This thesis analyses States’ duty to investigate grave breaches of humanitarian law and how it applies to deep reinforcement learning autonomous weapons. It identifies basic technologic intricacies related to deep reinforcement learning and discusses what issues may arise if such software is used in weapons systems. The thesis applies a legal doctrinal method to study how the technology could frustrate the grave breaches regime and hamper States’ ability to investigate suspected incidents. Furthermore, investigative standards under humanitarian law and human rights law are examined in the context of autonomous weapons systems. The main argument is that deep reinforcement learning algorithms create a black box that is virtually impossible to investigate and consequently causes accountability issues.