Autonomous weapons need to distinguish combatants from civilians but computer vision cannot reliably do this

An autonomous drone is programmed to identify and engage enemy combatants. It uses computer vision trained on images of soldiers in uniform carrying weapons. It encounters: a farmer carrying a long tool over his shoulder (classified as 'combatant with rifle' — false positive), a child holding a toy gun (classified as 'combatant' — false positive), a combatant wearing civilian clothes (classified as 'civilian' — false negative). In testing, the system has a 95% accuracy rate. In a scenario with 1,000 people, 50 of whom are combatants, a 5% error rate means: 47-48 combatants correctly identified, 2-3 combatants missed, and 47-48 civilians falsely identified as combatants. Each false positive is a potential war crime. So what? The fundamental challenge of autonomous weapons is not 'can AI aim a gun' — it is 'can AI make the legal and moral judgment of who to shoot.' International Humanitarian Law requires distinguishing combatants from civilians (Additional Protocol I, Article 48). This is difficult for humans (Rules of Engagement require visual identification of a weapon, hostile intent, and hostile action). For AI, it is currently impossible in the real world: combatants do not wear uniforms in asymmetric conflicts, weapons can be concealed, and 'hostile intent' is a subjective judgment. A 95% accuracy rate that sounds good in a lab means dozens of dead civilians in deployment. Why does this persist? Militaries want autonomous weapons for speed (faster than human decision-making) and scale (operate thousands of drones simultaneously). The ethical/legal barrier is real but the competitive pressure is stronger — if China deploys autonomous weapons, the US feels compelled to match. The result is a race to field systems that are not reliable enough for the legal standard they must meet.