Source: Ethica Societas

Military artificial intelligence remains subject to international humanitarian law, Cristina Di Silvio argues in an analysis of human oversight and legal responsibility. Di Silvio distinguishes decision-support, image-recognition and planning tools from autonomous weapons, stressing their different functions, autonomy and operational effects. Incomplete, outdated or biased data can distort target identification, making data provenance and the reconstruction of algorithm-assisted decisions important safeguards. Di Silvio argues that effective human control requires sufficient understanding, time, information and authority to modify, suspend or terminate an action. The analysis explains that proportionality concerns whether reasonably expected incidental civilian deaths, injuries and damage to civilian objects would be excessive relative to the concrete and direct military advantage anticipated. That assessment relies on information reasonably available when the decision is made; AI does not remove obligations to take feasible precautions, verify targets or cancel or suspend incompatible attacks. Di Silvio also cautions that satisfactory experimental results cannot alone establish a system’s suitability for a particular operational use.

The analysis identifies weapons reviews under Article 36 of Additional Protocol I as an obligation for States Parties and examines their application to AI systems. Di Silvio argues that reviews should address predictability and foreseeable uses, while software updates, data changes and operational conditions may require renewed assessment. The International Committee of the Red Cross has proposed binding prohibitions on unpredictable autonomous weapons and systems designed or used to apply force against persons; the analysis notes that these proposals are not a universally adopted legal framework. Di Silvio recommends lifecycle documentation, operator training and verification under realistic conditions, including operational stress and sensor degradation. The analysis explains that the EU AI Act’s exclusion for exclusively military, defence or national-security uses does not displace applicable international humanitarian law or responsibility rules. Records of inputs, model changes and human decisions can help distinguish design failures from improper deployment and support subsequent investigations. Di Silvio argues that responsibility must be assessed under the applicable legal requirements, distinguishing State responsibility, individual criminal responsibility and other forms of liability.