# The human was in the loop: what September revealed about AI in the kill chain

URL: https://www.thedeepfeed.ai/posts/2026-10-16-ai-in-the-kill-chain/
Category: Policy
Published: 2026-10-16
Author: the-deep-feed
Tags: policy, military-ai, pentagon, palantir, maven, accountability, iran, intelligence, automation-bias
Kind: deep

> September brought the first detailed accounts of AI contributing to US military harm, a school strike in Iran and a chatbot-written intel report, but they came through leaks, the Pentagon is holding its report past the midterms, and the rules still govern the wrong machine.

## TL;DR

- **No official admission exists.** The September accounts tying AI to the **Minab school strike** (more than 150 dead, most of them children) and to a false intelligence report came from anonymous officials via **Bloomberg** and **CNN**. The Pentagon's on-record position is still *under investigation*, and on Oct 15 Bloomberg reported it will hold the report until after the **Nov 3** midterms.
- Both incidents describe **decision support**, not autonomous weapons: a stale record that **Maven** turned into a day-one target in minutes, with **CENTCOM's** civilian-harm team cut from **10 people to one**, and an analyst's chatbot that fused open-source and signals intelligence into a report that nearly led troops to board a **Chinese ship**.
- A human was in the loop both times. The failure was **automation bias at speed**: users expected Maven to flag stale data, a check its maker says was never its job, and a second AI pass turned a wrong answer into a report in the format officers trust.
- The main rulebook misses this. **DoD Directive 3000.09** excludes *systems that are not weapon systems*; **NSPM-11** ordered it rewritten within 90 days without reaching decision support. The provisions that do (AI incident reporting, training to question or override outputs, named accountable humans) sit in the House's **FY27 NDAA**, which passed **216-212** and is stalled in the Senate.

On February 28, the first day of the war with Iran, Tomahawk missiles struck the Shajareh Tayyebeh elementary school in the southern Iranian city of Minab. The school sat next to a Revolutionary Guard naval compound. Counts of the dead vary by source: Bloomberg's reporting puts them at more than 150, including at least 123 children; [Iranian state media has said 168](https://www.npr.org/2026/09/17/nx-s1-5972934/us-%EC%9D%B4%EB%9E%80-war-un-report-alleged-war-crimes). In March, [121 House Democrats](https://ansari.house.gov/media/press-releases/03/12/2026/reps-ansari-jacobs-crow-lead-majority-of-house-democratic-caucus-demanding-answers-on-reported-us-strike-on-iranian-school) described it as what appeared to be the deadliest attack on civilians of the campaign. More than seven months later, the United States has not formally accepted responsibility for it.

In the week of September 17, the story changed shape. A UN fact-finding mission said there were reasonable grounds to believe the strike was a war crime. CNN reported that an analyst's chatbot had produced a false intelligence report that sent US forces toward boarding a Chinese ship. And Bloomberg, citing officials involved in an unreleased internal Pentagon review, reported that a "cascade of preventable failures" behind the Minab strike included an overreliance on Maven, the AI targeting platform built by Palantir. On September 22 the Bloomberg story reached the front page of Hacker News under the headline ["Pentagon says overreliance on AI contributed to missile strike on Iran school"](https://news.ycombinator.com/item?id=49806430), where it drew 974 points and 553 comments.

That headline is the first thing this piece has to correct. The Pentagon has said no such thing. Every account of AI's role in either incident comes from anonymous officials and sources describing documents the public has not seen. The Department's on-record position has been the same sentence since March: the incident is under investigation. On October 15, Bloomberg reported that senior officials intend to [hold its report until after the November 3 midterm elections](https://ph.investing.com/news/stock-market-news/hegseth-advisers-plan-to-delay-iran-school-attack-report-release--bloomberg-93CH-2631166).

The starting hypothesis for this piece was that September brought the first official admissions that AI contributed to military harm, and that they show over-reliance on decision support, rather than autonomous weapons, to be the failure mode policy has not caught up with. The first half does not survive the record. There are no admissions, only leaks, a refusal to publish and a corporate denial. The second half holds up better than expected, because the two incidents are different in nearly every operational detail and identical in one: a human was in the loop both times, and the loop was where it failed.

