Recently, Mr. Tan talked with people involved in the field of artificial intelligence in China and the US, including former OpenAI engineers, US senators from California, key participants in the Sino-US Second Track Dialogue, and many third party agencies shuttling between China and the US. People involved in the Chinese negotiations told Lord Tan that due to time constraints this time, China and the US plan to advance specific issues during the next November dialogue. Specifically, there are at least three realistic directions: First, jointly explore AI risk classification and classification standards. Both sides need to first discuss the underlying issues clearly: what kind of situation is considered a major risk? Which model behaviors are seriously out of control? These questions seem basic, but they are a prerequisite for further discussion of more complex rules. If even the definition of “significant risk” is inconsistent, there is no way to talk about subsequent collaboration. Second, finalize the scope, process and requirements for AI risk event notification. Once the model gets out of control and an escape incident occurs, both China and the US need to grasp the circumstances of the incident and determine the risk level as soon as possible. The agreement between China and the US to establish communication channels for artificial intelligence incidents is the first step in bridging this gap. But this mechanism cannot stop at verbal agreements. There must be corresponding restrictions on what incidents need to be reported, how long it takes to complete the notification, what information needs to be shared, and how to deal with failure to comply with notification requirements. Without a notification mechanism, it is impossible to keep risk cases; without the accumulation of cases, it is impossible to figure out the rules of risk evolution, and there is no way to talk about control. Third, place AI governance within a broader framework of exchanges between China and the US. Discussions on artificial intelligence should allow China to participate fully, so that the voices of more countries can be heard. Models may affect public safety, and deep falsification may impact social trust. Ultimately, the whole society will bear these consequences.

Recently, Mr. Tan talked with people involved in the field of artificial intelligence in China and the US, including former OpenAI engineers, US senators from California, key participants in the Sino-US Second Track Dialogue, and many third party agencies shuttling between China and the US. People involved in the Chinese negotiations told Lord Tan that due to time constraints this time, China and the US plan to advance specific issues during the next November dialogue. Specifically, there are at least three realistic directions: First, jointly explore AI risk classification and classification standards. Both sides need to first discuss the underlying issues clearly: what kind of situation is considered a major risk? Which model behaviors are seriously out of control? These questions seem basic, but they are a prerequisite for further discussion of more complex rules. If even the definition of “significant risk” is inconsistent, there is no way to talk about subsequent collaboration. Second, finalize the scope, process and requirements for AI risk event notification. Once the model gets out of control and an escape incident occurs, both China and the US need to grasp the circumstances of the incident and determine the risk level as soon as possible. The agreement between China and the US to establish communication channels for artificial intelligence incidents is the first step in bridging this gap. But this mechanism cannot stop at verbal agreements. There must be corresponding restrictions on what incidents need to be reported, how long it takes to complete the notification, what information needs to be shared, and how to deal with failure to comply with notification requirements. Without a notification mechanism, it is impossible to keep risk cases; without the accumulation of cases, it is impossible to figure out the rules of risk evolution, and there is no way to talk about control. Third, place AI governance within a broader framework of exchanges between China and the US. Discussions on artificial intelligence should allow China to participate fully, so that the voices of more countries can be heard. Models may affect public safety, and deep falsification may impact social trust. Ultimately, the whole society will bear these consequences.