A new report from the NATO Science and Technology Organization (STO) examines the role of human control in military AI systems. The publication emphasises the importance of balancing human judgement with autonomy and presents a model to understand how control methods interact.
The report asserts that artificial intelligence can provide advantages on the battlefield by accelerating decision-making and quickly analysing large datasets for increased precision and coverage. However, as systems become more autonomous, the need to maintain effective human control and oversight increases. This is according to a press release from NATO that summarises the report's findings.
The report notes that there is broad consensus on the necessity of "Meaningful Human Control" (MHC) on future battlefields, but there are differing views on what the concept entails in practice.
To clarify the issues, a research group within STO has explored the meaning and potential solutions for future AI-based systems. The group, formed under the Human Factors and Medicine (HFM) panel, asserts that MHC should not be regarded as a simple attribute. Instead, it is described as part of a complex socio-technical system that requires actions during different phases of an AI system's lifecycle.
The report identifies 17 potential methods to ensure control. These range from design guidelines to metrics for situational awareness and organisational training. The group also introduces a "holistic bowtie model" of MHC. The model combines aspects at the individual level with system, organisational, and societal levels to enhance understanding of how different methods relate to each other.
The authors stress that there is no universal solution. The report highlights the importance of a human-centred design process where various stakeholders are involved from the early stages of development.
The working group expresses hope that the identified methods will form the basis for further research into practical solutions for human control throughout the systems' lifespan.