# What is on the record, and who is speaking

Before the argument, the inventory. Grave claims deserve a clear account of who is making them.

| Claim | Who says it | On the record? |
|---|---|---|
| A US munition hit the school | NBC's March sources; NYT preliminary-findings report cited by 121 House Democrats; UN mission | No US official statement accepting responsibility |
| Targeting data on the site was years out of date | CNN (July, three sources); Bloomberg (September, officials involved in the review) | No |
| Senior commanders bypassed database warnings that the data needed re-vetting | CNN (July) | No |
| Overreliance on Maven contributed | Bloomberg (September), citing the internal review | No; Palantir disputes it |
| Civilian-harm staff cut by about 90%, CENTCOM's team from 10 to one | CNN, Bloomberg | No Pentagon response to either outlet |
| An analyst's chatbot produced a false report on a Chinese ship's cargo | CNN (four sources) | No; SOCPAC and the Pentagon did not respond |
| CENTCOM changed its targeting process after Minab | Bloomberg (September 22), unnamed officials | No |
| The report will be held past the midterms | Bloomberg (October 15) | Pentagon: "no updates to announce at this time" |

Two things follow from that table. The first is that the reporting is consistent across outlets that had different sources at different times: CNN in July and Bloomberg in September describe the same stale record, the same compressed timeline and the same gutted review team. The second is that none of it has been tested against the document that would settle it. The internal review has, according to officials who spoke to Bloomberg, been all but finished for months.

# Minab: a kill chain that ran faster than its checks

Read together, the [CNN account from July](https://www.cnn.com/2026/07/07/politics/us-commanders-intelligence-iran-school) and [Bloomberg's September investigation](https://gizmodo.com/pentagon-investigators-say-overreliance-on-palantir-ai-tech-contributed-to-u-s-strike-that-killed-123-iranian-children-2000814477) describe a sequence with several failures and no single cause.

The record came first. US databases had listed the Minab compound as a military site for years. Commercial imagery told a different story: by CNN's account, a fence separating the school from the base appears in 2016 imagery along with a separate entrance, and in December 2025 imagery showed dozens of people apparently playing in the courtyard. Bloomberg's sources described walls and separate entrances complete by 2017, and a 2018 image showing painted walls, a soccer pitch and playground markings. The school had its own website and a Google Maps listing. An analyst had noticed changes at the site as early as 2019 and logged them, but in a tool that was not connected to the database targeteers draw on. The note never reached the people building strike packages.

Then came the clock. The administration wanted an overwhelming opening assault, and more than 1,000 targets were hit in the first 24 hours. CNN's sources said analysts scrambled to refresh records before the war began and triaged: mobile threats and "upper-tier" sites first, fixed sites later. The intelligence for many targets on the strike list was more than ten years old, including the record for the compound next to the school. Bloomberg's account adds the machine. The Minab site, still catalogued as a Revolutionary Guard facility, was fed into Maven alongside other candidates and came out as a recommended day-one target. Target-list work that once took hours was condensed into minutes.

Then the people who were supposed to catch it were mostly gone. Both outlets report that Hegseth's cuts reduced civilian-harm mitigation staffing by roughly 90% (Bloomberg puts the Department-wide total at fewer than 20 people) and that Central Command's team shrank from ten people to one. Bloomberg's sources said no member of those teams reviewed the Minab site before launch.

![Minab decision chain: stale record, Maven fusion, day-one target list, approval, and a civilian-harm review that never ran.](/post-images/2026-10-16-ai-in-the-kill-chain/minab-kill-chain.jpg)

The two outlets differ on one detail, and the difference matters for the argument. Bloomberg's sources said some users *expected Maven to flag stale records or contradictions* in the intelligence assembled for a target, and could not explain why they expected that. CNN's July sources described the opposite arrangement. Two targeting databases, the decades-old Modernized Integrated Database and its newer AI-powered replacement, the Machine-Assisted Analytic Rapid-Repository System (MARS), clearly marked the Iranian target data as needing an update before use, and adding such a site to the strike list required a senior officer's approval. Senior commanders approved anyway, two sources said, for "expediency."

These accounts can both be true, since they concern different systems at different stages. But they point to the same conclusion from opposite directions. In one telling, people trusted a tool to perform a check its maker says was never its job. In the other, a tool performed the check and people overrode it because the pace demanded it. Either way, the machine sat inside a process that had been tuned for speed and stripped of the human redundancy that would have caught the error. "The Pentagon is pressing everyone to move faster," one of CNN's sources said.

