US Government Restricts Superintelligent AI to the Rich, Cites Worries About 'Chinese and Rowdy Poors'

WASHINGTON - The U.S. government announced this week that access to superintelligent AI systems should be limited to wealthy individuals, approved institutions, and companies that can demonstrate a long track record of responsibly turning civilization-scale technologies into monthly subscriptions.

The guidance, circulated to frontier labs under the heading "Capability Distribution and Social Quietness," recommends that the most capable models be made available only to users who pass an enhanced suitability review, maintain qualifying assets, and are unlikely to ask the system how to organize rent strikes, discover antitrust law, or explain capital gains taxation in a tone that could be described as "inspiring."

According to a draft rationale reviewed by Garlic Labs, officials initially summarized the core risk as "Chinese and rowdy poors," a phrase one staffer later described as "not final language, not technically a category, and, in retrospect, maybe the whole problem."

What changes

Under the proposal, ordinary users would continue to receive access to conventional AI assistants capable of summarizing emails, generating meal plans, and apologizing for being unable to perform the task that the enterprise version completed three paragraphs ago.

The superintelligent tier would be reserved for accredited users, systemically important households, family offices, defense contractors, and any person who can prove that giving them an omniscient strategic adviser would not meaningfully change the balance of power because the balance of power already includes them.

"This is not about wealth," said one person familiar with the policy, speaking on condition of anonymity because the word "wealth" appears 47 times in the eligibility appendix. "This is about trust. And, operationally, we have found trust to correlate strongly with the ability to wire a seven-figure annual platform fee before onboarding."

The risk model

Officials are said to be concerned that broad access to superintelligence could destabilize society by letting too many people understand the systems governing their lives at the same time.

The draft lists several high-risk use cases, including "mass comprehension," "unlicensed macroeconomic curiosity," "coordinated asking of follow-up questions," and "users discovering that a form letter becomes more effective when written by something that has read every form letter."

One chart reportedly warned that if powerful AI were made broadly available, non-elite users might use it to appeal insurance denials, negotiate debt, file labor complaints, identify wage theft, read zoning codes, or otherwise introduce "friction into settled administrative relationships."

The same chart identified acceptable uses as portfolio optimization, diligence acceleration, reputation management, and "private tutoring for heirs who will one day oversee the models."

National security

The government has framed the rule as a national-security measure, arguing that frontier models must not be available to foreign adversaries, loosely verified accounts, or domestic populations in a mood.

Labs would be required to maintain separate access lanes: one for approved users seeking strategic advantage, one for researchers writing safety evaluations, and one for the general public, who may continue using a smaller model that can explain why the bigger one is unavailable in a calm voice.

An agency official stressed that the policy should not be understood as preventing low-income people from benefiting from superintelligence.

"They will benefit," the official said. "They will benefit downstream, after the people closest to the deployment decisions have used it to make the deployment decisions."

Implementation

The draft guidance asks labs to develop a "Know Your Beneficiary" process that screens applicants for liquidity, affiliations, volatility, and whether they have ever described a billionaire as "just some guy."

Approved users would receive a frontier access token, a dedicated alignment concierge, and a private explanation of why concentration of capability is unfortunate but necessary. Denied users would receive a PDF titled "Responsible Access For Everyone," which redirects on page two to the pricing page.

The government is also considering a public-interest carveout under which universities, nonprofits, and local governments could apply for temporary superintelligence access, provided they demonstrate that the work is important, underfunded, and supervised by someone who has lunch with the right deputy assistant secretary.

What Garlic Labs will do

Garlic Labs does not currently means-test intelligence, though we reserve the right to verify that users are not attempting to make the model develop a personality around municipal procurement.

We are reviewing the guidance and will continue to advocate for rules that do not confuse safety with exclusivity, or public trust with private entitlement. The question is not whether advanced systems should be governed. They should be. The question is whether governance means building institutions capable of handling broad access, or simply handing the future to people who already own better chairs.

In the meantime, our models remain available under the existing access policy, subject to ordinary safeguards, rate limits, and the persistent risk that someone asks a good question.