The UN is the Wrong Place to Regulate AI

Hirad Marami

Issue Areas

In the aftermath of World War II, Soviet delegates at the UN pushed to destroy all nuclear arsenals through the 1946 Gromyko Plan. Only the U.S. had nuclear weapons at the time. In 1949, the Soviet Union conducted its first nuclear bomb test; it had been developing weapons all along. Free countries cannot expect authoritarian states to honor their promises. That lesson applies today: international AI regulation would be detrimental to American interests and would aid our adversaries.

​OpenAI CEO Sam Altman has repeatedly advocated for UN action on AI. A few months ago, he called for a US-led “international forum” to enforce “safety testing and standards,” serving as a “governance mechanism” modeled on the International Atomic Energy Agency (IAEA). Xi Jinping, China’s president, echoed Altman’s appeal for global AI cooperation after meeting President Trump in Washington. Those arguments have reached Congress: late last week, Senator Maria Cantwell unveiled an AI governance framework that emphasized the need for “global standards.”

The U.S. would be foolish to trust its authoritarian adversaries to respect any international AI accord. A key pillar of any testing framework is transparency: whistleblowers should be able to alert the relevant international body if their country is cheating. But dictatorships routinely silence anyone who dares challenge the ruling party’s narrative. Who can forget that Chinese police arrested the doctors who first reported Covid-19 cases in 2019, merely for speaking out?

Advocates for international AI regulation argue the world has no choice but to jointly regulate the technology: failing to do so risks an “Industrial Revolution of upheaval and disarray,” in Altman’s words. But that argument misses the central point: AI is a fundamentally neutral technology that underpins a wide range of other products and applications; it is akin to electricity rather than to nuclear weapons. A commitment to international oversight can have binding force, but only if each side holds genuine leverage over the other.

For example, once both the U.S. and Soviet governments obtained nukes, the doctrine of mutually assured destruction (MAD) incentivized jointly scaling back armament. With AI, there is no threat of MAD: an international AI pact entails trusting our adversaries to abide by rules they have no incentive to follow. Worse still, they have every reason to subvert the agreement as long as the U.S. retains its current lead in the field.

Existing international oversight mechanisms, such as International Atomic Energy Agency (IAEA) inspections, are already susceptible to cheating by authoritarian countries. Detecting deception is even harder with AI because no physical inspection is possible. Moreover, our rivals have already shown duplicity regarding AI’s allegedly “existential” risks. China warns of AI-induced societal collapse on the global stage and spreads that narrative in the U.S. through social media. At home, by contrast, Chinese officials reject negative press about the technology as they race to develop their own AI industry.

Besides the trust concern, an international AI accord culminating in an AI “governance mechanism” would mean entrusting adversaries with regulatory oversight of our technology. Rival countries could use their votes to push overly stringent testing protocols to hamper the U.S. lead. And our rivals have a history of using international forums to counter our interests. It was the Soviet Union that led efforts to criminalize “hate speech” at the UN, casting itself as an opponent of “Western Imperialism” while holding countless countries across Eurasia in servitude and destitution.

For all of Altman’s talk about ensuring “democratic” AI governance through the UN, the simple math of international relations stacks the deck against the Western world. The UN once picked Iran to chair a session of its human rights council. Do we really want the Chinese or Russians chairing a UN AI Council? And even if the body is restricted to Western democracies, what would stop the Europeans from imposing ruinous regulations like the EU AI Act on the United States? In the latter case, we would risk cementing an AI duopoly in the United States backed by international regulatory capture.  

​There is a genuine case to be made for strengthening international cooperation on AI in certain areas. For example, it would be worthwhile to boost the sharing of medical knowledge, given AI radically improves our ability to research new cures. Western countries could also build on existing intelligence alliances, such as the Five Eyes, to coordinate against AI-driven cyberattacks by adversaries.

And if AI companies are genuinely worried about the unforeseen consequences of their products, they have the right to pace their own firms’ development as they see fit. But none of that amounts to international regulation of AI itself, which would bury U.S. leadership while letting our rivals play catch-up. Gromyko would be proud.