UN Fact-Finding Mission on Iran: U.S. War Crimes and Precautions in AI-Enabled Warfare

[Jessica Dorsey is an Assistant Professor of International Law at Utrecht University School of Law and Managing Editor of Opinio Juris.]

The UN Independent International Fact-Finding Mission on the Islamic Republic of Iran (FFMI) released its 18-page report to the Human Rights Council on 17 September 2026. In the report, among other findings, the Mission outlines reasonable grounds to believe that the United States was responsible for two attacks on 28 February 2026: a strike on an elementary school in Minab, which killed at least 157 people including 120 children and at least 26 school staff and wounded 95 others, and a strike on a sports center in Lamerd, causing a further 20 civilian deaths and at least 100 injuries. The FFMI’s conclusion was that these attacks constituted the war crime of launching indiscriminate attacks (paragraphs 126, 128).

Though the report is worth reading in its entirety, this post focuses mainly on one aspect of the FFMI report: the analysis of feasibility of precautions, largely outlined in paragraphs 123-128 of the report. This analysis reflects an issue increasingly at the center of modern warfare, namely what it means to take all feasible precautions when the conduct of hostilities is increasingly being shaped by technological complexity that enables speed and scale through systems that affect qualitative legal decision-making. This report provides a concrete avenue to examine lines of argumentation in recent scholarship by myself and others.

The Primacy of Precautions

For a recent article for the International Review of the Red Cross, I examined what I termed “quantification logics” or the translation of complex legal and ethical judgments into numerical or statistical models that underpin algorithmic and AI-enabled technologies. I also analyzed how those logics can shift human cognitive focus toward what is able to be calculated on a battlefield and therefore away from what is not (I reference the McNamara or quantification fallacy: prioritizing

“measurable variables while disregarding critical qualitative or human factors simply because they are not easily captured through measurable elements. At its core, the fallacy assumes that what cannot be quantifiably measured is irrelevant, which results in flawed reasoning.”

Some examples are the estimated civilian harm, the anticipated military advantage, confidence scores and other probabilistic modeling aspects provided by AI-enabled systems. As I mention,

“[p]articularly in the context of armed conflict, where the environment is highly complex and dynamic, this can produce grave operational, legal and ethical consequences, including risks of civilian harm.”

With the focus on these quantitative metrics, I argue, the over-reliance on computational models can skew and displace contextual qualitative judgment that is essential to IHL compliance.

Specifically in the rules and principles governing taking precautions under Article 57 of the First Additional Protocol to the Geneva Conventions (API),  which impose a duty of constant care to spare the civilian population throughout military operations and the taking of all feasible precautions or to any extent minimize civilian harm during attacks. Both obligations form part of customary IHL. These primary obligations impose specific constraints on both the planning and conducting of attacks and undertaking a proportionality assessment does not supplant a party’s obligations to minimize or avoid harm prior to carrying out such an assessment. As targeting decisions are increasingly affected by AI-enabled systems, particular attention must be paid to this ordering.

From a legal perspective, it is insufficient to only interrogate the civilian harm estimated right at the point of attack. Instead, information about whether the target was properly identified in the first place, what kind of information was available and was it reliable (e.g., not outdated or stale), what uncertainties were included in the analysis and what kinds of measures could have been taken earlier should be at the heart of the inquiry. The sequencing is not just relevant to when a party invokes proportionality to justify anticipated harm, but equally to situations like this one where a party may claim after the fact that a strike was a mistake. Whether a mistake is reasonable (and in that sense not an IHL violation) is a question that turns on the precautions taken beforehand, not as an ex post judgment to be made after the fact.

Minab

This brings me to the strike in Minab, Iran. The FFMI’s analysis is concise and to the point. Paragraph 124 outlines that the U.S. failed in doing everything feasible to verify that the Shajareh Tayyebeh school was not a civilian object. And then the paragraph continues: “updating targeting information and acting upon such information falls within the category of feasible precautions an attacking party is required to take under IHL,” before stating in paragraph 126 the “Mission finds reasonable grounds to believe that the U.S. committed the war crime of launching an indiscriminate attack resulting in loss of life or injury to civilians or damage to civilian objects.” This largely tracks with a recent piece Taylor Kate Woodcock and I published in July, arguing that the continued use of known outdated targeting intelligence could be considered a failure to take feasible precautions and ensure distinction, raising questions about the reckless disregard for foreseeable civilian harm (especially in this case, where authorities were reportedly aware of the outdated satellite imagery). As Reed Brody mentions, “keeping your targeting file current is not optional.”

