When the Rome Statute was adopted on 17 July 1998, its drafters gave the International Criminal Court jurisdiction over natural persons only. Article 25(1) says so in one line, and the question of trying companies was left to national law. For a quarter of a century that choice mattered mostly for mining firms, arms brokers and chemical suppliers. It matters differently now that the product a company ships to an army can be a system that proposes targets.
On 29 October 2026 the ICRC's Humanitarian Law & Policy blog published “Contemporary accomplices? Considering the potential liability of AI targeting system providers under the Rome Statute” by Rosa Villar Orero, a PhD candidate at the University of Aberdeen. It appears in the blog's Emerging Voices series, so it carries the author's argument and should not be read as an institutional ICRC statement. Two weeks earlier the same blog ran Cordula Droege's “AI will not set its own limits in warfare. States must.” of 15 October 2026, which addresses what states owe. Villar Orero's piece turns to the people who sell the systems.
Practical takeaway. The argument rests on Rome Statute Article 25(3)(d)(ii), which requires knowledge of a group's criminal intention and does not require purpose. Only individuals can be charged, so the exposure falls on the executive with power to approve or veto a contract. On this reading, widely reported civilian harm linked to a client's operations is material from which a court might infer knowledge; a prosecutor would still have to prove that crimes were committed and that a specific executive was aware of the group's intention to commit them. No court has applied this provision to a technology supplier.
What Villar Orero argues: an executive who supplies an AI targeting system as an accessory to war crimes
The piece opens with Bruno Tesch, whose firm Tesch & Stabenow supplied Zyklon B to the concentration camps. A British military court convicted him in 1946 although he never entered a camp; the author's point is that the tribunal looked at what he knowingly supplied, and his distance from the killing did not save him. She then asks what Tesch would sell today and answers with algorithms: systems that correlate data, find patterns and recommend military objectives at a speed and scale that human analysts cannot match.
Her question is whether the senior executives of a technology company whose AI system contributes to war crimes can be held criminally liable under the Rome Statute. The company itself cannot be, because Article 25(1) of the Statute reaches natural persons only. The author therefore looks for the individual with effective decision-making power, borrowing the post-war I.G. Farben standard: a defendant who took part in the act, or who knew of it, had authority to prevent it and failed to do so.
Why Article 25(3)(c) aiding and abetting requires purpose, and Article 25(3)(d)(ii) requires knowledge
The obvious provision is Article 25(3)(c), which covers a person who aids, abets or otherwise assists a crime “for the purpose of facilitating” it. A vendor that sells a decision-support system under an ordinary procurement contract will say its purpose was revenue, and the author accepts that proving criminal purpose against a chief executive is, in practice, very hard. The result is a gap for commercially indifferent suppliers, the same neutral-contribution problem that scholars have debated for decades.
Article 25(3)(d) is drafted differently. It covers anyone who “in any other way contributes” to a crime committed by a group of persons acting with a common purpose, provided the contribution is intentional and either furthers the group's criminal purpose or, under limb (ii), is made in the knowledge of the intention of the group to commit the crime. Knowledge is defined in Article 30(3) as awareness that a circumstance exists or that a consequence will occur in the ordinary course of events. Villar Orero presents this as a return to the standard the post-war tribunals applied to Tesch, refined in the Krupp case, where knowledge could not be inferred from a defendant's position alone and had to be proved.
How the 2014 Katanga judgment's five elements would apply to an AI targeting-system supplier
ICC Trial Chamber II gave Article 25(3)(d) its first substantive reading in the Katanga judgment, which found Germain Katanga guilty on 7 March 2014 of four counts of war crimes and one crime against humanity for the 2003 attack on Bogoro in Ituri, after recharacterizing his mode of liability from principal to accessory. Villar Orero maps its five elements onto an AI supplier:
First, a crime within the Court's jurisdiction must have been committed; she points to ICC arrest warrants issued in current conflicts for intentionally attacking civilians under Article 8(2)(b). Second, the perpetrators must form a group acting with a common purpose, which Katanga held need not be a formal organization; the author treats a formal military structure as strong evidence of an organized group, while the common purpose itself would still have to be inferred from conduct such as targeting patterns. Third, the accused must have contributed. Here she drops strict but-for causation for a risk-based test: the system need not be indispensable, but its supply must have materially increased the risk that the crime would be committed. Fourth, the contribution must be intentional, which a signed commercial contract readily shows, because, as the author applies Katanga, intent attaches to the act of providing the system. Fifth, the supplier must know the group's intention at the moment of supply.
