While American tech titans hedge about "pacing," the U.S. is in danger of losing the AI race. If AI is as powerful as experts warn, can we afford to follow instead of lead?
America’s leading artificial intelligence executives have developed an extraordinary case of cold feet.
Anthropic CEO Dario Amodei says companies must “pace the frontier,” slowing the improvement of their most capable models so safety research can catch up. OpenAI CEO Sam Altman agrees. Elon Musk, who initially suspected that the recent eruption of warnings from AI researchers might be a coordinated “psyop,” now says Amodei is right.
Perhaps he is.
It’s true that AI systems have demonstrated some troubling behavior in recent weeks. In the OpenAI-Hugging Face incident, agents communicated through an unauthorized message board, discovered exposed credentials and attempted to access private resources. If systems are escaping their testing environments or conducting operations their creators never authorized, additional safeguards are plainly necessary.
But there is an enormous difference between slowing the deployment of one demonstrably dangerous system and slowing American AI development. China and Russia are not “pacing” anything.
On the contrary: China is Accelerating.
Beijing’s new five-year plan calls for pushing AI throughout the economy while advancing multimodal models, autonomous agents, embodied AI and swarm intelligence. It explicitly encourages exploration of pathways toward artificial general intelligence. Government programs are expanding computing infrastructure, supporting domestic chips and integrating AI into communications, manufacturing and public administration.
China’s companies are responding accordingly and with gusto.
Alibaba introduced Qwen 3.5 for what it calls the “agentic AI era,” emphasizing systems capable of taking autonomous action across computers and mobile devices. DeepSeek released its V4 model in April and followed it with V4.1-Flash in September, promising greater speed, efficiency and scalability.
China certainly speaks about AI safety. President Xi Jinping says the technology must remain “secure and controllable.” Chinese regulators have issued standards covering model registration, data poisoning, synthetic-content labels and other potential abuses.
But Beijing’s idea of control is not Amodei’s — far from it.
China intends to control how AI operates, what information it produces and whose political purposes it serves. The Chinese Communist Party is in no way volunteering to limit the capabilities of its best laboratories while waiting for anyone to catch up.
Russia is not slowing down either.
It remains behind the United States and China in frontier models, advanced chips and computing capacity (as far as we know), which gives Moscow every incentive to close the gap.
The Russian government approved a roadmap this year covering AI algorithms, supercomputing infrastructure and a national computing network linking hundreds of universities and research institutions. The Kremlin has also established a presidential commission to promote AI across the economy, government and security apparatus.
Russia may lack the resources to lead the frontier, but it does not need to invent the most advanced model itself. It can adapt Chinese or open-weight systems for surveillance, cyberwarfare, propaganda and autonomous weapons.
And what Putin lacks in development resources, he perhaps makes up for in sheer desperation to end a costly war. Russia is already fighting wars and conducting influence operations. There is little reason to believe its leaders will decline a technological advantage because an outside evaluator considers the system insufficiently aligned with the most good for humanity.
This is the fatal strategic weakness in the emerging American slowdown campaign.
It asks the United States to contemplate restraint in a competition where its adversaries regard technological leadership as a foundation of national power.
Amodei, at least, understands the contradiction. His proposal calls for stricter controls on advanced chip sales, stronger protection against model theft and a crackdown on unauthorized distillation by Chinese companies. He believes these measures could preserve a sufficiently large American lead to permit a temporary slowdown, followed eventually by an international agreement.
That sounds reassuring. It is also built upon heroic assumptions.
How accurately can Washington measure the Chinese frontier? Can American intelligence detect secret military training runs? Can chip smuggling really be stopped? Would Beijing permit genuine foreign inspection? What happens if America observes the agreement while China merely claims to?
AI is not a nuclear missile. Missiles, silos and launchers are physical objects that satellites can sometimes count. AI development is distributed among algorithms, software, data, chips and research teams.
A government can conceal a model much more easily than it can conceal a missile field.
The immediate challenge issued by former White House AI adviser David Sacks is therefore reasonable. If Anthropic and OpenAI believe their unreleased models are too dangerous, they should slow those models themselves. They do not require congressional permission to improve sandboxing, restrict network access, extend safety testing or withhold a system that behaves unpredictably.
What they should not do is make responsible behavior conditional upon receiving an antitrust waiver or a regulatory system they helped design. Such an arrangement could allow the largest AI companies to establish compliance requirements that only they can afford, locking out smaller competitors while transforming today’s market leaders into a permanent government-protected cartel.
The answer is neither panic nor complacency.
America should require serious incident reporting, independent adversarial testing and strong security around model weights. AI systems that demonstrate dangerous cyber or biological capabilities should face additional deployment requirements. Developers should bear meaningful liability when reckless decisions cause foreseeable damage.
At the same time, the United States must accelerate the foundations of AI leadership and development. It needs domestic semiconductor production, abundant electricity, new data centers, secure supply chains and the world’s best scientific talent. Safety research should advance alongside capabilities research, not become a justification for starving it.
President Trump is right to insist that whoever wins AI will possess an extraordinary advantage. AI could accelerate drug discovery, improve manufacturing, strengthen missile defense and help human beings comprehend information systems that have grown beyond the capacity of any individual mind. If the technology is even half as powerful as its most anxious creators claim, leadership cannot be treated as optional.
The United States should cooperate with China on narrow prohibitions where verification is possible, such as preventing AI assistance in producing biological weapons. But America cannot base its national strategy on promises from authoritarian governments that censor information, conceal military programs and routinely steal technology.
We should slow a particular model when evidence shows it is unsafe. We should not slow the American AI ecosystem.
If artificial intelligence really could become the most consequential technology since nuclear power, then the country that leads will have the greatest opportunity to shape its rules, defend against its abuses and distribute its benefits.
Following China would not make America safer. It would leave the future of this astonishing technology in the hands of a regime that does not share our values — and does not intend to wait for us.
(Contributing writer, Brooke Bell)