regions. Bitcoin mines and AI clusters compete for the same grid. E-waste from obsolete GPUs ends up in Agbogbloshie, Ghana, where young men burn cables for copper. The compute that powers a chatbot in Oslo may dry a well in Santiago. These are not externalities. They are design choices.
Nuclear Treaties and Fossil Fuels: What Governance Can and Cannot Do
History offers analogies, not blueprints. The Nuclear Non-Proliferation Treaty of 1968 created the IAEA safeguards system. It slowed some proliferation. It never disarmed the five recognized nuclear weapons states. India, Pakistan, Israel, and North Korea stayed outside or cheated. The treaty worked for the powerful and constrained the weak. Fossil fuel governance is similar. The UNFCCC and the Paris Agreement set targets. COP summits give oil majors observer status and backroom access. Emissions rise. The companies that caused the crisis help write the rules.
AI governance has no IAEA. It has the G7 Hiroshima Process, the UK AI Safety Institute, the US-EU Trade and Technology Council, and China's interim measures. These are clubs of rich states and their corporate advisers. The Global South appears as a recipient of capacity building, not as a rule maker. Private firms hold more power than many sovereign states. Nvidia's market capitalization exceeds the GDP of most African nations. TSMC and ASML control chokepoints that no UN agency can inspect. Export controls on chips are geopolitics, not safety.
Who is the 'humanity' being protected in these negotiations? Often it is an abstract future humanity, unborn and undifferentiated. It is not the smallholder farmer in Malawi whose crop insurance is denied by an opaque algorithm. It is not the refugee in Bangladesh whose biometric data is mismatched. It is not the disabled person in Indonesia denied benefits by an automated decision. The existential frame lets leaders sound grave while avoiding reparations. It also lets them avoid technology transfer, data labor protections, and compute redistribution.
Who Gets a Seat at the Table?
The people most affected by AI's rollout are absent from the room. Data annotators in Venezuela. Content moderators in Kenya. E-waste pickers in Ghana. Lithium miners in Chile and Bolivia. Call center workers in the Philippines. They are not at the G7. They are not at the White House. They are not at OpenAI's board meetings. When Altman says 'global standards,' he means standards written by people like him. When Kratsios says 'American advantage,' he means the same thing with a flag.
A different agenda would start with the immediate. Mandatory bias audits for healthcare algorithms, with public results. Portable labor protections for data workers, including a global minimum wage. Environmental impact assessments for data centers, with binding water and energy caps. A compute tax that funds digital repair in the Global South. Technology transfer for local model building. These are not utopian. They are boring, concrete, and enforceable. They also threaten the current business model.
The US rejection of OpenAI and Anthropic is not a defeat for regulation. It is a signal that even mild global standards threaten national champions. The real question is not whether to regulate or compete. The real question is whose humanity counts. If the answer excludes the Global South, then both the pleas and the rejections are theater. The audience pays. The architects leave. The rope snaps.