Palantir rejects the Bloomberg framing. A spokesperson told Bloomberg the company "is not responsible for the underlying data nor identifying intelligence deficiencies" and that there is no evidence its software was at fault; two people familiar with the contracts said the government holds primary responsibility for what goes into Maven. On CBS's *Face the Nation* on October 11, Palantir's chief technology officer, Shyam Sankar, was blunter: ["The Bloomberg reporting is wrong,"](https://aiweekly.co/alerts/palantirs-sankar-calls-bloomberg-strike-report-wrong) he said, adding that "the likely cause is human error." Wes Bryant, a former Pentagon official who led civilian harm assessments and advised on building those tools into Maven, told [Straight Arrow News](https://san.com/cc/staff-cuts-oversight-failure-to-blame-in-iran-school-strike-not-faulty-ai-targeting-ex-pentagon-official/) that staff cuts and oversight failures were the bigger cause, and that an open-source check AI could run in seconds would have flagged the school.

Notice what the defense concedes. If the likely cause is human error, the humans in question were working inside a workflow that turned hours into minutes, at an opening tempo of more than 1,000 targets in a day, with the review layer at Central Command reduced to one person. "Human error" and "overreliance on AI" are not competing explanations. They are the same explanation, described from the vendor's side and from the investigator's.

# The ship that wasn't carrying what the report said

The second incident is smaller in consequence and, in some ways, more alarming. [CNN reported on September 18](https://www.cnn.com/2026/09/18/politics/us-military-ai-false-intelligence-china-ship) that an intelligence report circulated across the military in the spring, during the war, warned that a Chinese ship in the Middle East was carrying components of a nuclear weapons program. According to four sources, the military moved to intercept it. Two said armed personnel were preparing to board. Aircraft were in the air.

Only shortly before the operation did officials look closely at how the report had been made. A special operations command analyst had asked a chatbot about reporting on the ship's manifest that originated with US Special Operations Command Pacific. The bot fused open-source intelligence with classified signals intelligence and reached a wrong conclusion about the cargo. The analyst then used AI a second time, to package the findings into a standard intelligence report, "the kind that is trusted by military officials," and disseminated it. One source called the report "entirely false" and said it "almost started a war." CNN could not learn what the cargo was, or whether the chatbot was a commercial product or a government one. "The internal tools are mostly just copies of the commercial stuff wearing lipstick," a former senior official said.

![Manifest, open-source and signals inputs feed an analyst's chatbot; a second AI pass formats a trusted report.](/post-images/2026-10-16-ai-in-the-kill-chain/report-laundering.jpg)

The second AI pass is the detail that deserves attention. A hallucinated conclusion in a chat window is an analyst's problem. The same conclusion rendered in the house format of a finished intelligence product becomes everyone's problem, because the format is itself a trust signal. Readers downstream had no ready way to see that the provenance was a model's synthesis rather than an analyst's judgment. The tool did not only produce an error. It produced the error in the form most likely to be acted on.

CNN's sources described this as a pattern rather than an isolated case. The hallucinations "have not been an isolated incident across the intelligence community," one said. Several described younger analysts as more likely to trust the tools uncritically, and older officials as worried that AI was pressuring analysts to produce and disseminate faster. Another source, familiar with current military policy, put the governance gap plainly: "AI in targeting is definitely something that is ramping up and there is no real guidance for how having a human in the loop will prevent civilian casualties or fratricide." The line that has travelled furthest from the story came from a third: "AI allows you to get to a bad idea faster."

# Same failure, twice

In March, as questions about Minab mounted, Central Command's commander, Adm. Brad Cooper, said in a video that ["humans will always make final decisions on what to shoot and what not to shoot and when to shoot."](https://www.nbcnews.com/politics/national-security/democrats-ask-pentagon-iran-school-strike-role-ai-rcna263083) On the evidence reported since, that was true. A human approved the Minab target. A human analyst produced and disseminated the ship report. The phrase "human in the loop" has carried most of the weight in public reassurance about military AI for a decade, and both September stories describe loops that contained humans and failed anyway.

