Researchers have identified a critical vulnerability in AI-powered satellite surveillance systems used for nuclear targeting and monitoring: they're fragile, constrained by physics, and could paradoxically increase the risk of conflict rather than prevent it. A comprehensive analysis from UC Berkeley's Risk and Security Lab and the Council on Strategic Risks examines how artificial intelligence applied to space-based intelligence, surveillance, and reconnaissance (ISR) capabilities affects strategic stability between the United States and China, revealing that the technology creates what experts call "precarious transparency".

The research was prompted by real-world events. When the United States struck Iran in early 2026, it used the Maven Smart System (MSS), a software platform developed by Palantir Technologies that integrates multiple ISR data sources and applies AI to target Iranian forces and nuclear sites at scale. Some experts suggest MSS provides a tenfold increase in targeting capacity compared to prior systems. However, the system's limitations became apparent when it mistook a girls' school in Minab for a military base, highlighting the risks of relying on AI for high-stakes decisions.

What Makes AI-Enhanced Satellite Surveillance So Unreliable?

The research identifies four fundamental weaknesses in AI-powered ISR systems that undermine their effectiveness in real-world conflict scenarios. These vulnerabilities reveal why even advanced machine learning models struggle with the demands of nuclear deterrence and strategic monitoring.

• Model Fragility: AI models used for ISR must classify objects like missiles, ships, and aircraft, but strategic applications require higher-level reasoning using vision-language models (VLMs), which are AI systems that combine image recognition with language understanding. These models are vulnerable to hallucinations, where they generate false information; data poisoning, where malicious inputs corrupt the training data; and misalignment, where the system's goals diverge from human intentions.
• Physical Constraints: Satellite sensing has inherent limitations that AI cannot overcome. Satellites cannot revisit the same location frequently enough for persistent tracking, and certain sensing technologies cannot penetrate cloud cover. These challenges persist even as satellite constellations expand to hundreds of satellites.
• Wartime Vulnerability: During conflicts, counterspace capabilities, which are weapons designed to disable or destroy satellites, severely limit the effectiveness of AI-enhanced ISR. Verification and monitoring tasks benefit more from AI-enhanced ISR because they occur in peacetime, cooperative contexts where satellites face less threat.
• Strategic Asymmetries: The technology favors the United States over China, particularly because China relies heavily on mobile land-based missiles. Improved AI-augmented ISR gives the US better warning of Chinese actions like missile dispersal and warhead mating, as well as non-intrusive verification capabilities.

How Could This Technology Destabilize Nuclear Deterrence?

The asymmetric advantage created by AI-enhanced ISR introduces a paradox: the same technology that could improve transparency and stability might actually trigger instability. Because improved surveillance makes China's nuclear forces more vulnerable to attack, China may respond by deploying counterspace weapons earlier and more intensively in a crisis. Additionally, China may accelerate its already-expanding nuclear force diversification to enhance survivability, creating a new arms race dynamic.

The research emphasizes that while space-based sensing and computer vision for satellite imagery offer novel capabilities, their impacts vary significantly depending on the specific task, the countermeasures employed, and the strategic context. This uneven distribution of advantages across different scenarios is why researchers term the phenomenon "precarious transparency" rather than genuine transparency.

What Steps Could Reduce the Risks?

The research team recommends a multi-faceted approach to mitigate risks while capitalizing on potential benefits. These recommendations address both technical and diplomatic dimensions of the problem.

• Improved Measurement and Evaluation: Develop better methods to assess AI model capabilities under realistic conflict conditions, moving beyond laboratory testing to understand how these systems perform when facing actual countermeasures and adversarial tactics.
• Diplomatic and Political Measures: Pursue negotiations and agreements designed to assuage China's concerns about the vulnerability of its nuclear deterrent in the face of advanced US ISR capabilities, potentially including arms control agreements or transparency measures.
• Launch Capacity Enhancement: Focus on improving the US capacity to launch and maintain proliferated satellite constellations, sustaining the American lead in space-based sensing architecture.
• Norms of Non-Interference: Promote international norms and agreements that prevent interference with space-based sensing systems, reducing the likelihood that nations will target each other's surveillance satellites during crises.

The research represents the output of the AI x Nuclear Research Fellowship, a collaboration between UC Berkeley's Risk and Security Lab and the Council on Strategic Risks, supported by Longview Philanthropy. The analysis draws on open-source research to establish an empirical baseline for ISR capabilities and how AI might augment them, focusing specifically on the development and deployment of space-based ISR capabilities by the United States and China, the two states with the most advanced AI capabilities and largest space-based sensing architectures.

The findings suggest that policymakers and military strategists should approach AI-enhanced ISR with cautious realism. While the technology offers genuine advantages for arms control verification and crisis monitoring in peacetime, its destabilizing potential in wartime scenarios demands careful management through both technical improvements and diplomatic engagement. The Minab school incident serves as a sobering reminder that even a tenfold increase in targeting capacity means little if the underlying data and decision-making processes remain flawed.