Nuclear weapons and poison drugs: Washington and China tread cautiously over AI talks

Lu said both sides of the debate she was a part of were engaged, eager to define what autonomous weapons are and what it means for humans to be involved in the loop of weapons of mass destruction. He said it looks like he is working on it.

“I feel the energy. We’re all trying to come up with ideas about this,” she said. “A common threat to the United States and China these days is something that AI can launch, such as nuclear weapons, such as nuclear missiles. This is a very dangerous threshold, and both sides have an interest in preventing unintended consequences. Masu.”

How that translates is less clear.

AI deal shows China and US can cooperate on technology rules despite conflict: analyst

“Autonomous weapons is really less about AI and more about some level of autonomous decision-making, such as command and control of nuclear weapons,” said Paul Triolo, head of technology policy at consultancy Albright Stonebridge Group. ”.

“If anything…it seems like the purpose is to reassure both sides that the other doesn’t have a doomsday machine secretly in development. That it’s a bad idea to allow any kind of automation here. It is generally agreed that.”

Arati Prabhakar, director of the White House Office of Science and Technology Policy, suggested in January that U.S.-China talks would also address unknown aspects of AI safety.

Analysts added that while the broad principles were fine, ironing out the details in bilateral talks was a much more difficult task. The Chinese government would likely prefer to talk about the safety of the frontier model rather than any restrictions on AI in the broader defense sector.

This reflects a long-running dispute between the two countries in recent years over “guardrails” and military hotlines involving conventional forces. Despite Washington’s eagerness to create rules to prevent inadvertent contact between ships and aircraft in the disputed and increasingly congested South China Sea, Beijing has given the United States assurances that would strengthen its assertiveness. They are reluctant to give to the government.

Bonnie Glaser, managing director of the Washington-based German Marshall Fund, said she is not sure what level of senior officials the Chinese government will send to any talks, especially whether they will be military-to-military talks. He said he did not yet know, and said it was considered more meaningful than the talks. Default diplomatic channel.

“You often see progress in Track 2 or Track 1.5 discussions, but it ends up going all the way to Beijing and goes nowhere,” said Glaser, who frequently participates in sub-official talks. “But if Xi Jinping says we will work together, that will motivate people to get things done.”

US-China decoupling will have a negative impact on AI governance: Henry Kissinger

So far, there is no evidence that militaries around the world are using or planning to use Frontier AI models for military purposes, analysts said.

Separately, the world’s two largest economies, key to a meaningful global agreement, are also debating a framework for governing AI in commercial sectors, with former Secretary of State Henry Kissinger I raised this issue during my previous visit to Beijing. He is reportedly in close consultation with former Google CEO Eric Schmidt.

Track 2 talks have made very limited progress on the commercial front, with China insisting the best way to ensure security is for both sides to fully share technology and end export restrictions on key AI technologies. are doing.

“It’s not the beginning, it’s a poison pill for America,” Lu said.

This comes as the Washington government in late January ordered cloud service providers such as Microsoft and Amazon to proactively identify foreign customers developing AI applications on their platforms that appear to be targeting China. This comes in conjunction with the announcement of a proposed rule requiring an investigation.

US asks cloud companies to report foreign users building AI applications

Chinese Foreign Ministry Spokesman Wang Wenbin countered at a press conference in Beijing that the United States should cooperate on AI “instead of separating, breaking chains, and building fences.”

But the two sides also share common interests in commercial areas, such as the idea of ​​traceable watermarks on AI images and concerns over data // and trademarks.

One area for discussion is the control of frontier models that rely heavily on graphics processing units (GPUs). Graphics processing units (GPUs) are sometimes called the rare earths of AI because of their role in “teaching” computers.

But in December, U.S. Commerce Secretary Gina Raimondo for the first time directly linked U.S. GPU export controls to efforts to prevent Chinese companies from training frontier AI models, rather than past references to unspecified military uses. And analysts expect more restrictions to be imposed this election year due to pressure from Congress.

“This makes it difficult to discuss with the Chinese side the safety of frontier AI models, the subject of the Bletchley Park agreement, and depletes the limited goodwill reserves built up by recent high-level diplomatic engagements. It seems like it will further limit progress,” said Toriolo, who took part in the Track 2 talks.

With prospects for formal progress limited and any negotiations expected to be an uphill battle, an informal Track 2 channel involving academics, think tanks and industry groups is being used to release pressure, test proposals and provide communication channels. important for keeping open.

Among the participants in recent AI Track 2 discussions were representatives from the Carnegie Endowment for International Peace, George Washington University, the Brookings Institution, several other US think tanks, Tsinghua University, the Department of Science and Technology and Chinese think tanks. was also included. Ministry of Industry.

While the immediate priority for Chinese regulators is almost exclusively content generation, the U.S. side is focused on key national security issues such as cybersecurity and the potential for AI models to design chemical, biological, and nuclear weapons. is guessing.

China and the US agree on AI risks, but can they take past military technology competition into account?

“U.S. officials are also concerned about traditional things like bias and disinformation associated with AI platforms, whereas Chinese regulators are less concerned about these things,” Triolo said. , Raimondo added that the idea for U.S.-China AI talks may have first been hatched when he visited Beijing in July for the Xi-Biden summit.

“And now both sides are scrambling to understand what they are trying to talk about.”

Disparate views on transparency, decision-making and centralized authority are also influencing the negotiations. Until now, as new technologies have advanced, from cell phones and fax machines to the internet and virtual currencies, the Chinese government has slowly worked to study and control their use and ensure they do not pose a threat to the Communist Party.

The United States has a more decentralized system, often regulating its use after problems or abuses surface and allowing companies and individuals to explore and exploit its use.

The People’s Liberation Army has become much more cautious about its thinking.

“The PLA is a long way from doing something similar, so it’s important that the PLA gets an understanding and says this is our framework,” said Martin Lasser, managing director of Dutch-based open company Datenna. It would be helpful if there was some kind of return on this front.” -Source Intelligence Software Company. “What the United States is most interested in is getting transparency about how the Chinese side thinks about these issues.”

US, China, EU and UK leaders agree to cooperate on AI risks

On other fronts, China has a growing number of quasi-nongovernmental organizations and think tanks working with major AI companies on issues related to AI safety. There are also relatively few civil society organizations in Western countries who tend to participate in these discussions. “All of this complicates China’s participation in multilateral and bilateral dialogue on AI,” Triolo said.

The summit meeting between Mr. Xi and Mr. Biden was aimed at halting the rapid deterioration of bilateral relations and lowering the temperature. “There’s nothing wrong with understanding the safety of AI,” said Lasser, a former intelligence analyst. “I wonder if we can establish some kind of dialogue.” “There is so much distrust on both sides that it will be a very steep climb to reach any agreement.

“But at least if they’re having a conversation, it’s better to have a dialogue than not. All in all, it’s not a bad thing that they’re exploring possibilities.”