The failure has a name in the human-factors literature: automation bias, the tendency to accept a system's output and stop looking for contrary evidence. It is not a new idea, and it does not require a sophisticated model. What the 2026 accounts add is scale and speed. Maven's own trajectory shows both. [CSIS describes the platform](https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do) as fusing drone video, radar, signals intelligence and documents into one map, with Central Command's 2024 deployment drawing on 179 data sources; in its design a human validates the labels and orders the strike. At a DefenseScoop event on September 22, the same day the Bloomberg story topped Hacker News, the deputy undersecretary for research and engineering, James Mazol, said Maven's user base had grown from about 50,000 in January to more than 100,000 after the Iran campaign began. The Pentagon's chief digital and AI officer, Cameron Stanley, said the capability helped the military strike ["13,000 targets in just 38 days."](https://defensescoop.com/2026/09/22/maven-smart-system-ai-james-mazol-cameron-stanley-defensetalks/) "That's data-centric warfare," he said. "That's using AI at scale."

Those numbers are the point. When a single human approval sits at the end of a pipeline producing targets at that rate, the approval becomes the step most exposed to automation bias, because it is the step with the least time per decision and the most confidence in what arrived. A data model with a stale record does not announce itself. A fluent intelligence report does not carry a watermark. In both incidents, the system's output looked like the product of the checks it had skipped.

![Ledger chart contrasting targeting throughput and Maven users against civilian-harm review staff during the 2026 Iran campaign.](/post-images/2026-10-16-ai-in-the-kill-chain/speed-versus-check.jpg)

The engineers and analysts arguing about the Bloomberg story on Hacker News split along the same line the officials did. One of the most-replied-to comments argued that AI was a scapegoat:

> Reading the details, "AI" doesn't really seem like the culprit -it's a scapegoat. The intelligence that it was no longer a military target never entered the target database, the team that was responsible for vetting the target list was gutted, and said team was never even consulted.
>
> — [legitster, Hacker News](https://news.ycombinator.com/item?id=49806777), Sep 22, 2026

A comment the next day made the more careful point, which is that the two readings are not exclusive:

> Two things can be true at once, the military can be incompetent, and AI can be used to make targeting decisions that are wrong - we can understand both the new world that we live in, and the absolution of responsibility it offers without ignoring these nuances.
>
> — [kamranjon, Hacker News](https://news.ycombinator.com/item?id=49816673), Sep 23, 2026

A third commenter read the headline as the Pentagon blaming Palantir and Palantir blaming the data, and asked ["where's the accountability by both organizations?"](https://news.ycombinator.com/item?id=49806897) The misreading is instructive. The Pentagon had not blamed anyone. But a public that hears about military AI failures only through leaks will fill the silence with the most legible story available, which is a vendor dispute. Lt. Gen. Karen Gibson, a former Central Command intelligence director, said in March that [responsibility would rest with "a commander somewhere"](https://www.defensenews.com/news/your-military/2026/03/24/deadly-iran-school-strike-casts-shadow-over-pentagons-ai-targeting-push/) rather than a machine or a software engineer. Seven months on, no commander has been named, and the document that would name one is on hold.

# The rulebook covers the wrong machine

The Pentagon does have a policy on AI and lethal force. It is DoD Directive 3000.09, *Autonomy in Weapon Systems*, last reissued in January 2023. Its central sentence is that autonomous and semi-autonomous weapon systems "will be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force." It defines an autonomous weapon system as one that, "once activated, can select and engage targets without further intervention by an operator."

Its applicability section is where the September stories fall out of scope. Among the things the directive [does not apply to](https://web.archive.org/web/2024id_/https://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodd/300009p.PDF), item seven reads: "Autonomous or semi-autonomous systems that are not weapon systems." Maven nominates targets and ranks options; it does not fire. MARS is a database. A chatbot that writes an intelligence report is not a weapon by any definition the directive uses. Every system implicated in the 2026 accounts sits outside the document the Pentagon cites when asked about human control of lethal AI.

The [Congressional Research Service's October 5 update](https://www.congress.gov/crs_external_products/IN/PDF/IN12669/IN12669.4.pdf) on the Pentagon's dispute with Anthropic makes the same gap visible from the other side. That fight was over Anthropic's refusal to allow two uses of Claude, fully autonomous weapons and mass domestic surveillance (this publication covers the litigation [separately](/posts/2026-10-14-the-assistant-files-a-report/)). CRS notes that the Department is not publicly known to be using Claude, or any other frontier model, inside autonomous weapon systems. The public argument about military AI in 2026 has been about the use case nobody has been shown to deploy. The harm reported this year came from the use case that is deployed everywhere.

