After years of observing U.S.-China dialogues on artificial intelligence, one pattern is hard to miss. The agenda has actually expanded over the past two years. From military AI to frontier model risks, from biosecurity to cybersecurity, the topics under discussion are not few. Disinformation, however, has barely entered the conversation, let alone become a subject of cooperation. In existing international discussions, it usually surfaces as accusation. I want to ask whether it can move from being a topic of accusation to being a problem the two countries handle together.
Today’s disinformation has a feature that is easy to overlook. Its damage often arrives before the truth is established. An unverified video, a screenshot of unknown origin, can run its full life in public discourse within a few hours. By the time the facts are pinned down, diplomatic protests have already been issued, public anger has flared, and platforms have already acted.
This kind of pressure is not new to journalism. Wait a few hours to verify, and the news cycle has moved on. Do not wait, and you risk circulating something false. But the dilemma is now spilling into diplomatic responses and policy decisions.
The current high level of mistrust between the U.S. and China amplifies this “cannot wait” condition. A feature of low-trust environments is that anything still unclear gets sorted into the worst-case interpretation by default. An unverified video can be read, within hours, as an organized operation directed by the other government. AI does more than spread rumors faster. It collapses the time between suspicion and conviction.
This does not mean cooperation is hopeless. It can begin from a relatively modest shared recognition. If synthetic content can shape a crisis narrative within hours, either side can become a victim of that dynamic regardless of intent. The first thing the two countries need to address is stability, not truth itself.
The Cold War hotline between Washington and Moscow offers a useful reference. During the Cuban Missile Crisis, diplomatic cables between the two capitals routinely took several hours to transmit. One critical message from Khrushchev to Kennedy took close to twelve hours to receive and decode. Both sides regarded the delay as a serious risk. In June 1963, after the crisis had passed, the two governments signed a Memorandum of Understanding establishing a direct communications link. The hotline was first formally used during the 1967 Six-Day War, and again during the 1971 Indo-Pakistani War and the 1973 Yom Kippur War. It resolved no fundamental Cold War dispute. But on more than one occasion, it gave leaders on both sides a few hours of breathing room, and kept them from making decisions that could not be taken back. That same logic of buying time for verification is what cooperation on AI disinformation could borrow today.
Beyond buying time, two other often overlooked issues belong in the same cooperative space. One is the speed gap between generation and verification. Producing a video that looks real now takes minutes. Professional forensic verification takes days or weeks. In a moment of public anger, asking a national leader to tell the public “let us wait a little longer” is politically close to impossible. Improving public media literacy will not solve this. The problem does not originate with the public.
The other issue is less visible. Future influence operations may aim less at human audiences and more at machines. As people around the world increasingly turn to large language models and AI assistants to understand U.S.-China relations, the Taiwan Strait, or trade frictions, the low-quality content circulating online today may end up appearing as “background material” in those models’ answers years from now.
The conventional response to disinformation has been takedown and labeling. But generation will keep accelerating as models improve, and removal cannot keep up. A more useful approach may be to build basic public information infrastructure. That means making a country’s authoritative records and primary documents available in machine-readable form on the open internet, where human fact checkers and machines alike can find them.
What can actually be done need not be grand. Three modest steps come to my mind. First, within the existing bilateral crisis communication mechanisms, add a specific arrangement for time-sensitive information events, so that during major incidents involving AI-generated content the two sides have a few hours to consult and verify before issuing a political response. Second, push for interoperability between the two countries’ technical standards on content provenance, so that the source label attached to a piece of generated imagery can still be read by the other side after it crosses the border. With that in place, no matter which side’s AI tool produced a given video, the other side could trace where it came from and whether it had been altered. Third, each country should expand the volume of authoritative material it makes available in the global public information space, so that when mainstream LLMs are asked about key U.S.-China issues, they have credible primary sources to draw on. None of these three steps requires either side to set aside existing disagreements. They share one underlying logic. When the information environment lacks basic verification capacity, the first party to lose response time is the government itself.
Structural disagreements between the U.S. and China will not disappear, and competition in AI is likely to last. Yet even at the most strained moments, the two governments still share at least one interest. Neither has any reason to let the global public information environment become a space where no one has the time to form judgments or correct mistakes.
