US-China Military AI Hotline Proposed by Experts
US-China military AI hotline proposals would create a dedicated crisis channel for accidents, unauthorized operations and cyberattacks involving autonomous systems. American and Chinese experts also want explicit red lines keeping artificial intelligence away from nuclear-use decisions and attacks on nuclear command networks.
The recommendations were published by Brookings Institution on September 9, 2026, before planned government-level AI talks. They extend a 2024 commitment by U.S. and Chinese leaders to preserve human control over nuclear weapons, but neither government has publicly adopted the new safeguards.
The experts propose three areas for bilateral action:
- Ban autonomous attacks on nuclear command systems.
- Define meaningful human control over military AI.
- Create a hotline for urgent AI incidents.
US-China Military AI Hotline Would Report Incidents
The proposed hotline would build on existing military communication channels while focusing on emergencies caused by AI errors or AI-enabled cyber operations. Its purpose would be to give Washington and Beijing a way to explain unusual machine activity before either side treats it as an intentional state attack.
An automated defensive system could detect suspicious traffic, isolate infrastructure or launch a countermeasure within seconds. The other country might interpret that response as offensive action and retaliate before officials understand whether the original event was authorized, accidental or still under investigation.
Tianjiao Jiang of Fudan University argues that a dedicated channel could reduce this risk of cascading escalation. He participated in the continuing Track II dialogue convened by Brookings and Tsinghua University's Center for International Security and Strategy, which has examined AI and national-security risks since 2019.
A hotline would not guarantee communication during a crisis. Existing U.S.–China military channels have sometimes failed when political approval was slow or officials declined to answer. The proposal therefore depends on procedures that identify who can send an alert, what information it must contain and how quickly each government should respond.
Nuclear Command Systems Define the Red Lines
The Brookings paper recommends a clear prohibition on AI independently deciding to use nuclear weapons or launching attacks against nuclear command, control and communications systems. Those networks help leaders detect threats, assess warnings, issue orders and maintain contact with nuclear forces.
A disruption could be catastrophic even when no government intended an attack. Model errors, unauthorized actors or agents exceeding their instructions might interrupt communications or create false indications of a strike. Leaders could then face nuclear decisions under extreme uncertainty and severe time pressure.
Melanie Sisson, a Brookings senior fellow, extends the human-control principle to decisions about AI-enabled cyberattacks against strategically important civilian infrastructure. The proposed boundary covers sectors such as energy, finance and health care, where automated interference could cause casualties, economic disruption or military retaliation.
The recommendations build on the November 2024 understanding between then-U.S. President Joe Biden and Chinese President Xi Jinping that humans should retain control over nuclear-use decisions. The new paper argues that this principle needs operational safeguards as agentic systems become more capable and enter military workflows.
Meaningful Human Control Faces a Speed Problem
Both countries use language supporting human control, but the authors warn that the phrase can conceal different technical and institutional practices. A person might formally approve an action while relying on an automated recommendation delivered too quickly, or with too little context, for meaningful independent judgment.
The paper recommends a terminology working group or regular dialogue to align definitions. Common language would need to address which decisions must remain human, what information operators receive, whether they can interrupt an action and how systems record responsibility when multiple models and commands interact.
Speed is the central technical constraint. AI-enabled cyber operations can occur in milliseconds, while attribution may take days or remain uncertain. A defensive agent and an attacking system could exchange automated responses faster than commanders can verify whether the activity came from a government, a nonstate group or a compromised model.
Opacity creates another problem. Governments may not know an adversary's capability, level of autonomy or safeguards, encouraging worst-case assumptions. A defensive deployment can resemble an offensive one from outside, making verification and confidence-building measures important even when neither side wants to limit all military AI development.
Proposals Await Government Adoption
The recommendations are expert proposals, not a treaty, executive policy or negotiated bilateral agreement. Reuters reported that neither government has publicly endorsed them, although the dialogue is intended to inform official discussions.
The Brookings publication presents parallel American and Chinese perspectives rather than a single institutional blueprint. It identifies areas of overlap while acknowledging different strategic cultures, command structures and concerns about restrictions that could weaken technological competitiveness.
The immediate test is whether planned government talks convert the ideas into concrete procedures. Progress would require agreed red lines, definitions, notification thresholds, authorized contacts and exercises showing that a hotline can function during an incident rather than only during routine diplomacy.
Even limited agreement could matter because military AI failures may be misread as deliberate escalation. A reliable channel and human-only decisions for nuclear and strategic cyber operations would not resolve wider competition, but they could create time for verification when autonomous systems move faster than political leaders.
