The February 28, 2026 strike on the Shajarah Tayyebeh elementary school in Minab, southern Iran, has become a major case study in the risks of using artificial intelligence to accelerate military targeting. A Pentagon investigation, according to officials familiar with its findings, identified a chain of failures involving outdated intelligence, an incorrect classification of the site, compressed targeting timelines and heavy reliance on AI-enabled systems, Bloomberg said early this week as revealed by Pentegon officias involved in an inside investigation.
The full investigation has not been publicly released, so some details remain based on accounts from officials involved in the inquiry. Here we are trying to explain the series of actions or inactions that led to the US strike on the Iranian school, and what is an AI-assisted 'Kill Chain' based on the Bloomberg report and other sources.
What happened at Minab?
The school was located on land that had previously been part of an Iranian military compound. US intelligence databases continued to identify the location as an Islamic Revolutionary Guard Corps facility even though the site had changed substantially over the years.
Satellite imagery showed that the school had been separated from the military area by walls and separate entrances. A playground, sports markings and other visible indications of civilian use had also appeared at the site years before the attack. According to officials familiar with the investigation, these changes were identified by at least some intelligence analysts, but the information was not incorporated into the principal military database used for targeting.
This created the basic problem that ultimately travelled through the targeting process: the physical site had changed, but the military intelligence record had not been adequately updated.
Where artificial intelligence entered the process?
The US military was using Palantir's Maven Smart System to process large volumes of intelligence and assist with military planning and targeting. The system brings together information from sources including satellite imagery, surveillance and other intelligence and presents it to military planners and commanders.
The Washington Post reported in March that Maven, with AI technology incorporated into the system, was being used to generate and prioritise targets during the Iran operation. The system could dramatically shorten processes that traditionally required much longer human analysis. US Central Command commander Adm. Brad Cooper said the military's AI systems helped personnel process vast quantities of information rapidly, while humans retained the final authority over decisions about what to attack and when.
The important distinction is that the available reporting does not establish that an AI system independently decided to attack a school. Rather, AI was incorporated into a broader human military targeting process. The problem was that the system was working with information that apparently contained a critical error about the identity of the site.
The first failure: outdated intelligence?
The Minab location had been classified as a military installation years earlier. By the time of the attack, however, the site had developed into a functioning school.
According to Bloomberg's reconstruction of the Pentagon investigation, satellite imagery showed changes at the site dating back almost a decade. An intelligence analyst reportedly identified some of these changes as early as 2019. However, the analyst's observations were recorded in a system that was not connected to the primary intelligence database used for targeting.
Consequently, information indicating that the site had changed did not effectively replace the older military classification. The targeting system continued to treat the location as an IRGC facility.
The second failure: an accelerated targeting process?
The circumstances surrounding the February operation added another layer of risk. US planners were preparing an exceptionally large opening assault against Iranian targets. More than 1,000 targets were reportedly struck during the first 24 hours.
According to officials cited by Bloomberg, the collapse of diplomatic negotiations shortly before the operation compressed work that would normally have taken much longer. More than 1,000 potential targets had to be assessed and approved in a matter of days.
This is where AI-assisted systems offered a major operational advantage: they could process and organise enormous quantities of information far faster than traditional teams working manually. But the speed also reduced the amount of time available for independent scrutiny of individual targets.
How the ‘kill chain’ worked?
The military term "kill chain" describes the sequence through which information becomes an authorised military action. In simplified form, it begins with intelligence, moves through target identification and assessment, proceeds to review and authorisation, and ends with the weapon being launched.
In the Minab case, the process reportedly began with the site's existing classification as an IRGC facility. That information entered the AI-assisted targeting architecture. Maven then helped integrate information and prioritise targets for the operation.
The proposed target subsequently passed through additional human reviews involving intelligence personnel, imagery specialists, targeteers, legal advisers and commanders. According to officials familiar with the Pentagon investigation, confidence in the site's military classification remained high as it moved through these stages.
