September AI security dialogue: shallow consensus is clear, hard problems remain

US and Chinese officials, according to Bessent's remarks to US media, favor an AI dialogue and a notification channel for shared risks and national security incidents. The proposed talks would exclude export controls on advanced-process AI chips and semiconductor equipment.

US-China talks flag an AI dialogue, with chip controls left out

Artificial intelligence was a core topic in recent Sino-US economic and trade talks. According to comments by Bessent to US media, both sides favor establishing an AI dialogue mechanism and building a notification channel for shared risks and national security incidents. Bessent stated that the dialogue would not address export controls on advanced-process AI chips or semiconductor equipment.

A 22 September 2026 commentary reviewed those positions and listed six observations. The piece said it was for academic discussion and market research, that all information came from public channels, that the judgments did not represent any company, and that they did not constitute investment advice.

Shared ground on cyber attacks, arms control, and CBRN

The commentary said the domains that both the United States and China treat as belonging to an AI safety dialogue are clear. They include at least cyber attacks, arms control covering chemical, biological, radiological, and nuclear (CBRN) issues, and biological and chemical risks. Those concerns have been central to US safety governance since 2024 and formed a core part of Biden Executive Order 14110. The commentary noted that although the Trump administration revoked that executive order, the underlying safety narrative did not change.

Disinformation, distillation, and open source

The commentary treated disinformation as a sub-issue of AI safety in Democratic Party discourse. It argued that under a two-party system the topic is contested inside the United States, and that unless an incident on the scale of Cambridge Analytica occurs, disinformation is mainly a matter of domestic governance. It said the need and feasibility of placing it in a state-to-state security dialogue are both weak.

Model distillation, the commentary said, is more a matter of industrial practice and commercial competition, with large disagreements, and should not be a topic of interstate security dialogue.

The commentary said the timing is not ripe for open source to enter the security dialogue. Safety reports from frontier laboratories and official bodies, it said, focus on frontier models for both autonomy-related safety concerns and human misuse. That focus, it argued, indicates that open-source models do not yet produce a material effect in cyber attacks, arms control, or CBRN.

The same piece said open-source risk is not absent. As frontier model capability improves, open-source model capability will rise, and open-source misuse is likely to become a topic. It argued that such issues would more often be handled through existing rules, for example open source combined with social fraud or financial fraud, and that the work needed is improvement of traditional sector regulation rather than a bundled open-source security agenda used to restrict open-source development.

Chip export controls as the disputed item

The commentary identified chip export controls as the disputed issue. Under the logic of the US Bureau of Industry and Security (BIS), the core trigger for controls since 2023 has been military and national security. On that basis, the piece said, preventing unordered diffusion of high-end chips and ensuring orderly, beneficial use should be part of a bilateral safety dialogue. It pointed to licensed H200 exports to China as a possible approach, for example through agreed limits on end users and strict process verification.

It also argued that if both sides make real progress on AI safety, that would amount to recognizing that the chance of severe AI-triggered security problems had been reduced, and that the need for continued strict controls would then decline.

To produce staged, practical results, leaving controls out of the safety dialogue mechanism is understandable, the commentary said. From industrial demand and legal logic, it added, a safety dialogue without a controls agenda is incomplete.

Human misuse versus autonomy risk

The commentary said the current AI safety dialogue still focuses mainly on human misuse of AI. It noted a divergence: when existential risk is emphasized, the stress is on destructive risk from AI autonomous evolution; when safety dialogue is the frame, the stress is on the human factor.

If the larger risk comes more from recursive AI improvement, the piece argued, safety cooperation should not stop at people or at after-the-fact information notification, and should move earlier into application and even research settings. Citing some practices associated with Fabel, it said even commercial activity in that area faces tight limits because of distrust, let alone state-to-state security cooperation, and that this gap reduces the substance of a safety dialogue.