The Traditional Model and the Risk-Based Approach in the International Responsibility of States: A Comparative Study in Artificial Intelligence Applications
Keywords:
State Responsibility, Artificial Intelligence, Risk, Based Approach, Due Diligence, Prevention, Risk ManagementAbstract
States now use artificial intelligence systems for border control, disaster forecasting and administrative decisions. When such a system causes harm, a familiar question returns: are the traditional rules of State responsibility still adequate, or should they give way to a new model built on risk management and prevention? Some writers argue that self-learning, unpredictable systems built by many actors call for liability based on risk, even without proof of fault. Others hold that the existing rules are sufficient if they are applied properly. This article compares the two models and asks how they relate to each other when AI is involved.
The study is descriptive, analytical and comparative. It relies on the 2001 Articles on State Responsibility of the International Law Commission and their commentary, the case law of the International Court of Justice and the International Tribunal for the Law of the Sea, and technical risk-management instruments such as the NIST AI Risk Management Framework and the OECD framework for classifying AI systems. The two models are compared on six points: when each enters the analysis, the central question it asks, the place of fault, the role of causation, the source of its standard, and its practical outcome. A hypothetical case of a flood-warning system on a border river is used throughout.
The traditional model is retrospective. It asks, after harm has occurred, whether conduct is attributable to the State and whether it breached an international obligation. Contrary to a common view, fault is not a general condition of responsibility in this model; it depends on the primary obligation. Causation matters mainly for reparation, and the Bosnian Genocide judgment shows that a failure to prove causation with sufficient certainty changes the form of reparation rather than removing responsibility. AI makes this model harder to apply, because the harmful conduct may be a system output, outputs are not always predictable, and several companies contribute to the system. These difficulties concern application and proof, not the existence of rules. The focus of attribution moves from the last operator to the decision-makers and from the moment of output to the stages of development and procurement.
The risk-based approach is prospective. It asks what the State did to identify and reduce risk before and during deployment. Its appeal lies in answering exactly the difficulties the traditional model faces. Its limits are also clear: risk measurement is imperfect, even binding regulation offers no clear method for assessing risk in concrete situations, and most technical instruments are not binding.
The two models share two premises. Neither treats the AI system as responsible, and both value prevention, since the traditional model has long recognised the obligation of due diligence. They differ in their basis. In the traditional model the basis is breach of an obligation; in some versions of the risk-based approach the basis is risk itself. Current international practice adopts the first option. ITLOS describes due diligence as a variable concept, and the ICJ in its 2025 climate change opinion treats the standard as stringent for significant harm to the climate system. Risk therefore enters the law as a measure of what due diligence requires, not as a new basis of responsibility. Three practical consequences follow: risk management defines the expected measures, performing it does not by itself shield the State, and failing to perform it is evidence of a lack of due diligence.
The two models are complementary rather than rival. The risk-based approach identifies the questions and the traditional model supplies the legal answer. What has occurred is a change within the system of responsibility, not a move to a different system. One question remains open: whether a fault-based model will suffice for systems whose behaviour is inherently and fully unpredictable. The article recommends that States assess and document the risks of AI systems before deploying them in public functions, continue that assessment during operation, give operators the information, training and authority to override faulty outputs, and use technical risk-management instruments as guides for measuring due diligence rather than as independent sources of obligation.
Downloads
Publication Timeline
Issue
Section
License
Copyright (c) 2026 فاطمه نادعلیزاده, عاطفه امینینیا, سیده ساناز ذبیحی شهری, فرانه زرافشان (نویسنده)

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.