The relationship between artificial intelligence systems and International Humanitarian Law has been extensively analysed in terms of whether AI systems may themselves comply with humanitarian obligations, whether autonomous weapons systems may lawfully be deployed, and whether meaningful human control may be preserved across the operational chain. This paper proposes a distinct analytical category that may complement existing scholarship.
The category proposed is mediated compliance: the condition under which the practical operability of an International Humanitarian Law obligation depends on technical infrastructure designed and procured by actors distinct from the human decision-maker who bears the formal obligation. Where sensor fusion, automated classification, and algorithmic prioritisation mediate the information that reaches the human operator, architectural choices made upstream of the engagement decision may determine which compliance-relevant signal classes the operator is in a position to perceive.
The paper develops the concept in five dimensions: analytical foundations, a taxonomy of architectural mediation forms (sensor filtering, taxonomic exclusion, threshold calibration, priority weighting), a comparative analysis with analogous doctrines in domestic legal orders (United States, Germany, Italy, Switzerland), implications for state responsibility under ARSIWA, and operational implications for procurement and audit. Special attention is given to the institutional position of Switzerland as depositary of the Geneva Conventions and host state of the ICRC.
International Humanitarian Law · Mediated Compliance · Architectural Mediation · State Responsibility · ARSIWA · Command Responsibility · Corporate Criminal Liability · Swiss Penal Code Art. 102 · AI-Mediated Operational Systems · Geneva Conventions
This paper develops a concept introduced in summary form in the author's earlier work: Lawful Operational Safeguards in AI Systems (2026).