AI systems are vital to overall deterrence postures and replacing conventional systems with reliable second-strike and self-isolating military capabilities.
Horowitz 19 (Michael C. Hororwitz is a former Adjunct Senior Fellow in the Technology and National Security Program at CNAS. Michael C. Horowitz is Richard Perry Professor and the Director of Perry World House at the University of Pennsylvania. He is also an adjunct senior fellow at the Center for a New American Security. His research interests include technology and global politics, military innovation, the role of leaders in international politics, and forecasting.) May 2019 “Artificial intelligence and nuclear stability” THE IMPACT OF ARTIFICIAL INTELLIGENCE ON STRATEGIC STABILITY AND NUCLEAR RISK Volume I Euro-Atlantic Perspectives p.60-70 https://www.sipri.org/sites/default/files/2019-05/sipri1905-ai-strategic-stability-nuclear-risk.pdf) // ZX
Nuclear-armed states, and also non-nuclear-armed states, could use machine learning and autonomy in non-nuclear applications with a strategic effect. Machine learning methods could significantly improve the targeting capability of conventional defensive systems. Missile and air defence systems have relied on automation for decades. The first automatic air defence system, the Mark 56 gun fire-control system, was invented during World War II.24 Since the 1970s, air defence systems have been using an AI technology known as automatic target recognition (ATR) that can detect, track, prioritize and select incoming air threats more rapidly and more accurately than a human possibly could. However, the progress of the target-identification capabilities of these systems has been slow, particularly due to the difficulties associated with the development of target libraries (i.e. the database of target signatures that an ATR system uses to recognize its target). With traditional AI programming methods, the designers of an ATR system have to upload a large and representative sample of data about the target in all conceivable variations of its operating environment (i.e. background and weather conditions). This is a challenging task for many target types and operational situations.25 Advances in machine learning, particularly deep learning and generative adversarial networks (GANs), could significantly simplify that process.26 With deep-learning methods, engineers could make ATR systems capable of learning independently not only the differences between types of target but also the differences between military and civilian objects (e.g. a commercial aeroplane and a strategic bomber).27 With GANs, engineers could generate realistic synthetic data on which an ATR system can be trained and tested in simulation. An ATR system trained with these machine learning techniques would perform comparatively much better than an ATR system trained with traditional methods. Equally, autonomous systems offer new defensive tools against incoming threats. Autonomous unmanned vehicles can be deployed as decoys or flying mines to complement traditional air defences.28 Advances in autonomy for swarming and for multi-vehicle control could also enable autonomous unmanned systems to operate in a coordinate way and conduct advanced A2/AD manoeuvres.29 Such systems would increase deterrence against both conventional and nuclear attack as they would increase the risks for an attack by manned platforms (e.g. combat aircraft and manned bombers) and make the outcome of an attack with unmanned systems (including missiles) more uncertain.