On February 28 2026, the US and Israel launched an unprovoked attack on Iran with more than 1,000 targets hit by bombers and cruise missiles in the first 24 hours. Within hours of the first strike there were media reports that a primary school in Minab, southern Iran, had been hit with reports of significant number of causalities including children.
In fact at least 157 people were killed in the US Tomahawk missiles strike on the Shajara Tayyiba school, the vast majority of them children (more than 100 under the age of 12) with more than 100 others also injured (see Airwars detailed report). In March, the Pentagon announced that it was undertaking an official investigation into the strike but more than six months later, no report has been forthcoming.
This week, however, a detailed investigation by Bloomberg News, written by Ben Bartenstein and Krishna Karra, was published, with anonymous contributions from US officials said to be involved in the Pentagon investigation. While the official report is still to be released, its clear that there are serious lessons here for those advocating the increased use of AI to speed up targeting.
Epic Fury Foolishness
Following Israeli and US strikes on Iranian nuclear facilities in June 2025 and a build up of US forces in the region, from early 2026 Israel began lobbying the US to launch airstrikes on Iran aimed at so-called regime change, with senior Iranian leaders being key targets. Nationwide protest in late December and early January in Iran had been met with ruthless force by Iranian security, with reports of thousands killed. Trump told Iranians he would ‘make Iran great again’.
Due to the involvement of IRGC naval units in the dispute around the Straits of Hormuz, it is highly likely that Iranian naval facilities were also among the hundreds of targets fed into the Maven Smart System (MSS) – the Palantir-developed US AI-enabled targeting and battlefield management system – for strikes on the first day.
Maven uses AI and machine learning tools (such as Claude) to rapidly process data and ‘support’ commanders. It combines satellite imagery, drone feeds, signals intelligence and a myriad of other information into a single interface, then generates targets and recommends weapons systems. According to Military Times, it also drafts automated legal justifications for each strike.
Epic Fury, the Pentagon name for the campaign, was the first major AI-driven war according to Palantir’s Chief Technology Officer, Shyam Sankar:
“People will reflect back and say this is the first large-scale combat operation that was really driven, enhanced and made substantially more productive with technology, with AI.”
Blame Game
The Shajareh Tayyebeh primary school in Minab was on the edge of an Islamic Revolutionary Guards Corps (IRGC) naval medical facility. While the school building and grounds had once formed part of the military facility, it had been walled off at least nine years before the strike, and according to satellite images, had brightly painted walls and school sports facilities visible clearly visible.
Almost immediately after the strike, commentators began to ponder whether the use of AI to rapidly develop such large number of targets contributed to the deaths. However, equally quickly, advocates of the use of AI pushed back insisting that as humans had fed data into the Maven, humans not AI were to blame. This blame game is likely to occur each time an incident such as this occurs.
Militaries generally maintain a library of potential targets – military bases, missile sites, airfields, munition factories, etc. – in anticipation of armed conflict but obviously over time, those designated targets can change. While trying to keep potential targets updated, planners always imagine they will have time before a conflict to check and recheck the targets. However, in this case, as Bloomberg’s investigation notes:
“Work that had unfolded in previous conflicts over weeks or months — matching sites from the library, such as buildings, airfields or bases — was completed in a matter of days, according to officials involved in the probe. More than a thousand potential targets had to be reassessed and approved.”
Advocates of the use of AI in military targeting insist that using such systems can speed up ‘the kill chain’ and there is likely to have been pressure from the internal Pentagon pro-AI lobby to demonstrate the speed with which large numbers of targets could be generated.
Human Machine Interface: Who’s in control?
Palantir has repeatedly said that its software does not make lethal decisions and humans remain responsible for selecting and approving targets. However, numerous reports have detailed how humans overly rely and trust in computers and AI.
It is clear that such automation bias is present in the military use of AI. Bloomberg reports the officials they spoke to, who were involved in the investigation of the school strike, as saying
Some Centcom personnel expected Maven to flag stale information or inconsistencies in the underlying intelligence assembled for potential targets
Separately, the article states:
Two people familiar with the firm’s Pentagon contracts said the government retains primary responsibility for the quality of the information that’s fed into Maven. But they said it’s not uncommon for personnel who use tools such as Maven to develop operational understandings that differ from contractual terms.
The officials Bloomberg spoke to went:
Inside Centcom, which conducted the US attack, some personnel relied too much on the artificial intelligence embedded in Maven Smart System.
Since the Minab strike, Bloomberg reports “Palantir built new capabilities into Maven that ‘re-review underlying intelligence to identify factors that would disqualify a target and flag inconsistencies and inaccuracies that human review may have missed’.