![DoD Directive 3000.09 covers autonomous weapons; Maven, MARS and analyst chatbots sit outside its scope.](/post-images/2026-10-16-ai-in-the-kill-chain/rule-coverage-map.jpg)

The directive is being rewritten, but not in that direction. [NSPM-11](https://www.whitehouse.gov/presidential-actions/2026/06/national-security-presidential-memorandum-nspm-11/), signed June 5, ordered the Secretary of War to issue an update to 3000.09 within 90 days, to be reviewed annually "to account for the rapidly evolving capabilities of AI systems." The memorandum's first policy pillar instructs the national security enterprise to accelerate adoption by "eliminating unnecessary barriers to rapid deployment." Its accountability pillar concerns free speech, ideological bias, surveillance and civil liberties; its definition of an "AI incident" covers performance degradation, data loss, technical malfunction and adversarial attack. Neither reaches the failure in which the system performs as designed and a person trusts it too far.

Sen. Ruben Gallego saw the gap a week later. His [June 12 letter](https://www.gallego.senate.gov/wp-content/uploads/2026/06/Final_Gallego-Letter_Directive-3000.09.pdf) asked how the updated directive would prevent friendly fire from autonomous weapons, semi-autonomous weapons "or targeting recommendations produced by AI-enabled decision support systems," and what civilian-harm capacity had been assigned to decision support. He asked for answers by June 26. [DefenseScoop reported](https://defensescoop.com/2026/06/15/lawmaker-questions-pentagons-plan-to-revise-autonomous-weapons-policy/) that the Department had been largely unforthcoming about how it applies the directive. The 90-day deadline passed in early September. No revised text has been published in any source this publication could find, and CRS's October update still describes the 2023 version.

The Department's own strategy explains the direction of travel. Secretary Pete Hegseth's [January 9 AI strategy memo](https://www.afacpo.com/mpm/wp-content/uploads/2026/01/ARTIFICIAL-INTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.pdf) is explicit about its priorities. One of its seven "pace-setting projects," called Agent Network, is "AI agent development and experimentation for AI-enabled battle management and decision support, from campaign planning to kill chain execution." Under the heading "Speed Wins," it states: "We must accept that the risks of not moving fast enough outweigh the risks of imperfect alignment." A section titled "Out with Utopian Idealism, In with Hard-Nosed Realism" redefines responsible AI around model "objectivity" and "any lawful use" contract language. Two months later, Deputy Secretary Steve Feinberg's March 9 memo, [reported by DefenseScoop](https://defensescoop.com/2026/04/03/palantir-maven-feinberg-directive/), set out to "establish AI-enabled decision-making" as the center of the department's strategy for joint command and control, and directed that Maven become a formal program of record by the end of fiscal 2026.

None of this is hidden. The policy record says, in its own words, that speed is the priority, decision support is the vehicle and the binding human-control rule applies to weapons. The September accounts describe what that combination produced.

# What changed after September

Three things moved after the stories broke, and none of them is an accounting.

The first is operational. On September 22, Bloomberg reported, again citing unnamed officials, that Central Command had [changed its targeting process](https://aiweekly.co/alerts/pentagon-rewires-palantir-maven-after-iran-school-strike) in three areas: how planned targets are refreshed and vetted, new open-source feeds to track civilian movement, and dozens of upgrades to Maven, including features that re-review underlying intelligence for factors that would disqualify a target. Bloomberg's earlier story said Palantir added such checks after the strike, and a person familiar with the work said they had already caught some anomalies. These may be sound changes. They are also, in structure, more automation layered over a process whose failure was over-trust in automation, and they arrived without a published account of what the first layer got wrong.

The second is institutional, and it runs the other way. In a [September 30 address at Quantico](https://breakingdefense.com/2026/09/hegseth-announces-plans-for-new-4-star-drone-command-further-general-officer-cuts/), Hegseth announced a new four-star Autonomous Warfare Command, targeted for October 2027 and designed to hold its own acquisition authority, which would require legislation. ["The pace of war is changing faster than the processes to support it,"](https://theintercept.com/2026/10/08/hegseth-military-ai-drones-autonomous-war/) he said. Critics quoted by The Intercept argued the command's design is meant to sidestep oversight.

