Dual-Use Technology

A dual-use technology serves legitimate civilian purposes but can also be adapted for harmful, coercive or military purposes. It includes goods, software, knowledge and infrastructure whose effects depend on capability, access, user and context. Dual-use potential is not inherent harmfulness; misuse is the actual improper or harmful application.

Why advanced AI models are dual-use

Advanced AI models are often general-purpose: the same underlying capability can be transferred across sectors, combined with other digital tools and deployed at scale. Their security impact depends on capability, user, purpose, deployment environment, access to model weights or interfaces, and connection to external tools.

• Models can detect anomalies, summarise threat intelligence, support secure coding and identify vulnerabilities; similar functions can assist reconnaissance, social engineering or cyber-enabled harm.
• Generative AI supports education, research and software development, but can also facilitate impersonation, automated persuasion and disinformation.
• Computer vision and multimodal models can assist medical diagnosis, disaster response and security monitoring, but may also enable intrusive surveillance or military targeting.
• Language and predictive systems can improve public-service delivery and fraud detection, while biased data or inappropriate deployment may produce discrimination and exclusion.
• Rapid replication and automated operation allow useful or harmful outputs to be produced at speed and scale, while wider access can reduce a developer's control over downstream modification and use.

Advanced AI chips as enabling infrastructure

Advanced AI chips, including powerful graphics processors and specialised accelerators, provide the computing capacity needed to train and run complex AI models. The same hardware can support scientific research, healthcare, industrial automation and public services, or enable military, intelligence, surveillance and cyber applications.

• Highly parallel processing accelerates the mathematical operations used in AI training and inference, without tying the chip to a single end use.
• Powerful chips can support large-scale data analysis, image recognition, autonomous systems and decision-support tools in civilian and defence settings.
• The risk lies mainly in the downstream capabilities enabled by concentrated computing power; specified advanced items may therefore be subject to export controls.

Satellite communication as another example

Satellite communication relays signals through space between fixed or mobile terminals. It supports broadband, broadcasting, remote-area connectivity, emergency communication and disaster response; the same capability can provide military command links, operational communications and data relay.

• Its wide geographic coverage and reduced reliance on terrestrial networks are valuable during disasters and in contested or infrastructure-poor areas.
• Civilian and military services may use related satellites, ground stations, radio links, antennas, terminals, encryption and network-management capabilities.
• Mobility, rapid deployment and connectivity beyond national terrestrial networks make satellite systems useful to civilian agencies, armed forces and potentially harmful actors.

Managing dual-use risks

Governance should be proportionate to capability, likely use and potential harm. For AI, safeguards should cover design, training, access, deployment and monitoring while preserving innovation through safety, transparency, fairness and accountability.

• Developers should conduct risk and impact assessments, evaluate dangerous capabilities, test system behaviour and document known limitations.
• High-risk systems may require controlled access, privacy and security protections, human oversight and traceable accountability.
• Post-deployment monitoring, incident reporting and periodic evaluation help identify unforeseen or evolving misuse.
• Advanced computing and satellite systems may require applicable export controls, licensing, cybersecurity and supply-chain assurance; satellite operations also require spectrum coordination.