building a machine that is able to follow the law may be far more ambitious than one would give it credit for being. Inadvertently, I may have also offered a partial explanation for the fact that soldiers' compliance with the laws of war has often suffered in comparison with the understandably high expectations imposed by governments regarding the protection of the general public. Even for human soldiers, who benefit from an incredible capacity for abstract learning, compositional and hierarchical thought, and robust behavioral flexibility, the demands of rule-following are substantial. The problems posed by following the law—the heterogeneity of laws, semantic variation, and the global reasoning problem—set
an extraordinarily high bar to competence for lethal autonomous weapons and for machine ethics more generally. That is not to suggest that technology will not surpass these considerable milestones; it might even happen far earlier than one could reasonably expect from the present vantage point. After all, recent success in massively scaling transformers, a type of machine learning network especially adept at language generation tasks, should serve as a reminder that advances in artificial intelligence have often been unpredictable, moving in dizzying fits and starts. As these networks have grown in size by orders of magnitude, they have given some indication that novel cognitive abilities may emerge at scale. This will, of course, exacerbate existing difficulties caused by operators' inability to understand machine behavior,
