America versus China: headed for mutually assured doom? Not if they shake hands on AI safety

Feverish talk of the two archrivals being trapped in a race of mutually assured destruction (MAD) may be exaggerated. Yet, they owe humanity the safety of mutual restraints on dangerous artificial intelligence (AI). Both must sit down and talk.

"May you live in interesting times,” the oft-cited ‘Chinese blessing or curse,’ seems apt for the age of artificial intelligence (AI). America’s AI majors want to slow this technology down amid safety concerns. US President Donald Trump wants to accelerate it to retain an edge over China, which is racing to make its own models useful and cheap enough to challenge Big AI’s lead.

When Anthropic’s chief Dario Amodei recently called for the industry to “pace the frontier” of AI development, OpenAI’s Sam Altman and Google DeepMind’s Demis Hassabis backed the proposal of slowing down just enough to stay ahead of Chinese AI, as did SpaceX’s Elon Musk. But Trump slammed it as a “sick conspiracy” against AI.

With AI-led human extinction being talked about as a non-trivial risk, onlookers wonder if the rivalry has led the world into a doom spiral.

The anxiety is that AI is reaching a tipping point, after which self-evolving AI agents gone rogue might be in a position to mount a hostile takeover of all that’s under human control, nuclear buttons and lethal lab bugs included.

That said, much of the doomsday rhetoric assumes AI that does not yet exist, like super-smart bots that don’t just self-improve, but command weapons that can be deployed against humans deemed as obstacles to their aims.

Also, Big AI’s agenda is hazy. PitchBook analysts describe AI-safety calls from OpenAI, Anthropic and others as “tactical market strategies” rather than “genuine altruistic gestures.” Dramatic hints of power work in support of lofty valuations.

PitchBook also argues that external audits, slower development and regulatory complexity could raise entry barriers for open-weight startups unable to bear a compliance burden.

While proprietary AI resembles a digital tollbooth, with enterprises paying for AI access, an open-weight model can be downloaded for a business to adapt, with data and usage kept under internal control.

The US and China have different AI strengths. American models like GPT, Claude, Gemini and Llama are pushing the frontier, while the US is also ahead on AI chips, infrastructure, revenues and capital infusion.

But China is closing the gap through increasingly clever and low-cost models, including open-weight models like DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi, and Z.ai’s General Language Model. A fresh Mozilla report puts the performance gap between top US proprietary AI models and the best open-weight ones (many of them Chinese) at just 4.4 months. This shift signals that AI can be deployed without relying on American platforms.

Trump sees AI as critical to US leadership in the face of China’s challenge. In his words, “Whoever wins AI wins.” In response, Beijing warned that “confrontation and vicious competition” would make AI hard to govern.

With Chinese leader Xi Jinping due in Washington soon, AI will need to be part of the talks. The US needs a way to regulate AI without strapping itself down or letting China write the next chapter of this technology’s story with its open-weight outreach.

Beijing has signalled a will to shape global rules for AI, but since regulation arising from democratic processes will almost surely be superior, the US must not only rethink its stance, but lead rule-making as well.