China-US AI talks Tuesday have absurdly low expectations

When key players from the US and China meet today in Switzerland to discuss AI, the stakes couldn’t be higher, and the expectations couldn’t be lower, observers say.

Officials from the White House, the US Department of State, and the US Department of Commerce will meet Chinese representatives in Geneva for the talks, which are aimed at exchanging views on understanding and addressing the risks of advanced AI systems.  

The US is seen as having a significant advantage in generative AI capabilities for the moment, but China’s massive resources could allow it to surpass the US within a few years. The US is attempting to extend its lead a little longer with attempts to cut deals with AI leaders in other regions, such as the Middle East. And China officials are hardly sitting still in this AI race.

CIOs typically worry about AI risks such as copyright, cross-border exchange of training data, privacy, and access to the best AI creations from both countries, but the participants in the talks are likely to have other things on their minds.

 “Although the talks are a good first step towards establishing potential guidelines for AI safety, the talks will most likely be focused on existential risks, instead of the more mundane, but significant risks, that IT leaders are facing on a daily basis,” said Andrew Gamino-Cheong, CTO at Trustible, an AI governance strategy firm. “Regardless of the outcome of the talks, we’re still likely to see a further divide in the AI space as generative AI models that meet China’s regulations for models would make them unsuitable to release in the US and vice-versa.”

That diplomatic paradox — that what works for one side is almost certain to be rejected by the other — is what many pointed to as a reason little if anything of practical AI value is likely to emerge from the talks.

Brian Levine, a managing partner with Ernst & Young who was one of the US Department of Justice’s representatives in the US law enforcement Joint Liaison Group (JLG) with China, was one of those who said that he didn’t expect anything to come from the talks. He served on the Intellectual Property Crime Committee.

“Although I’d like to be optimistic, when either country smiles and proclaims that ‘this’ aspect of AI is certainly an area where both countries can work together in harmony, the other country may smile, nod, and then cross it off the list of acceptable areas for collaboration.”

Another technology consultant who has strong ties to China is Michael Hasse, whose clients trade with China and who also owns a company with a branch office in Taipei. 

Implicit threat

Hasse argued that one of the reasons both countries might be able to find small slivers of potential cooperation is the implicit mutual cyberattack threat.

“All of the major nation-states — and some minor — have sleeper personnel and unleveraged cybersecurity penetrations in their rivals’ key corporations and government agencies,” Hasse said. “In a scenario where China, or a rogue element within a Chinese corporation, decides to activate one or more of their options in that realm, then affected CIOs could have data leaks and/or security breaches that are far deeper and harder to detect than they may have realized was even possible.”

This is potentially a possibility because the relationship between China’s government and Chinese companies are exponentially more susceptible to government influence than in the US.

“What I frequently see missing in these analyses is the simple fact that the Chinese government and enterprise are inextricably intertwined, to an extent unimaginable to most Americans. Combine that with being joined at the hip to American trade interests and you have an automatic damper on anything that could cause major waves,” Hasse said. “We’ve been at war over AI technology for a few years now. Companies in the AI arena are already very well aware of the corporate espionage target on their backs.”

For the moment, the US is in the lead with generative AI developments, but that may not last.

Language advantage

Geopolitical analyst Irina Tsukerman put the US about five years ahead of China today but added that China’s investments may change that. One of the complicating factors is that most generative AI global content today is in English, which puts US efforts at a major advantage. “This (English language advantage) limits China’s physical reach in some ways for the time being,” Tsukerman said. 

Tsukerman added that the nature of generative AI might make that lead easier to retain than other tech areas. For China, the US generative AI lead “is more difficult to surmount than the lead in semiconductors,” which China has rapidly erased. 

Another key concern, Tsukerman said, is whether violations of global laws involving copyright and related issues can realistically be enforced in China and with Chinese interests. That could prove to be a significant CIO concern if executives doubt that the US can deliver what China agrees to at a negotiating table.

“China’s propensity for intellectual property theft raises questions about its willingness and ability to carry through any potential commitments. There is also the view that China’s geopolitical agenda is nefarious and inherently anti-American,” Tsukerman said, adding that whatever the general agreements are regarding ethics, “China will likely push for the use of AI in ways ultimately disfavorable or damaging to the US or even US companies and businesses.”

Some suggested that the US is likely to push for agreements in less sensitive areas to try and get something agreed to. 

“Hopefully the US and China can identify areas to collaborate in AI, and not just compete, particularly where there might be mutually shared goals such as climate change, pandemic response and economic cooperation,” said Ronen Cohen, VP of product and strategy at Duality.

Material matter

A critical area for discussion is one that may be the least possible to address meaningfully: identifying the material that generative AI trains on. Even if both countries agree, which is highly unlikely, companies in both geographies are training with data that they gather, but also with data from a large number of third parties. 

Unless those companies took the time and effort to identify every source that those third parties used, it would be almost impossible to ever truly know all of the training material. That means that copyright, compliance, and privacy issues can likely never be truly resolved. 

On top of that, China has cracked down on companies’ generative AI accuracy in a way that the US has not even tried. 

For example, Gamino-Cheong said that he has seen Chinese regulators holding vendors responsible for any hallucinations. That has made some Chinese firms understandably skittish about generative AI development in anything other than highly limited implementations. “China has clamped down, imposed a lot of regulations, put out a lot of guardrails. It has slowed things down. To release a GenAI model in China, it must meet strict government standards. They are telling model makers that they are responsible for anything it outputs.”

Hasse agreed, but stressed that China may have little choice but to loosen its grip if it wants to effectively compete with the US in generative AI — which China absolutely wants.  “If relations continue to grow more strained, we could see situations where the PRC mandates internal technology development, which covers a much broader spectrum as AI does not operate in a vacuum, at a rate that is simply not possible with their present capabilities,” Hasse said. “This may pressure corporations, or individuals within corporations, to act independently of PRC oversight in a desperate bid to achieve goals.”