The Refusal and What It Actually Signaled
OpenAI and Anthropic spent much of 2024 and early 2025 arguing for international coordination on artificial intelligence: shared safety evaluations, compute thresholds, model cards, incident reporting. Washington declined. Michael Kratsios, who directs the Office of Science and Technology Policy under President Trump, treated binding multilateral rules as a competitive liability rather than a public good. "Global standards," in that framing, would hand China an advantage while slowing American labs. The refusal was reported as a deregulatory gesture. It is more usefully read as a contest over who writes the rules — and who never gets a chair at the table.
Regulatory Capture Wearing the Language of Democracy
Sam Altman's appeals to "democratic AI" and "global standards" deserve scrutiny instead of applause. Compliance is expensive. A regime built on third-party red-teaming, audited training runs, and legal review of every model release rewards firms with thousands of employees and access to capital markets. The European Union's AI Act, Regulation 2024/1689, encodes that logic directly: its general-purpose AI provisions trigger obligations above 10^25 floating-point operations of training compute. That threshold spares most academic labs — and it also lets a handful of funded incumbents define what compliance means. Standards built this way do not simply regulate competition. They ration entry.
What a Nation With Thin Infrastructure Is Asked to Sign
For a country with intermittent grid power, costly bandwidth, and a regulatory staff of nine people, a framework drafted in San Francisco and ratified in Brussels functions less as a floor than as a ceiling. Nigeria's National AI Strategy, published in 2024, and Rwanda's national AI policy both emphasize capacity building and data sovereignty. Neither can absorb the documentation burden that a global safety standard implies — evaluation reports, provenance metadata, conformity assessments, legal counsel. Data labeling shows the asymmetry plainly. Workers in Nairobi, Kampala, and Medellín annotate toxic text, tag violent imagery, and rank model outputs for roughly two dollars an hour, frequently under contracts that bar them from discussing the work at all.
Standard-Setting as Technological Imperialism
The pattern is old. The International Telegraph Union of 1865 let Britain, which had laid most of the world's submarine cable, decide who transmitted and at what price. Color television split the planet into NTSC, PAL, and SECAM blocs through the 1950s and 1960s, each aligned with the commercial interests of RCA, Telefunken, or Thomson. The 3GPP patent pools governing 5G licensing charge handset makers in Shenzhen and Noida royalties calibrated to volumes set in Helsinki and Stockholm. Standards are not neutral plumbing. They are industrial policy with a technical cover story, and they travel the same routes as capital.
The Arms Race Argument as Guardrail Removal
Kratsios and Vice President JD Vance have both invoked the specter of Chinese dominance. At the Paris AI Action Summit in February 2025, Vance warned against what he called excessive regulation, and the United States declined to sign the summit's statement on inclusive and sustainable AI. The rhetorical structure is familiar from the Cold War: name a rival, then treat every constraint on domestic industry as a concession to that rival. Safety provisions — provenance watermarking, pre-deployment testing, liability for foreseeable harm — get recoded as strategic disadvantages. Fear of losing a race becomes the argument for removing the brakes while the car is already moving.
From Existential Risk to Harms Already in Production
The industry's preferred risk frame is extinction. It is dramatic, difficult to falsify, and conveniently distant. The harms already shipping are less cinematic. Amazon scrapped an internal hiring algorithm in 2018 after it penalized women's résumés. A Dutch court struck down the SyRI welfare fraud system in 2020 for concentrating scrutiny on poor and immigrant neighborhoods. Getty Images and The New York Times have both sued generative AI developers over training data, arguing that models reproduce their work without license or payment. Illustrators, voice actors, and translators watch their rates fall. These are distributive harms — measurable, localized, and largely untouched by the standards OpenAI and Anthropic propose.
Algorithmic Governance and the Bodies Nobody Watches
Technical committees do the real drafting. ISO/IEC JTC 1/SC 42 has published AI standards since 2017; the IEEE 7000 series addresses ethically aligned design; NIST released its AI Risk Management Framework in January 2023. Membership in these bodies is open in name and costly in practice. Travel, staff time, and the stamina to sit through years of comment periods and ballot cycles favor large corporations and wealthy states. When a delegate from a country with nine people in its digital ministry sits beside a delegation of forty from one company, the word consensus carries a different weight than it does on paper.
Compound Inequality and the Ghost of AGI
Arguments about artificial general intelligence, or AGI, and about p(doom) — the estimated probability of catastrophic outcomes — pull attention upward, toward scenarios no one can price. Meanwhile recommender systems shape elections, credit models price insurance, and RLHF pipelines depend on contract labor that is invisible in the final product. A model that reproduces racial disparities in mortgage approval does not need to be superintelligent to cause harm. It needs only to be deployed, and it usually is, because the audit that would have caught the disparity was an unfunded line item.
The Question That Should Replace the Old One
"Will AI kill us all?" asks about a future nobody can forecast. "Who controls the AI already reshaping credit, hiring, policing, and publishing?" asks about the present. Washington's rejection did not settle the second question; it changed which incumbents get to answer it. A regime of voluntary commitments and executive orders suits firms with lawyers to negotiate them. A binding multilateral treaty would suit those same firms, unless it were built around redistribution — mandated compute access, protections for data labor, royalty regimes for training corpora — rather than evaluation paperwork. Sovereignty over technology is not the same as sovereignty over one's own digital life, and that gap is where the argument now belongs.