UK pushes ahead with AI targeting
Over the past two years the UK has been pressing ahead with developing its own AI-enabled targeting and battlefield command system, currently designated the ‘Digital Targeting Web’. Giving evidence to the Defence Select Committee in July 2026, retired General Sir Richard Barrons, one of the authors of the recent Strategic Defence Review said:
“Essentially, what the web does is to take in every form of thing or capability that senses the operating environment—whether it is satellite, aircraft, drone, mobile phone, ship, vehicle or database, it does not matter; everything that allows you to understand your operating environment—and networks that into a database and secure cloud managed by AI—without AI, you cannot manage it—supervised by clever people. That capability allows you to assemble an understanding of your operating environment to support decision making, including ranking targets and allocating them to weapons systems, whether they are kinetic, from any service, or non-kinetic, like a cyber-attack.
The Defence Investment Plan (DIP), published in June 2026, re-stated the 2025 Strategic Defence Review’s embrace of drones and AI. Funding is being cut from non-military and military budgets to enable this to happen, with a whole variety of UK military AI projects being developed.
A new MoD taskforce – Taskforce RAID (‘Rapid AI Development’) – has been allocated £100m to speed up the military’s adoption of AI, with ministers giving the taskforce “exemptions from standard financial and procedural controls”. At the same time, the MoD has replaced its AI Ethics Advisory panel with a new AI Expert Advisory Group. Experts apparently being preferred to ethics.
Will lessons be learned?
Britain is pushing rapidly ahead with integrating AI and machine learning tools into is military targeting processes, with the FT even declaring over the summer ‘UK military looks at allowing lethal strikes without human approval‘. There is a real danger that ministers and military officials, dazzled by the possibilities that AI companies set before them, will give primacy to speed over care.
The British public however, wants and expects that its military will, at the very least, stick to the long-established global rules around the conduct of warfare – international law and in particular, International Humanitarian Law. IHL mandates that war must be a last resort and that that any use of force must ensure that civilians are protected and that a clear distinction is made between military and civilian targets.
Yet even that most pro-military voices accept that civilians are regularly and repeatedly killed in aerial bombing. The public are continually told in the aftermath of such strikes that they are an ‘aberration’, or ‘honest’ and ‘genuine’ mistakes. But when ‘mistakes’ repeatedly happen, not to see the systemic failure is simply to be in denial.
Now, when AI is already enabling 1,000 targets to be identified and struck in a single day, even a 1% ‘failure rate’ could result in 10 civilian casualties incidents a day. The way that AI is both increasing the number of targets and enabling a huge increase in the number of strikes will simply and inevitably lead to more civilian casualties.
As we approach an important moment in efforts to establish a treaty on the development and use of lethal autonomous weapons – the November 2026 Review Meeting of the Convention on Certain Conventional Weapons (CCW) – signs are not good. Reports from the final preparatory meeting in Geneva detail how the larger military states watered down language and removed key elements from a draft document.
The UK should be standing up for adherence to international law and primacy of the protection of civilians at the heart of its defence and military policies. Yet it should be remembered, that on the very same day that news began to circulate about the awful strike on the Minab school, the UK gave permission to the US to use the USAF base at Fairford in Gloucestershire for strikes on Iran, a military attack which international law experts describe as “a manifest violation of the United Nations Charter.“
While some argue that miliary AI systems can somehow be made to be ‘ethical’ the real lesson to be learned from Minab is that we are on the precipice of great danger. The rush to develop and deploy AI systems in warfare has to be resisted.
In other respects the Lords have challenged MoD’s approach more substantially, and in such cases these challenges are rejected in the government response. This is so in relation to the Lords’ recommendation that the government should adopt a definition for autonomous weapons systems (AWS). The section of the response dealing with this point lays bare the fact that the government’s priority “is to maximise our military capability in the face of growing threats”. A rather unconvincing assertion that “the irresponsible and unethical behaviours and outcomes about which the Committee is rightly concerned are already prohibited under existing legal mechanisms” is followed by the real reason for the government’s opposition: “there is a strong tendency in the ongoing debate about autonomous weapons to assert that any official AWS definition should serve as the starting point for a new legal instrument prohibiting certain types of systems”. Any international treaty which would outlaw autonomous weapon systems “represents a threat to UK Defence interests” the government argues. The argument ends with a side-swipe at Russia and an attempt to shut down further debate by claiming that the debate is taking place “at the worst possible time, given Russia’s action in Ukraine and a general increase in bellicosity from potential adversaries.” This basically seems to be saying that in adopting a definition for autonomous weapon systems the UK would be making itself more vulnerable to Russian military action. Really? 