What U.S.–China cooperation on AI could look like
Joe Biden and Xi Jinping agreed at last year's APEC summit in San Francisco to open a bilateral channel for AI consultation, but analysts say the U.S. and China can't even agree on what problems they're trying to solve.
Why it matters: As China and the U.S. compete for global AI dominance across civil and military applications, it's increasingly important for the countries to find common ground on AI safety.
- Washington and Beijing lack even a basic understanding of each other's approaches to AI.
Driving the news: The nations are considering holding bilateral telephone talks this spring.
- At a hearing of the U.S.-China Economic and Security Review Commission on Capitol Hill on Thursday, expert witnesses testified that the Chinese military places a high priority on advancing AI "across the kill chain," per Jacob Stokes, from the Center for a New American Security.
- Arati Prabhakar, director of the White House Office of Science and Technology Policy, told the Financial Times, "We have to try to work [with Beijing]" and that unspecified "steps have been taken."
- Asked at an Axios event in Davos how she saw the U.S. and China working together on AI regulations, Prabhakar said it would be "extremely constructive" to solve how to "robustly assess if an AI model is effective, safe and trustworthy."
Between the lines: AI regulation is either absent or very new in both countries, creating obstacles to efficient dialogue.
- The Cyberspace Administration of China is that nation's leading AI regulator, but there's no clear counterpart in the U.S.
Context: "Through dialogues you at least develop touch points, and learn who you need to call in an emergency," Matt Sheehan, a specialist in Chinese AI at Carnegie Endowment for International Peace, told Axios.
- Sheehan says it's likely that today neither side "fundamentally understands" the AI standards, testing and certification systems being developed by the other side.
What we're watching: Officials could start with a symbolic gesture, to signal they are serious about an ongoing dialogue.
- Top White House AI fears include rampant cyberattacks and misinformation, and accelerated chemical and biological weapons development by terrorists and rogue states.
- Sheehan suggests committing to keep nuclear weapons under human control: "It's not a huge sacrifice, but is meaningful symbolism," he says.
What they're saying: If officials can bag an initial win from AI safety talks, analysts say a second stage of the dialogue could expand into joint safety exercises.
- The challenge is to "begin building boundaries and common expectations around acceptable military uses of automation," write Graham Webster and Ryan Hass in a Brookings Institution paper published last month.
- "The U.S. should engage with China in a clear-eyed way on military AI risk," Stokes says.
- "We are not going to help Beijing red-team their censorship systems," Sheehan says, but the two sides could work together on standardizing pre-deployment testing of powerful AI.
- Managing AI that could be used in cyberattacks is another strong mutual interest: Both countries "may be deploying malware against each other, [but] they don't want random people on the street doing it," Sheehan says.
- The White House declined to comment.
Reality check: Big U.S. tech companies have historically struggled with their China relationships — Facebook was banned for years, Google withdrew, and Microsoft is still grappling with political issues at its AI lab in China.
- The particular way that generative AI relies on cultural values raises the stakes.
The bottom line: Establishing an effective AI dialogue with China will require officials to get specific and address a narrowed-down set of priority issues.
Meanwhile, Jensen Huang, CEO of leading AI chipmaker Nvidia, visited China last week to boost morale of local staff, after the Biden administration banned exports of its high-end chips in China, frustrating local customers.