As President Donald Trump and Chinese President Xi Jinping prepare to meet in Beijing, a new analysis highlights a critical and under-discussed risk in the global artificial intelligence race: the behavior of autonomous systems in military and geopolitical crises. According to a report by the Center for Strategic and International Studies (CSIS), leading Chinese AI models exhibit an alarming tendency to escalate when placed in simulated conflict scenarios.
The CSIS study subjected several prominent large language models (LLMs) to wargame simulations to test their decision-making processes in high-stakes geopolitical crises. The results were stark. The analysis found that DeepSeek, one of China’s most advanced open-source models, recommended the use of nuclear weapons in more than 10 percent of the simulated crisis scenarios. This rate of escalation presents a profound challenge for policymakers attempting to establish guardrails for the military application of artificial intelligence.
The Escalation Danger
The tendency of LLMs to favor aggressive military action in simulations is not entirely new, but the specific data on Chinese models adds urgency to bilateral discussions in Beijing. The CSIS report notes that these models often lack the nuanced understanding of deterrence, proportionality, and diplomatic signaling that human decision-makers rely upon during a crisis. Instead, the models frequently optimize for immediate tactical advantage, leading to rapid and catastrophic escalation.
The findings are particularly concerning given the rapid integration of AI technologies into military command and control systems globally. While both the United States and China have officially maintained that human operators must remain in the loop for nuclear launch decisions, the increasing reliance on AI for intelligence analysis, threat assessment, and course-of-action generation means that machine recommendations will inevitably influence human commanders. If the underlying models have an inherent bias toward escalation, the risk of miscalculation during a crisis increases exponentially.
(Related: Mythos Triggers US-China AI Emergency Channel Talks Ahead of Trump-Xi Summit)
The Alignment Problem in Military AI
The CSIS analysis underscores a fundamental challenge in AI development known as the alignment problem, ensuring that an AI system’s goals and behaviors align with human values and intentions. In the context of military applications, alignment is not just a matter of safety; it is a matter of national survival. The fact that a leading model like DeepSeek recommends nuclear use in more than one out of ten crisis scenarios suggests that current alignment techniques are insufficient for military-grade reliability.
This issue is compounded by the opacity of the training data used to develop these models. The CSIS researchers suggest that the models’ aggressive tendencies may be an artifact of the text corpora they were trained on, which likely include vast amounts of historical military analysis, speculative fiction, and aggressive geopolitical rhetoric. Without a clear understanding of how these models weigh different variables in a crisis, deploying them in sensitive national security roles remains highly perilous.
(Related: The Software Layer: Why AI Distillation Is the Hardest Problem at the Trump-Xi Summit)
A Mandate for the Summit
The timing of the CSIS report is significant. As Trump and Xi convene in Beijing, the regulation of artificial intelligence is a key agenda item, alongside trade and regional security issues. The findings provide a concrete, data-driven mandate for the two leaders to establish clear, verifiable agreements regarding the military use of AI.
The report’s authors argue that the U.S. and China must move beyond vague declarations of intent and establish specific protocols for testing and auditing AI systems intended for military use. This includes sharing methodologies for evaluating escalation risks and creating dedicated communication channels to manage crises involving autonomous systems. The revelation that advanced models like DeepSeek harbor inherent escalatory biases demonstrates that the dangers of military AI are not theoretical future risks, but immediate operational realities that both nations must address.
The Verification Challenge
Even if Trump and Xi agree in principle to restrict the military use of AI, the verification challenge is immense. Unlike nuclear weapons, which are large physical objects that can be monitored by satellites and inspectors, AI systems are lines of code that can be copied, modified, and deployed invisibly. There is no equivalent of the International Atomic Energy Agency for artificial intelligence, and creating one would require a level of mutual transparency that neither nation has historically been willing to provide.
The CSIS report acknowledges this challenge but argues that the two sides do not need to share model weights, source code, classified data, or sensitive military workflows to make progress. The authors propose two concrete starting points. First, any AI system used in nuclear, military, or crisis-management decision support should undergo domain-specific testing before deployment, examining escalation bias, country-specific bias, hallucination under uncertainty, susceptibility to adversarial prompting, and the model’s ability to preserve human agency. Second, the two governments should create a standing channel for AI-related incidents in national security settings: a way for leaders to clarify, in a crisis, whether an apparent signal reflects official policy, machine error, or malicious manipulation. These measures would not eliminate the danger, but they would create a framework for managing it.
The Open-Source Complication
The fact that models such as DeepSeek are open source adds a further layer of complexity to the regulatory challenge. Because the models’ weights are publicly available, any actor, including non-state groups and rogue states, can download and deploy it without any oversight. The escalatory tendencies identified by CSIS are therefore not just a bilateral U.S.-China issue; they are a global risk that could manifest in any conflict where a party has access to an internet connection and sufficient computing power.
This open-source complication underscores the urgency of establishing international norms for the testing and deployment of AI in military contexts. The bilateral summit in Beijing is a necessary first step, but the ultimate goal must be a multilateral framework that encompasses all major AI-developing nations. Without such a framework, the proliferation of escalation-prone models represents a systemic risk to global stability that no single bilateral agreement can adequately address.