The third is the report itself. On October 15, Bloomberg reported that Hegseth's advisers had told Pentagon personnel the Minab findings would be held until after the midterms, with a Pentagon official saying only that the investigation is ongoing, that there were "no updates to announce at this time," and that no final decision on timing had been made. Sen. Kirsten Gillibrand, who [led 24 colleagues in a July 13 letter](https://www.gillibrand.senate.gov/news/press/release/gillibrand-colleagues-demand-pentagon-release-long-awaited-report-on-u-s-bombing-of-iranian-elementary-school/) noting the investigation had reportedly been submitted in April, called the delay ["not only disgraceful, it is cowardly"](https://x.com/DropSiteNews/status/2111209317424578628) in a statement. Rep. Sara Jacobs, who co-led the House letter in March, responded the next day:

> The Pentagon is trying to avoid accountability, but I won't let them. I passed an amendment through the House to require the Pentagon to share its report of the Iran school strike with Congress and the public – and I will keep working until it's law.
>
> — [@RepSaraJacobs](https://x.com/RepSaraJacobs/status/2111245033680556423), Oct 16, 2026

![Timeline from Feb 28 to Oct 15: congressional letters, NSPM-11, UN findings, the September leaks, and the report held past the midterms.](/post-images/2026-10-16-ai-in-the-kill-chain/accountability-timeline.jpg)

That amendment is the most specific response to the failure mode on any public record, and its fate is the clearest measure of where accountability stands. Jacobs's office lists the provisions it secured in the House Armed Services Committee's [FY2027 defense authorization bill](https://sarajacobs.house.gov/news/press-releases/rep-sara-jacobs-votes-against-ndaa-due-to-lack-of-constraints-on-weaponization-and-politicization-of-military). One requires the Pentagon to share its Minab investigation with Congress and publish an unclassified version. Others read like a direct answer to September: policies that identify "the human commanders or operators responsible for authorizing, supervising, and terminating" AI-enabled use of force; a Department-wide AI incident reporting program covering "risks arising from human-machine teaming"; and training "to help operators appropriately rely on, question, or override AI outputs."

The House passed its bill [216-212 on July 22](https://breakingdefense.com/2026/07/house-passes-2027-ndaa-while-measure-remains-stalled-in-the-senate/). Senate Democrats have blocked the Senate version from reaching the floor, citing opposition to the war with Iran and the scale of the spending increase, and no final bill has been reported enacted. The House [went home in September](https://www.cnn.com/2026/09/18/politics/congress-ai-inaction) and will not return until after the election. The provisions that would turn automation bias into a reportable, trainable, attributable failure exist only as text in a stalled bill.

# The unanswered file

Strip away the leaks and the denials, and the accountability record for AI in this year's kill chain is a list of questions with no published answers.

| Date | Who asked | What they asked for | Public answer |
|---|---|---|---|
| Mar 12 | [121 House Democrats](https://ansari.house.gov/media/press-releases/03/12/2026/reps-ansari-jacobs-crow-lead-majority-of-house-democratic-caucus-demanding-answers-on-reported-us-strike-on-iranian-school) | Whether AI, including Maven, identified the school; whether a human verified it; release of findings (due Mar 20) | "The incident is under investigation" |
| March | [46 Senate Democrats](https://www.defensenews.com/news/your-military/2026/03/24/deadly-iran-school-strike-casts-shadow-over-pentagons-ai-targeting-push/) | Whether AI, including Maven, identified the school | None published |
| Jun 12 | Sen. Gallego | Safeguards for AI decision support in the 3000.09 rewrite (due Jun 26) | None reported |
| Jun | UN fact-finding mission | Information and meetings with US officials | Unanswered, per the mission |
| Jul 13 | Sen. Gillibrand and 24 colleagues | Release of the report, reportedly submitted in April, and a plan to prevent recurrence | Report held |
| Sep 19 | Sens. Warner, Reed and Coons | Inspector General investigations of "both instances this year" of AI-enabled targeting errors | None reported |
| Oct 15 | Bloomberg | Release timeline | "No updates to announce at this time" |

The March 12 House letter asked the narrowest version of the right question: ["If artificial intelligence is used, is it subject to human review and at what point?"](https://www.nbcnews.com/politics/national-security/democrats-ask-pentagon-iran-school-strike-role-ai-rcna263083) The September 19 letter from Sens. Mark Warner, Jack Reed and Chris Coons asked the broadest. It requested an "immediate investigation by relevant Inspectors General, with a higher degree of public transparency," and [warned that](https://www.cnn.com/2026/09/19/politics/democrats-letter-ai-investigation-military) "public reports that AI targeting platforms have repeatedly generated spurious outputs (based on outdated input data or outright hallucination) risk serious operational error." It also asked for inspectors general to see "any additional instances that may have so far not been publicly reported." CNN's sources had already said there were some.

