once performed by humans, the balance of agency tilts. Incremental AI advances could steadily erode human influence over the economy, culture, governance and societal systems.165 The more that AI agents themselves are used in R&D to develop AI agents further, the greater the risk that the technology companies managing them could cease to understand how those AI systems work. Such R&D automation could accelerate the timeline for progress in AI, making it even more difficult for humans to build the technical and regulatory capabilities to keep pace.166 Military misuse or mistakes Following Russia’s invasion of Ukraine, both sides in the conflict have pushed forward the boundaries of AI use in military conflict. AI technologies have played important roles in geospatial intelligence, autonomous systems, and cyber warfare, among other areas.167 As militaries embed AI deeper into intelligence, surveillance, logistics, and command functions, the risk landscape will shift from tactical to systemic. AI will increasingly influence how militaries perceive threats, make decisions, and take actions. AI system failures could propagate through entire chains of command and deterrence systems. Without humans firmly in the loop, AI-powered platforms may misidentify threats,168 respond to biased data,169 or behave unpredictably in conditions outside their training parameters.170 Adversaries might use data poisoning – introducing corrupted data during model training – as a covert weapon to undermine military AI systems.171 When humans are in the loop, an additional set of risks needs to be considered. Weaponized generative AI models can instantly fabricate executive orders or create synthetic, convincing battlefield footage, potentially confusing both humans and technology-based responses. Human decision-making is influenced by cognitive biases, such as confirmation bias or recency bias, when interpreting AI outputs. This can become especially challenging in conflict conditions, when it might also be tempting to over-rely on AI systems even if these are not yet fully equipped to provide nuanced decision-making support.172 The speed at which AI systems operate, when applied without checks and balances, can itself be a source of risk. Military crises that once unfolded over days or hours could instead escalate in seconds. An automated early-warning system misinterpreting a missile test, for instance, could trigger defensive responses from an adversary's AI system, leading to a conflict started by technical error rather than strategic intent. Traditional deterrence, built on human deliberation and diplomacy channels, may not hold when algorithms initiate actions before leaders can act. With countries starting to implement AI tools for managing nuclear weapons stockpiles and in some areas of nuclear weapons command, control, and communications, addressing such risks becomes especially critical.173 However, major powers are rushing to integrate AI across military domains, each fearing strategic disadvantage if rivals move first. This dynamic incentivizes rapid deployment over rigorous safety testing, increasing the probability of failures precisely where consequences are most severe. The intense pace of innovation makes it unlikely Juli Kosolapova, Unsplash Global Risks Report 2026 65