The Refusal in Washington
US officials have rejected pleas from OpenAI and Anthropic for global AI standards, according to reporting on the dispute. The rejection is presented as a choice between American advantage and international rules. Michael Kratsios and figures around Donald Trump argue that binding standards would hamstring US laboratories while China scales its own models. OpenAI and Anthropic counter that common rules are needed for safety. Both sides accept a premise that rarely gets named: artificial intelligence is a tool of statecraft and hegemony, not a neutral public good.
Two Faces of the Same Hegemony
The disagreement between Washington and Silicon Valley is about method, not metaphysics. US policymakers want to preserve compute advantage, export controls, and chip supply chains. OpenAI CEO Sam Altman wants a global regulator that can license and audit frontier models. Anthropic wants safety commitments. Yet each position treats AI as an instrument for national or corporate power. The Global South appears as a market, a data source, or a security threat. It rarely appears as a co-author of the rules.
The National Security Trap
The logic of US rejection of global AI standards is often explained as geopolitical realism. China's AI ambitions, export controls on advanced semiconductors, and military applications make Washington wary of any external authority. That realism is not wrong about power, but it hides a domestic choice. National security framing converts public infrastructure and public data into classified assets. It also weakens antitrust, privacy, and labor enforcement at home. The same agencies that reject international standards may also resist local democratic oversight. Security becomes a shield for corporate concentration.
The Representation Gap
The global AI governance debate is a closed dialogue among San Francisco labs, US agencies, and Chinese leadership. Missing are the Global South, indigenous communities, and labor movements in developing nations. For these groups, global standards can look like digital colonialism. Historical parallels include the enclosure of common lands and the Bretton Woods institutions, where poorer states had little voice. When standards are written by incumbents, they can lock in access to data, GPU clusters, and model weights. Data sovereignty becomes a slogan only for rich states.
Regulatory Capture by Another Name
Sam Altman's claim that AI should be democratic deserves scrutiny. His call for common standards may be regulatory capture. Compliance regimes with audits, licenses, and FLOP thresholds cost millions. Large labs can absorb those costs; open-source projects, community labs, and public-interest researchers cannot. The result is a moat around incumbents. Regulatory capture is not a conspiracy theory; it is a predictable outcome of rulemaking by the regulated. Abbreviations such as API and GPU describe the infrastructure that smaller actors already struggle to afford.
Existential Risk as Moral Shelter
Anthropic's safety focus on existential risk can distract from immediate harms. Algorithmic bias in hiring, credit, policing, and welfare is already documented. Mass surveillance spreads through facial recognition and predictive analytics. Low-wage data laborers in Kenya, the Philippines, and Venezuela label toxic content for RLHF pipelines, often for a few dollars a day. A focus on humanity-ending scenarios lets labs speak in universal terms while ignoring the people already harmed. The question is not whether superintelligence is risky. It is whose suffering counts as evidence.
AI Sovereignty and Human Rights
Framing the issue as an arms race between the US and China narrows the political imagination. AI sovereignty offers a better frame. States and peoples should control their data, infrastructure, and models. Indigenous data sovereignty, expressed in CARE principles and OCAP, demands that communities govern how their knowledge is used. Labor movements demand algorithmic accountability, living wages, and the right to organize against automated management. Human rights law already provides tools: impact assessments, transparency, and remedies. These are not barriers to innovation; they are barriers to impunity.
What Would a Real Global Standard Look Like?
A real global standard would not be a treaty among powerful states and labs. It would include affected communities at the table with voting power. It would set binding labor protections for data work. It would require transparency about training data, model behavior, and compute use. It would fund public compute utilities and open-source ecosystems. It would ban biometric surveillance in public spaces. It would treat standards as a floor, not a ceiling that entrenches monopolies. Without these elements, global standards are a brand, not a safeguard.
From Consultation to Co-Authorship
Token consultations, listening sessions, and advisory councils do not fix the representation gap. Co-authorship means shared authority over budgets, datasets, and deployment. It means community consent before data extraction, not after harm. It means technology transfer without debt traps. It means reparations for the low-wage labor that made large language models possible. Abbreviations like LLM and RLHF name technical systems, but they also name supply chains that cross borders. A global standard that ignores those chains is not neutral. It is an agreement among beneficiaries.
Who Is Humanity?
The humanity invoked by AI leaders is narrow. It includes consumers in rich countries, shareholders, and security elites. It excludes refugees, informal workers, indigenous nations, disabled people, and the stateless. If AI standards protect that narrow humanity, they are not global. They are provincial rules with universal rhetoric. The US rejection of OpenAI and Anthropic is not the end of the story. It is a sign that even the powerful cannot agree on who gets to write the future. The rest of us should refuse to be written out.