The UN mission's findings, released September 17, were the only formal findings an official body has published about Minab, and they did not turn on the software. The experts found that the US failed in its obligation to do everything feasible to verify that the school was a military objective, a failure they said went beyond mere negligence, and concluded that ["the building of the school was the intended point of impact."](https://www.npr.org/2026/09/17/nx-s1-5972934/us-%EC%9D%B4%EB%9E%80-war-un-report-alleged-war-crimes) A White House spokeswoman said the Human Rights Council had spent decades "spouting unserious nonsense." Neither the mission's conclusions nor the White House's dismissal engaged with what the Pentagon's own review is reported to say about Maven, because nobody outside the building has read it.

# Speed was the specification

The honest summary of September is narrower than the headlines. No official admitted anything. The role of AI in the Minab strike is disputed by the vendor, described differently by different sets of sources, and entangled with stale data, compressed timelines and a review team cut to a single person. The ship incident rests on four anonymous sources and an unidentified chatbot. Readers should hold all of it at the confidence anonymous sourcing deserves until the report is published.

What the record does establish is the shape of the failure. In neither incident did a machine decide to use force. In both, a person approved a recommendation that arrived faster, looked more finished and carried more apparent authority than the checks behind it justified. The governing rule on AI and lethal force does not cover the systems involved. The presidential memorandum ordering that rule rewritten says nothing about them. The strategy that put them at the center of the targeting cycle says, in writing, that moving too slowly is the greater risk. The only legislative text that names the problem is stalled in the Senate.

"Human in the loop" was meant to be the safeguard. The 2026 record suggests it has become the place where accountability goes to wait: present at every decision, responsible for each one, and given less time, less staff and more confident machinery with every increase in tempo. The Pentagon has the document that would show whether that is what happened at Minab. By Bloomberg's account, its senior officials intend to let the public read it after the votes are counted.

