Anthropic’s AI Slowdown Plan Doesn’t Say Who Else Decides

This month, Anthropic CEO Dario Amodei published an essay arguing that AI companies should slow down how fast their models get more capable, so safety work can catch up. His plan has three steps. Anthropic says it will let outside evaluators work inside the company with access close to what its own staff have. AI companies in democratic countries would agree on shared safety standards and speed limits. And the U.S. and its allies would try to reach agreements with China.

Amodei says society deserves a say in how AI gets used, and that more time for public debate would help. Then he names the players. American companies. Democratic governments. China. One proposed global step has “both sides” testing models before release, possibly through a global standards body. He doesn’t say who else would sit on it.

If the risks are global, the power to decide how much risk is acceptable should reach past the countries and companies building the biggest models. My training in public diplomacy taught me to watch the distance between being addressed and being heard. Growing up Black in America taught me to ask one thing about any promise of inclusion. What can the people being included actually change?

Roughly eight months of travel across Southeast Asia taught me I’m not always on the losing end of that question. In a Ho Chi Minh City nail salon, I tipped each of the three women working on me 200,000 dong, about $8, and watched their eyes widen. I didn’t grow up with money, and in that chair it didn’t matter. Geography made me privileged in ways that had nothing to do with my own history. Since then, I pay closer attention to who sets the terms and who has to accept them.

American AI already reaches people who had no say in how it works. In Hoi An, a man I was messaging asked whether I was “one of those woke Americans.” When I asked what he meant, he sent me a definition signed “Courtesy of ChatGPT.” He was 30 and had never left Vietnam. An American chatbot is where he went to learn what a word from Black American life means.

Emrys Schoemaker, who leads technology policy work at the United Nations University’s Centre for Policy Research, points out that many governments lack the experts, staff, and money to take part in talks they’re already invited to. His focus is the 2027 review of the Global Digital Compact, a 2024 UN agreement on how countries cooperate on digital issues. How widely the UN draws that review, he argues, will help decide who gets a voice. A government can’t push back on a technical proposal just because someone sent it the documents. Somebody has to be paid to read them, talk to the people affected, write alternatives, and keep showing up.

In July, the Group of 77 and China, a bloc of more than 130 developing countries, called for long-term help building up universities, research centers, and regulators, plus the money to pay for it. Its statement warned that AI rules can’t work if only the countries with the most technology shape them.

Some of that expertise already exists. Southeast Asian governments put out a guide in January 2025 on testing AI systems, reporting when they go wrong, and labeling AI-made content, with examples from Vietnam, Thailand, and Singapore. A global body that sets testing rules should have to answer to that work.

The same gap shows up with the people AI companies depend on. Sarah Gilbert, a Cornell researcher who moderates Reddit’s AskHistorians community, writes that Reddit sells its data to OpenAI and Google, while the volunteers who keep that data worth buying work unpaid. Those volunteers now also weed out AI-generated posts.

Taken together, these are people handed responsibility without the money or power to carry it out.

If Anthropic and the governments considering a slowdown want the extra time to count, they should pay for independent research in more languages, regulators who can keep up, and room for workers and communities to take part, with no strings that stop anyone from criticizing the funder. A country’s seat at the table also doesn’t guarantee its marginalized communities get heard. Whatever standards body comes out of Amodei’s plan should let outsiders propose rules, challenge releases, and see a public record of how their objections were answered.

Slowing down may be the right call. But extra time also helps whoever already writes the rules. Amodei lays out how the U.S., its allies, and China might coordinate. He doesn’t explain how countries outside the leading AI powers would help set the terms.