The critical problem was therefore not necessarily the absence of human involvement. Rather, the same incorrect underlying assumption appears to have survived multiple stages of human and technological review.
Why human oversight did not catch the mistake?
Human involvement is often presented as the principal safeguard against errors made by AI-enabled military systems. In the Minab case, however, the humans reviewing the target were reportedly operating with the belief that the site was a legitimate military installation.
That distinction is important. Human oversight can prevent an AI system from making an autonomous decision, but it is less effective if the humans are presented with incorrect information and have no independent reason to challenge it.
Bloomberg reported that Pentagon investigators found high confidence in the military classification throughout the targeting chain. The issue was therefore not simply that there was insufficient information; rather, decision-makers apparently had considerable confidence in information that was wrong.
The civilian-harm safeguard?
Another issue identified in the reporting concerns the Pentagon's civilian-harm mitigation structure.
According to officials familiar with the investigation, staffing for civilian-harm mitigation teams had been substantially reduced, including at US Central Command. Bloomberg reported that the CENTCOM team had been reduced from 10 personnel to one and that no member of the team reviewed the Minab site before the strike.
Such teams are intended to provide an additional layer of scrutiny by assessing potential civilian risks and identifying precautions that might reduce them. Their absence from the Minab targeting process meant that another potential opportunity to challenge the site's classification was missing.
Why AI can amplify an intelligence error?
The Minab episode illustrates a broader problem with AI-assisted decision-making. Artificial intelligence can process information much faster than a human team, but speed does not automatically make the underlying information more accurate.
If an outdated intelligence record identifies a civilian site as military infrastructure, an AI-enabled system may process that information efficiently, connect it with other data and place the location into a target-selection workflow. The technology can therefore accelerate a mistaken conclusion without necessarily recognising that the original premise is wrong.
This is fundamentally different from saying that the AI "decided to bomb a school." The available evidence instead points to an interaction between flawed intelligence, automated data processing and human decision-making.
The role of Maven remains contested?
The precise contribution of Maven to the Minab strike remains an important question. The Washington Post reported that Maven was being used extensively in the Iran campaign and that AI tools helped generate and prioritise targets. Bloomberg's investigation found that some military personnel expected Maven to identify inconsistencies or outdated information in the intelligence underlying potential targets.
Palantir, the company behind Maven, has disputed the suggestion that its software was responsible for the intelligence failure. The company has said that it is not responsible for the underlying data or for identifying deficiencies in intelligence and that there is no evidence that its software was at fault in the Minab strike.
That distinction matters because the quality of an AI system's output depends heavily on the information and instructions provided to it. Determining responsibility therefore requires examining the entire chain rather than attributing the outcome to the software alone.
What changed after the strike?
According to Bloomberg, Palantir subsequently developed additional capabilities for Maven intended to re-examine underlying intelligence and identify information that could disqualify a target or expose inconsistencies that human reviewers might have missed.
The reported change reflects one of the central lessons emerging from the incident: AI systems used in targeting may need to do more than assemble information and recommend targets. They may also need mechanisms specifically designed to challenge the assumptions behind those recommendations.
What is the larger lesson?
The Minab strike demonstrates why the military use of AI cannot be assessed simply by asking whether a human ultimately authorised the attack. The more important question is how information reaches that human decision-maker and whether the system creates genuine opportunities to challenge an erroneous conclusion.
In this case, the reported chain contained several interconnected weaknesses: an old military classification, intelligence that was not adequately integrated into the main targeting database, an accelerated targeting timetable, extensive use of AI-enabled systems and the absence of a civilian-harm review before the strike.
The case therefore represents less a story of an AI system independently selecting a school and more a warning about how technology can make an existing intelligence error travel through a military decision-making system faster and with greater apparent confidence.
The Pentagon's full investigation remains important because it should clarify precisely where the decisive failures occurred and how much responsibility should be attributed to intelligence processes, database management, AI-assisted targeting, human review and command decisions. Until the complete findings are publicly released, the precise role