On that fifth element she sides with the 2011 confirmation-of-charges decision in Mbarushimana, which accepted general knowledge of a group's criminal intentions, over Katanga's crime-by-crime knowledge. Her reason is structural: a supplier who knew of each specific attack would effectively control it and would face a heavier mode of liability than Article 25(3)(d)(ii).
Where the argument is weakest: causation, knowledge of intent, and the absence of case law
The contribution element carries most of the weight, and the author herself calls it the most contested. A risk-based test fits a targeting system well on paper, because target generation at volume, combined with error rates and loose collateral-damage thresholds, plausibly shapes how often and how systematically unlawful strikes occur. A prosecutor would still need to show what the specific system did in specific operations, and that information sits inside classified military workflows the vendor's executives may never see.
The knowledge element turns on public documentation. The author argues that where significant civilian harm is widely reported from operations in which a technology plays a central role, an executive with authority over the contract “could struggle to credibly claim ignorance of the general pattern of violence.” That is a strong claim about inference, and it would make NGO reports, investigative journalism and ICC filings the de facto notice file for every defense-technology board. It also sits uneasily with the Krupp refinement she relies on, which required proof of knowing participation beyond position.
Two practical limits remain. The ICC has never applied Article 25(3)(d)(ii) to a corporate supplier of any kind. The Court's jurisdiction depends on territory, nationality or a Security Council referral; several major AI-producing states are not parties to the Statute, although their nationals can still fall within jurisdiction for conduct on the territory of a state party. The EU's AI rules stop short of these systems: Article 2(3) of the EU AI Act excludes AI systems placed on the market, put into service or used exclusively for military, defense or national security purposes. Export controls under the EU Dual-Use Regulation (EU) 2021/821, sanctions regimes and national law can still reach such suppliers; the AI Act's own conformity and transparency duties do not.
How Quentir Reads It
Villar Orero's argument belongs to a wider question that Quentir has followed this month from a different direction: who answers when an AI system acts for someone else. Delaware's draft AI Company Act, which an earlier Quentir post read against its text, routes that answer through a capitalized shell and a liability shield for the member. The Rome Statute route runs the opposite way. It sets the corporate form aside and looks for the individual who approved the contract while knowing what the client was doing.
For companies that sell analytics, computer vision or decision-support software into defense markets, the practical consequence is about knowledge. If the author's reading gained traction, the material a compliance team already monitors for export-control and sanctions screening, public reporting on a client's conduct, could become evidence from which a court infers what executives knew. Article 30(3) still requires proof of an individual's actual awareness, so the reporting would support that inference without settling it, and reported civilian harm would also have to be tied to conduct that meets the elements of a war crime.
The argument also exposes a mismatch between regimes. Under the AI Act, civilian high-risk systems, such as those listed in Annex III for law enforcement and border control, are to carry conformity assessments, logging duties and post-market monitoring once the Act's staggered high-risk obligations apply, while the military systems with the highest stakes for civilians fall outside the AI Act and are governed by humanitarian law that was written for weapons and soldiers. Quentir follows these threads across its blog, Defense Monitor and Medicine Monitor, and the All-access membership keeps that archive in one place, including every Signature Brief and Signature Report as it publishes.
If the ICC Office of the Prosecutor opens a case that reaches a technology supplier, Article 25(3)(d)(ii) will move from a blog argument to a charging decision, and the five Katanga elements will be argued on real procurement files. A national prosecutor acting under universal jurisdiction would charge under domestic complicity law, which may set a different mental element.
Sources: Rosa Villar Orero, Contemporary accomplices? Considering the potential liability of AI targeting system providers under the Rome Statute, ICRC Humanitarian Law & Policy blog, Emerging Voices series, 29 October 2026; Cordula Droege, AI will not set its own limits in warfare. States must., ICRC Humanitarian Law & Policy blog, 15 October 2026; Rome Statute of the International Criminal Court, adopted 17 July 1998, Articles 8, 25 and 30; ICC Trial Chamber II, Judgment pursuant to article 74 of the Statute, The Prosecutor v. Germain Katanga, ICC-01/04-01/07-3436, 7 March 2014, and the ICC Katanga case page; ICC Pre-Trial Chamber I, Decision on the confirmation of charges, The Prosecutor v. Callixte Mbarushimana, ICC-01/04-01/10-465-Red, December 2011; Trial of Bruno Tesch and two others (the Zyklon B case), British Military Court, Hamburg, 1946, via the ICC Legal Tools Database; Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 2(3) and Annex III; Regulation (EU) 2021/821 (Dual-Use Regulation). Links checked 30 October 2026.
Published intelligence, built to inform your own decisions. Published: October 30, 2026.