AI-enabled systems, like those used by the U.S. in Iran, can exacerbate this problem through taking incomplete or outdated information, feeding that into a database making it easier to retrieve, prioritize and operationalize at speed and scale. In such an environment, outdated imagery can be reproduced or pushed through these systems and presented as actionable at the same time ironing out opportunities to detect or interrogate errors that might arise. This case shows a clear and tragic reason not only to scrutinize the final target nomination but also the process, especially in a high-tempo targeting environment, of how that nomination came to become part of operational decision-making. This strike also raises the issue of institutional learning. Previous incidents like two 2025U.S.strikes in Yemen, and the Kabul strike in August 2021, show that civilian harm risks are often associated with faulty or outdated intelligence. The fact these kinds of risks persist also highlights the importance of interrogating not only the information itself but whether lessons from the past have been integrated into subsequent targeting practices.

Constant Care and Upstream Legal Obligations in AI-Enabled Warfare:

In thinking about the broader duty imposed by Article 57 of sparing the civilian population across the conduct of military operations in terms of increasingly autonomous warfare, there is legal space here to consider precautions much earlier in the operational use of these systems. If targeting information must be updated, attention given to system limitations and opportunities for verification preserved, the decision-making around precautions likely also arise well before the planning or deciding upon an attack. This is an aspect I have also recently examined in a draft article, on precautions by design, in which I argue that feasibility is related to technological capacity that can expand, not only limit, what a military is able to do (or in IHL parlance: what is “practicable or practically possible”) under particular operational conditions. Neil Cameron also analyzes that angle in respect to U.S. strikes in Iran here.

If a party is operating with systems that can cross-check, identify inconsistencies, preserve audit trails or flag particular uncertainties, this may well implicate additional opportunities for compliance with its precautionary obligations. Another related argument I highlight in that piece is that a military is not at liberty to interpret the understanding of feasibility due to design choices or constraints it sets up itself. This underscores the legal significance of robust and meaningful civilian harm mitigation and response (CHMR) policies. Measures developed through CHMR approaches, which include institutionalized mechanisms for learning from civilian harm incidents, improving intelligence practices and strengthening internal review mechanisms are not just policy aspirations but, more compellingly, practical ways to fulfil existing precautionary obligations under IHL.

The fact that a military chooses to accelerate its targeting process does not establish on its own that additional verification or validation has become impractical or practically impossible. Prioritizing rapid decision-making and execution of targets (in this case, nearly 5,200 targets reportedly struck in the first 96 hours of Operation Epic Fury) is not a legal requirement, it is a normative and operational choice. If that choice renders insufficient time for verification and validation, the response requires adjusting the operational tempo and not the requirements of IHL.

Lamerd: Choice of Means and Methods

I now turn briefly to another aspect of the FFMI’s report with respect to the strike in Lamerd, for which the findings implicate war crimes by the US. This report cites investigative work by Airwars, which showed that the PrSM missile used in striking a sports center, an adjacent elementary school and residential buildings, designed to disperse 180,000 tungsten pellets over a wide area with a blast radius of 170 meters, could not in the circumstances prevailing at the time be directed at a specific military objective or have its effects limited as IHL requires. IHL also prescribes constraints on means and method of attack, requiring parties to determine prior to employing a new weapon, whether its characteristics and effects allow the user to comply with IHL requirements of distinction and precautions in the circumstances of its use. In this case, the FFMI report resoundingly concludes that the U.S. did not, in paragraph 128:

“Accordingly, the strike on Lamerd also amounted to the war crime of launching an indiscriminate attack resulting in loss of life or injury to civilians or damage to civilian objects. The Mission concluded that the U.S., by choosing to use the PrSM missile in a populated civilian area, under the prevailing circumstances, launched an attack that struck civilians and civilian objects reckless as to the likely extent of death or injury.”

Accountability and disclosure

Finally, in paragraph 129, the FFMI report joins calls for theU.S.to take steps immediately to disclose its own internal investigation findings. As of 18 September 2026, 202 days have passed since the strikes. It is in the interest of the U.S.to substantively respond to this call in order to have a better understanding of what information might have been available, how the targets were identified, which precautions were considered and taken and whether and which internal problems or concerns might have been flagged-up internally (and why those may not have found their way into further stages within the targeting process). The findings of the FFMI are of course investigatory in nature and not judicial findings and criminal responsibility at the individual level is distinct from State responsibility under international law. However, transparency regarding the internal investigations would help pierce the veil of opacity surrounding the circumstances of the attacks and develop measures for the future that would prevent their recurrence.

Conclusion

The findings of the FFMI report on the strikes in Minab and Lamerd allow for a more granular focus on the precautionary framework embedded within IHL, a framework that can sometimes get lost in the complex machinery and high tempo of increasingly sophisticated targeting systems and processes. Feasibility as a concept is contextual but cannot be defined by the operational tempo a military creates itself. Technology in this sense should be seen to offer both constraints as well as possibilities for new verification, review and monitoring. This demonstrates how the obligation to take feasible precautions and ensure constant care across military operations can reach much earlier than just the time of planning and deciding on an attack. In terms of AI-enabled targeting, there is opportunity for information maintenance, design and testing of systems, and building in moments for meaningful opportunities for contextual qualitative human judgment.