Abstract
Background The adoption of AI for military targeting, defence and engagement operations changes the tempo, magnitude and scope of lethal decision making. Both Iran and Israel have introduced systems that incorporate AI in different ways, technologically and doctrinally, in the Iranian-Israeli confrontation. The issue that is being analysed here is not about which state is technologically more advanced but if meaningful human control of the application of force can be maintained in a high volume, under compressed decision time algorithmic system. Method The individual AI-enabled system or application is the primary unit of analysis, whereas Israel and Iran are higher-level comparative contexts used for interpreting the findings at the system level. Systems are first broken down into five functional categories (AI-supported target generation, automated defence interception, assisted fire control at the weapon, autonomous platforms and swarms, and cyber and information applications) based on decision tempo, the location of the human in the decision cycle, and the legal test that must be used. A six indicator assessment is conducted for each system across four dimensions of the system: technical capability, operational employment and effectiveness, human control, and legal compliance. Five indicators are measured as an ordinal capability band with a within band rubric, while legal compliance is not scored but is measured qualitatively. There is an explicit level of confidence attached to each assessment, differentiating between documented and assessed competence. Sources were identified and sourced from peer-reviewed literature, OSINT, official documents and investigative journalism and then prioritized on a five-tiered scale. Results The two national sets don't diverge quite as evenly as the aggregate comparison implies within the categories that make up functional sets. The undetermined Iranian capability in target generation is in contrast with the documented high level of Israeli capability. In defensive interception, high automation comes from the physics of the threat. With assisted fire control, AI will be added to the fire control system, but it will only reinforce human control – its authority will be purely negative. For both autonomous platforms and swarms, the capability sets are closest and the control problem is most acute for them. In cyber and information applications the evidentiary asymmetry inverts. Technical accuracy, operational effectiveness, legal compliance and humanitarian outcome are treated as separate constructs throughout. Conclusion However, the limitation on meaningful human control is not the sophistication of the algorithm, but the ratio of the number of engagement options that can be provided by a system in a given time to the time taken to interrogate them. Two paths of degraded human control are at work, one from advanced military forces, one from resource-poor ones, both moving toward the same end point. Governance should thus directly control the tempo of decisions, the speed at which decisions pass through, and the review of decisions, and not leave autonomy as an on/off switch on the weapon. A set of 17 specific assignable mechanisms are suggested.