<p>Fully autonomous weapon systems need to comply with International Humanitarian Law and underlying ethical principles. This requires the ability to recognize not only objects or persons to be targeted but also protected persons or objects. Such sophisticated object classification abilities, if achievable at all, would have to utilize machine learning techniques. These come with well-known limitations to predictability, reliability and explainability. This article argues such limitations could be overcome to a sufficient extent by thorough and realistic testing efforts. These efforts, even if they detected satisfactory performance, could not generate certainty of compliance. Yet they could result in compliance having been achieved becoming the best explanation of empirical evidence and the most reasonable belief to have about the specific AWS model. However, the limitations of this approach – its reliance on quantity and quality of data unavailable to most actors and the fragility and limited durability of positive outcomes – would limit its practical utility. Issues of predictability are thus likely to limit AWS use to a highly restricted, closely supervised kinds of operations analogous to the current use of area-affecting munitions or to environments virtually free of noncombatants. </p>