Artificial intelligence combined with space-based satellite imagery is giving the United States and China unprecedented ability to detect and track nuclear forces, but researchers warn this "precarious transparency" could backfire during crises if the AI systems misidentify targets or if satellites are disabled in conflict. A new analysis from UC Berkeley's Risk and Security Lab and the Council on Strategic Risks examines how computer vision and satellite surveillance are reshaping the strategic balance between the world's two leading nuclear powers.

The stakes became clear in early 2026 when the United States used the Maven Smart System, a Palantir Technologies platform that integrates intelligence, surveillance, and reconnaissance data with artificial intelligence, to target Iranian forces and nuclear sites. The system reportedly provides a tenfold increase in targeting capacity compared to previous methods. However, the same system mistook a girls' school in Minab for a military base, highlighting the risks of relying on AI-enhanced satellite imagery for critical decisions.

What Makes AI-Enhanced Satellite Surveillance Different?

Computer vision, the technology that allows machines to interpret and analyze images, has become central to how military and intelligence agencies process the flood of data from orbiting satellites. Rather than having human analysts examine thousands of satellite photos, AI systems can now automatically detect objects, classify them as missiles or ships or aircraft, and track their movements over time. Vision-language models, a more advanced form of AI that combines image recognition with language understanding, can even analyze patterns and make inferences about military activity.

The appeal is obvious: better surveillance of nuclear forces could theoretically improve stability by making surprise attacks less likely and by helping verify arms control agreements without intrusive on-site inspections. But the researchers identified four critical vulnerabilities that make this transparency precarious and uneven.

Why AI Satellite Systems Are More Fragile Than They Appear?

The first vulnerability is that AI models themselves are fragile. They can fail in unexpected contexts or when faced with countermeasures. Vision-language models, which are needed for higher-level analysis beyond simple object classification, are particularly vulnerable to hallucinations, where the AI generates false information, data poisoning attacks, and misalignment with intended goals. The Minab school misidentification is a real-world example of how these systems can fail catastrophically.

Second, satellite sensing has hard physical limits. Satellites cannot continuously watch the same location; they can only revisit areas periodically. Some types of sensors cannot see through clouds. Even constellations of hundreds of satellites cannot overcome these constraints, making persistent tracking of mobile missiles extremely difficult.

Third, during actual conflicts, adversaries can disable or destroy satellites using counterspace weapons, eliminating the surveillance advantage entirely. This means AI-enhanced ISR, or intelligence, surveillance, and reconnaissance, works best during peacetime for verification and monitoring, but becomes unreliable during crises when it matters most.

Fourth, the technology asymmetrically favors the United States over China. Because China relies heavily on mobile land-based missiles that are harder to track, improved AI-powered satellite surveillance gives the US better ability to detect when China disperses its forces or prepares for conflict. This advantage could paradoxically make China more likely to use counterspace weapons earlier in a crisis, destabilizing the very situation the surveillance was meant to stabilize.

How Experts Recommend Managing These Risks

• Improve Testing: Develop better methods to measure how AI models perform under realistic conflict conditions, including against countermeasures and in degraded environments where satellites may be damaged or destroyed.
• Diplomatic Reassurance: Pursue diplomatic and political measures to address China's concerns about the vulnerability of its nuclear deterrent in the face of advanced US surveillance capabilities.
• Strengthen Space Infrastructure: Focus on improving launch capacity to maintain the US lead in proliferated satellite constellations, ensuring continued access to space-based sensing even if some satellites are destroyed.
• Establish Norms: Promote international norms against interfering with space-based sensing systems to reduce the incentive for counterspace attacks during crises.

The researchers argue that while AI-enhanced satellite imagery offers genuine benefits for arms control verification and crisis monitoring, policymakers must be realistic about its limitations. The technology is not a silver bullet for strategic stability. Instead, it creates a mixed picture where some capabilities improve, others remain constrained by physics, and the overall strategic effect depends heavily on the specific scenario and the countermeasures available to adversaries.

As space-based surveillance becomes more central to how nuclear powers monitor each other, understanding both the promise and the peril of AI-enhanced ISR will be critical to maintaining strategic stability in an era of great power competition.