A human commander approved every strike in Operation Epic Fury, the US campaign against Iran from late February to early May.
By the book, that satisfied the standard the Pentagon treats as sufficient for autonomous weapons: a person stays on the loop and signs off on each engagement. But at the campaign’s peak, American forces were striking hundreds of targets a day, on the order of one every few minutes, around the clock.
No human can independently classify a target, weigh proportionality, and estimate civilian harm in that window, not once every few minutes for days on end. Artificial intelligence did that work. The commander approved the result.
Authority in Name Only
That is the lesson of Epic Fury that most after-action assessments have skipped. They asked whether the campaign met its objectives. The harder question is whether the chain of command still functioned at machine speed, or whether the authority assigned to a human on paper had quietly migrated into the software.
This is not an indictment of any commander or administration. The dynamic is structural. Once a targeting pipeline runs faster than a person can verify what it produces, the real decision shifts to the system that produces it, whatever the doctrine says about who is in charge.
Consider how a single strike worked. AI pulled in intelligence feeds, identified a candidate target, judged it a lawful military objective, estimated collateral damage, assigned a confidence score, and generated a recommendation. The commander reviewed that recommendation and authorized the strike.
Directive 3000.09, last revised in 2023, calls this human-on-the-loop control and treats it as adequate. But the commander did not independently reproduce any of those judgments. He read what the machine produced and approved it. At Epic Fury’s tempo, that was the only workflow available.
The result is a gap between authority on paper and authority in practice. The directive was written for a commander who approves one engagement and has time to think. It does not address a commander approving hundreds a day, each in minutes.
CENTCOM’s commander described the change as a breakthrough: AI turned work that once took hours or days into seconds. He meant it as praise, though it also describes the problem. Speed is not the same as oversight: a six-hour targeting cycle leaves room for review; a six-second cycle does not.
When the Systems Fail
The cost became clear on the first day. A Tomahawk missile struck the Shajareh Tayyebeh elementary school in Minab. Amnesty International found that at least 156 people were killed, more than 120 of them children.
The preliminary US military investigation found that Central Command built the target coordinates from outdated Defense Intelligence Agency data, dating to when the site was part of an adjacent Revolutionary Guard base. By the time of the strike, it was a school.
The question investigators pressed was whether a human had verified the target. Nothing in the process flagged that the data was stale before the package was approved.
What followed matters as much as the strike. International condemnation, diplomatic pressure, a Senate investigation request. Tempo was reined in, civilian screening was tightened, and the campaign ended weeks later in a ceasefire.
Read generously, that is a system correcting itself. Read honestly, it is a system corrected from the outside.
Nothing in the AI architecture caught Minab, prevented it, or proposed a fix. The correction came from politics. A governance model that depends on a catastrophe being visible and embarrassing enough to force change is not a governance model. It is a hope. This time the pressure came. Next time it may not.
Engineering Oversight In
The answer is not to pull AI out of targeting, which gave US forces an edge against a mobile Iranian missile force. It is to stop treating human approval as a safeguard the human cannot actually provide, and to engineer the safeguards into the system.
Five steps follow directly.
Automated currency checks. The Pentagon should require an automated currency check on the intelligence behind every strike package, so that data older than a set threshold is flagged or held before a recommendation reaches a human. No one reviewing packages at Epic Fury’s pace could catch what happened at Minab.
Circuit breakers inside the software. Acquisition and program offices should require automatic pauses inside AI targeting systems when anomalies appear. A flag should trigger a halt and human escalation, not one more approval. Recovery should not wait for the evening news.
Tempo tied to verification. The Joint Staff should set a verification standard and link operational tempo to it. If human review cannot keep up, either the pace comes down or the review is replaced by constraints built into the system. Epic Fury chose neither and deferred the question.
An updated directive. The Office of the Secretary of Defense should revise Directive 3000.09 for mass-engagement scenarios, specifying what human-on-the-loop control requires when a commander cannot independently verify each strike. The 2023 text assumes a tempo the last campaign already exceeded.
Congressional visibility. Congress should require, for any AI-enabled operation, reporting on how much time commanders actually had to review each engagement and what they independently verified. Oversight cannot judge human control it cannot see.
What the Record Demands
The next AI-enabled campaign is coming, perhaps in the Pacific, where US forces have been building AI-assisted targeting for years. The pressures Epic Fury exposed will travel with it.
The Pentagon now holds an unusually detailed record of how AI targeting performs under operational pressure. It can use that record to build oversight that is real — or it can keep signing the screen and discover, too late, that the authority was never there to begin with.
Burak Oktenli holds an MBA and is pursuing a Master of Professional Studies in Applied Intelligence at Georgetown University, where his research focuses on the governance of autonomous and AI-enabled military systems.
The views and opinions expressed here are those of the author and do not necessarily reflect the editorial position of The Defense Post.
The Defense Post aims to publish a wide range of high-quality opinion and analysis from a diverse array of people – do you want to send us yours? Click here to submit an op-ed.