## Sources

- [CNN — Exclusive: US military had close call after using AI for false intelligence report, sources say (Sep 18, 2026)](https://www.cnn.com/2026/09/18/politics/us-military-ai-false-intelligence-china-ship)
- [CNN — Democrats call for investigation into faulty AI-assisted intel report (Sep 19, 2026)](https://www.cnn.com/2026/09/19/politics/democrats-letter-ai-investigation-military)
- [CNN — US commanders bypassed warnings that intelligence on Iran targets was out of date (Jul 7, 2026)](https://www.cnn.com/2026/07/07/politics/us-commanders-intelligence-iran-school)
- [Gizmodo — Pentagon investigators say overreliance on Palantir AI tech contributed to US strike (Sep 20, 2026), summarizing Bloomberg's Sep 18 investigation](https://gizmodo.com/pentagon-investigators-say-overreliance-on-palantir-ai-tech-contributed-to-u-s-strike-that-killed-123-iranian-children-2000814477)
- [AI Weekly — Pentagon rewires Palantir Maven after Iran school strike (Sep 22, 2026), summarizing Bloomberg](https://aiweekly.co/alerts/pentagon-rewires-palantir-maven-after-iran-school-strike)
- [Investing.com — Hegseth advisers plan to delay Iran school attack report release, Bloomberg (Oct 15, 2026)](https://ph.investing.com/news/stock-market-news/hegseth-advisers-plan-to-delay-iran-school-attack-report-release--bloomberg-93CH-2631166)
- [NPR — U.N.-backed experts cite possible U.S. war crimes in Iran, including a school attack (Sep 17, 2026)](https://www.npr.org/2026/09/17/nx-s1-5972934/us-%EC%9D%B4%EB%9E%80-war-un-report-alleged-war-crimes)
- [DoD Directive 3000.09, Autonomy in Weapon Systems (Jan 25, 2023), archived copy](https://web.archive.org/web/2024id_/https://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodd/300009p.PDF)
- [The White House — National Security Presidential Memorandum NSPM-11 (Jun 5, 2026)](https://www.whitehouse.gov/presidential-actions/2026/06/national-security-presidential-memorandum-nspm-11/)
- [Secretary of War — Artificial Intelligence Strategy for the Department of War memo (Jan 9, 2026), mirror copy](https://www.afacpo.com/mpm/wp-content/uploads/2026/01/ARTIFICIAL-INTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.pdf)
- [Congressional Research Service — Pentagon-Anthropic Dispute over Autonomous Weapon Systems, IN12669 (updated Oct 5, 2026)](https://www.congress.gov/crs_external_products/IN/PDF/IN12669/IN12669.4.pdf)
- [Sen. Ruben Gallego — Letter to Secretary Hegseth on Directive 3000.09 (Jun 12, 2026)](https://www.gallego.senate.gov/wp-content/uploads/2026/06/Final_Gallego-Letter_Directive-3000.09.pdf)
- [Rep. Yassamin Ansari — 121 House Democrats demand answers on reported US strike on Iranian school (Mar 12, 2026)](https://ansari.house.gov/media/press-releases/03/12/2026/reps-ansari-jacobs-crow-lead-majority-of-house-democratic-caucus-demanding-answers-on-reported-us-strike-on-iranian-school)
- [Rep. Sara Jacobs — FY27 NDAA provisions, including Minab report and AI accountability (Jun 2026)](https://sarajacobs.house.gov/news/press-releases/rep-sara-jacobs-votes-against-ndaa-due-to-lack-of-constraints-on-weaponization-and-politicization-of-military)
- [NBC News — Democrats ask Pentagon about Iran school strike and role of AI (Mar 12, 2026)](https://www.nbcnews.com/politics/national-security/democrats-ask-pentagon-iran-school-strike-role-ai-rcna263083)
- [DefenseScoop — More than 100K personnel use Maven Smart System: Pentagon official (Sep 22, 2026)](https://defensescoop.com/2026/09/22/maven-smart-system-ai-james-mazol-cameron-stanley-defensetalks/)
- [DefenseScoop — Feinberg's new Maven directive sets AI-enabled decision-making as the cornerstone for CJADC2 (Apr 3, 2026)](https://defensescoop.com/2026/04/03/palantir-maven-feinberg-directive/)
- [DefenseScoop — Senator questions Pentagon's plan to revise autonomous weapons policy (Jun 15, 2026)](https://defensescoop.com/2026/06/15/lawmaker-questions-pentagons-plan-to-revise-autonomous-weapons-policy/)
- [CSIS — What Is Maven Smart System, and What Does It Do? (Jun 2, 2026)](https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do)
- [Straight Arrow News — Staff cuts, oversight failure to blame in Iran school strike, not faulty AI targeting: ex-Pentagon official (Oct 16, 2026)](https://san.com/cc/staff-cuts-oversight-failure-to-blame-in-iran-school-strike-not-faulty-ai-targeting-ex-pentagon-official/)
- [AI Weekly — Palantir's Sankar calls Bloomberg strike report wrong (Oct 12, 2026)](https://aiweekly.co/alerts/palantirs-sankar-calls-bloomberg-strike-report-wrong)
- [Defense News — Deadly Iran school strike casts shadow over Pentagon's AI targeting push (Mar 24, 2026)](https://www.defensenews.com/news/your-military/2026/03/24/deadly-iran-school-strike-casts-shadow-over-pentagons-ai-targeting-push/)
- [Breaking Defense — House passes 2027 NDAA while measure remains stalled in the Senate (Jul 22, 2026)](https://breakingdefense.com/2026/07/house-passes-2027-ndaa-while-measure-remains-stalled-in-the-senate/)
- [The Intercept — Pete Hegseth's military of the future: defined by AI and an utter lack of oversight (Oct 8, 2026)](https://theintercept.com/2026/10/08/hegseth-military-ai-drones-autonomous-war/)
- [Hacker News — Pentagon says overreliance on AI contributed to missile strike on Iran school (Sep 22, 2026)](https://news.ycombinator.com/item?id=49806430)
- [Sen. Kirsten Gillibrand — Gillibrand, colleagues demand Pentagon release long-awaited report on US bombing of Iranian elementary school (Jul 13, 2026)](https://www.gillibrand.senate.gov/news/press/release/gillibrand-colleagues-demand-pentagon-release-long-awaited-report-on-u-s-bombing-of-iranian-elementary-school/)
- [Breaking Defense — Hegseth announces plans for new 4-star drone command (Sep 30, 2026)](https://breakingdefense.com/2026/09/hegseth-announces-plans-for-new-4-star-drone-command-further-general-officer-cuts/)
- [CNN — Congress remains stuck on AI legislation (Sep 18, 2026)](https://www.cnn.com/2026/09/18/politics/congress-ai-inaction)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-16-ai-in-the-kill-chain/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt