Future Warfare Risks.

 

The conflicts profiled above show how AI is currently being used in warfare. However, the rapid advancement and uptake of AI technologies in warfare means that the number of potential threats and applications is likely to increase significantly in the next decade. These applications and threats are likely to come in one of three forms: • Tactical threats, such as the use of AI-controlled drones warms. • Strategic threats, such as using AI to coordinate entirewarfare operations. • Existential threats, where AI control of critical decision-making systems could lead to mass-casualty events. 

 By 2030, the baseline usage of drones as seen in Ukraine, Gaza, and Iran is likely to spread. Counter-drone systems like Epirus’s Leonidas, naval surface drones like the Magura family, and dedicated swarm-coordination software will mean that the cost of lethality keeps falling and the response time of human commanders keeps shrinking. Drone swarms are a particularly striking example. A drone swarm is a group of unmanned systems that operate as a coordinated whole rather than as a collection of individually piloted vehicles. Each drone runs onboard software for sensing, navigation, and target selection, and the drones share data with one another through a shared communication network rather than reporting back to a single operator. Coordination is distributed: the swarm allocates roles among its members, adapts to losses by reassigning tasks, and can converge on a target from multiple vectors simultaneously. Swarm sizes in operational use today range from several units to several dozen, while demonstration programs have flown hundreds of drones at once, and concept-of-operations documents in major militaries assume swarms in the thousands within the next decade. The military significance of swarms rests on two properties. The first is mass: low unit cost makes each drone expendable, so a defender must either intercept every incoming drone or accept the strike, while the attacker only needs a fraction to get through. The second is saturation: existing air-defence systems, designed against small numbers of high-value missiles or aircraft, are not built to engage dozens of cheap targets in the same minute, and per-shot interceptor costs are often much higher than the drones they destroy. The combination shifts the cost curve of offence and defence in favour of the attacker and compresses decision time for the defender, since coordinated arrival patterns leave little room for human review of individual engagements. Ukraine, Russia, Israel, the United States, China, Iran, and Türkiye have all conducted operational tests or limited combat use of swarm-capable systems since 2021, leading to therapid development of counter-swarm capability such as high-power microwave weapons and interceptor drones.

 Development of AI systems with general cognitive capability has accelerated since mid-2024. Frontier labs have begun activating their highest internal safety protocols, citing the risk that AI could make chemical, biological, radiological, and nuclear weapons easier to build. A US congressional commission recommended in late 2024 that Congress fund a Manhattan Project-style program to race to Artificial General Intelligence. The growth in capacity of AI systems means that they can be used in warfare for applications that go far beyond the tactical engagement layer. The most developed is theatre-level intelligence fusion: systems that pull satellite imagery, signals intelligence, open-source data, and battlefield reports into a single picture, using machine learning to spot patterns at a scale no human analyst can match. Course-of-action analysis and computational wargaming are heading the same way. US programs including DARPA's Strategic Chaos Engine for Planning and Air Combat Evolution are already running thousands of simulated campaigns and recommending forceemployment options to commanders. Logistics and predictive maintenance also become strategic when AI can forecast ammunition use, anticipate equipment failure, and re-allocate supply across a theatre, because sustained operations depend on this layer rather than on the engagement layer above it.

Over and above strategic threats, the use of AI in warfare could also have consequences so severe that they lead to catastrophic, irreversible outcomes. The most consequential of these concerns nuclear command and control. Fears about the potential consequences from the militarisation of AI are shown in Figure 5.4, which traces how close the Doomsday Clock is to midnight. The Doomsday Clock is a symbolic measure established at the beginning of the Cold War indicating how close humanity is to self-inflicted catastrophe, which is represented by the notion of midnight. It was originally used in the context of nuclear war, but it now includes threats related to climate change and emerging technologies. The Bulletin of the Atomic Scientists set the Doomsday Clock at 89 seconds to midnight on 28 January 2025 and moved it to 85 seconds on 27 January 2026, the closest in its 79-year history, citing the expiration of the New START nuclear arms reduction treaty, climate trajectories, AI integration into militaries, and biosecurity concerns. New START expired on 5 February 2026with no successor agreement, despite the Russian Federation’s September 2025 proposal for a one-year voluntary observance ofcentral limits.

Two recent empirical studies highlight the risk of AI involvement in nuclear weapon decision making. Research presented in 2024 found that all five large language models tested escalated conflict in simulated diplomatic and military scenarios, in some cases to nuclear use. A more recent study released in 2026 of 21 simulated nuclear-crisis scenarios across three frontier models found that 95 per cent of games featured nuclear signalling by at least one side, 95 per cent involved tactical nuclear use, and 76 per cent involved strategic nuclear threats. The research summary observed that the nuclear taboo does not seem to be as powerful for machines as for humans. ‘It is almost like the AI understands escalation, but not de-escalation. We do not really know why that is.’ Jacquelyn Schneider, Hoover Institute Wargaming and Crisis Simulation Initiative, 2025 Both superpowers have recognised this issue, with the Biden-Xi meeting in November 2024 affirming the need to keep human control over the decision to use nuclear weapons. The same principle was reaffirmed by 61 states in the REAIM Seoul Blueprint of September 2024, and was the explicit subject of a dedicated UN resolution on AI in nuclear command, control,and communications adopted in November 2025. Whether this principle holds, particularly under crisis instability between nuclear-armed states whose conventional forces increasingly rely on AI-mediated decision support, will define the stability of the nuclear order in the late 2020s. A 2025 assessment notes that integrating AI into nuclear decision pipelines, even where humans retain formal authority, increases the risk that decisions are made too quickly at exactly the points where deliberation is most needed.