General-Purpose AI and the Future of Drone Autonomy in Contested Environments
Militaries are racing to field increasingly sophisticated autonomous drones. Future concepts for drone warfare envision the mass employment of unmanned systems that can reliably execute missions on complex battlefields without human control. Today’s drones cannot yet deliver these capabilities consistently, but advances in general-purpose AI may allow militaries to take these concepts from theory to practice.
For drones to be combat-effective, they will have to operate in contested environments, where internet and cloud compute access are denied, degraded, intermittent, or limited. Technical feasibility, then, does not guarantee operational usefulness: AI models powering autonomous systems need to be small enough to run locally, at the tactical edge, and fast enough to operate at the speed of combat.
How militaries pursue drone autonomy will be shaped by unavoidable trade-offs in where they access computing power and which AI models they field. The most sophisticated models will likely require drones to draw on external computing power, making them dependent on links that enemies can jam or exploit. Smaller models can avoid this vulnerability by running on the drone itself, but they offer narrower capabilities. Militaries will therefore likely pursue a combination of approaches.
How drone autonomy advances has critical implications for military planning, global competition, and the risks posed by future autonomous weapons. This paper assesses the feasibility of using general-purpose AI for drone autonomy today, forecasts when models matching today’s frontier could be deployed on drones, and offers recommendations for the US military, AI developers, and policymakers.
Findings
Despite rapid progress, general-purpose AI systems still lack the capabilities and processing speed to match the performance of human drone operators. Smaller, narrower models can operate with low latency, but they are not equipped to handle many complex tasks and decisions on real-world battlefields. Larger, more powerful general-purpose systems are capable of complex, contextual decision making but have not yet shown that they can perform reliably under variable conditions, and they are too slow for many fast-paced drone tasks (Figure ES1). Militaries will likely still field less robust autonomous systems, however, especially where decisions are lower-risk, like non-lethal use cases and operational environments without civilians or friendly forces present.
Compute and power requirements create additional barriers to fielding the most capable general-purpose models on small drones. Stark trade-offs exist between model capabilities and the size, weight, and power a platform can bear. Because more capable models need more computing power and a heavier, costlier, more power-hungry chip, small drones are unlikely to be able to field the most capable models in the immediate term (Figure ES2). Militaries are therefore likely to field sophisticated autonomy first from offboard compute sources and onboard small numbers of large, expensive systems, while smaller, mass-produced systems possess more limited onboard capabilities.
If current hardware constraints and trends in model efficiency hold, it will be at least one to three years before systems as capable as today’s frontier AI can fit onboard most drones. Frontier models require too much memory to fit onboard drones today, but models matching their performance that are small enough to fit onboard drones will likely become available within the next few years (Figure ES3). This indicates only when drones could physically carry today’s peak level of intelligence; operational effectiveness will also require progress in drone task performance and latency.
The US currently holds advantages in producing chips for onboard drone autonomy, but America’s leverage is not guaranteed. The most widely fielded edge chips are Nvidia’s Jetsons, which are also used by US adversaries in China and Russia. The best chip for onboard autonomy is not the one with the best raw performance, but the one that performs most efficiently within the limits of its platform. Jetsons are currently preferred, but a variety of competitor systems claim more peak compute, better efficiency, or lower prices.
If frontier AI companies are unwilling to support edge deployment of their systems, Chinese open-weight models may outclass fieldable American models. America’s leading AI companies could help it secure a lead in drone autonomy, but their current approach poses challenges for developers seeking to use their models for these purposes. These companies currently restrict their models with safeguards to prevent misuse and by keeping their model weights private. Because of these restrictions, drone developers face barriers in using them to write code for military autonomy applications and cannot fine-tune them on military data. Furthermore, these companies do not prioritize making their models efficient enough to run outside of data centers, and they may be unwilling to allow their models to be deployed locally in combat zones, due to the risk of model-weight theft. Open-weight models solve many of these issues, but the best of these systems come from China, forcing the US military to choose between Chinese models and potentially inferior American technologies.
Recommendations
The US military should plan around today’s technical and operational challenges to drone autonomy, but prepare for rapid capability gains. Given the continued brittleness of AI models in unpredictable and contested environments, the US military should prepare to train and maintain a large number of drone pilots for the foreseeable future. Meanwhile, the Pentagon should lay the groundwork for fielding more advanced systems in the future by developing evaluations to test whether autonomous systems meet human decision-making standards under harsh real-world conditions. Since it remains unclear which technical paradigms will deliver the most robust drone autonomy, the military should experiment with a variety of model types and compute configurations. The Pentagon should also develop programs to collect and share high-quality data that developers can use to train autonomy models.
Frontier AI developers should consider supporting the development of safe and effective drone autonomy models. US frontier AI companies should consider offering trusted access programs that let vetted defense firms use frontier coding tools with fewer guardrails to build military autonomy models. They should also consider developing models tailored for compute-constrained operations for military purchase. Otherwise, the US military may have to rely on less robust and reliable American models or Chinese products that could introduce unacceptable vulnerabilities. Valid safety concerns exist about accelerating the development of autonomous drones, but frontier AI companies should at minimum weigh those risks against those discussed here. Finally, frontier AI companies should closely monitor the use of coding agents for drone development by US adversaries and, given the risk of adversarial distillation, consider withholding their most drone-relevant capabilities from public release.
US policymakers should seek to secure an American lead in drone autonomy through policies that (1) encourage models suitable for edge deployment and (2) restrict adversary access to leading American chips. Policymakers should support funding for general-purpose AI that developers in the US military and defense industry are able to freely customize for defense applications. Open-weight models offer one approach, but given the risks of enabling adversary access to these systems, policymakers should also consider support for proprietary models that only vetted defense developers can access and modify. The US should also tighten export controls not just on the top chips used for developing frontier AI, but also on those designed for edge deployment, and increase the enforcement capacity of the Bureau of Industry and Security accordingly.
