The Demo
On March 13, at Palantir's AIPCON conference, Pentagon Chief Digital and AI Officer Cameron Stanley gave a live demonstration of the Maven Smart System. The audience watched a real-time map of Iran, studded with red targeting icons, some labeled "HQ." One of those locations corresponded to the Minab area.
Stanley's pitch was efficiency. Maven consolidated what used to require eight or nine separate systems into a single platform. Operations that once demanded approximately 2,000 intelligence officers for targeting work now require approximately 20. A 99% reduction in human analysts, executed in rapid succession.
"We've gone from identifying the target to now coming up with a course of action, to now actioning that target, all from one system," Stanley told reporters. "This is revolutionary."
In the audience, Palantir CEO Alex Karp put it more plainly: "No fair fights. If I can avoid it, let's not have fair fights. Our guys win and we come home."
The numbers back up the speed claim. During Operation Epic Fury, Maven helped execute 900 strikes in 12 hours, a pace that previously took weeks. The system generated approximately 1,000 targets in the first 24 hours, more than double the air power deployed during the entire opening phase of the 2003 Iraq invasion.
By every metric the Pentagon chose to measure, Maven is working.
The problem is the metric.
What 2,000 Analysts Actually Did
The 2,000-to-20 figure is being reported across the defense press as a triumph of automation. Almost no one is asking what those 1,980 eliminated analysts were actually doing.
They weren't just pushing data between systems. Intelligence targeting is a layered process involving collection, collation, validation, contextualization, and proportionality assessment. Analysts cross-reference signals intelligence with human intelligence. They verify that a target identified by satellite imagery still serves its original function. They assess whether a military target is proximate to civilian infrastructure. They evaluate whether the expected military advantage of a strike is proportional to the anticipated civilian harm.
This isn't bureaucratic overhead. Under international humanitarian law, specifically Additional Protocol I to the Geneva Conventions, parties to a conflict are required to take "constant care" to spare civilians. Commanders must do "everything feasible to verify" that targets are military objectives. They must assess proportionality. They must cancel or suspend an attack if it becomes apparent that the target is not military or that civilian damage would be excessive.
Those legal obligations don't become optional because you built a faster system. But when 20 people are doing the work of 2,000, in "rapid succession," with targets being identified, actioned, and struck from a single interface, the time available for each verification step compresses to near zero. The legal framework assumes deliberation. The system is designed to eliminate it.
Minab
On February 28, a Tomahawk cruise missile struck the Shajareh Tayyebeh primary school in Minab, Iran. 165 people were killed, including 110 schoolgirls. The school had previously been used as a military installation. The intelligence designating it as a military target appears to have been outdated.
Over 120 Democratic members of Congress are now demanding answers about the role of AI in target selection for that strike. The Pentagon has not confirmed or denied whether Maven generated the target. But the timeline is damning: Maven was actively producing target lists for Epic Fury, the school appeared on a map displayed to reporters on March 11, and the intelligence underlying the target was stale.
This is not a failure of AI capability. It is a failure of the system working exactly as designed.
Maven processes intelligence inputs and generates targeting recommendations at machine speed. If those inputs contain outdated data, a building that was a military base in 2024 but is a functioning school in 2026, the system processes it at the same speed as everything else. The consolidation that Stanley celebrated, eight or nine systems reduced to one, means there are fewer independent verification points where someone might catch the discrepancy. The personnel reduction, 2,000 analysts down to 20, means there are fewer human beings with contextual knowledge who might recognize that a location has changed use.
The speed that makes the system "revolutionary" is the same speed that prevents the error from being caught.
The Metric Nobody's Measuring
I spent eight years in Navy EOD. Every operation involved a decision with lethal consequences, either for the disposal team or for civilians in the area. The imperative was always speed; there is no safe amount of time to spend near a live explosive device. But speed was never the goal. Accuracy was the goal. Speed was the constraint you managed so that accuracy didn't suffer.
We had a principle: slow is smooth, smooth is fast. It sounds like a cliché until you're standing over an IED and the pressure to rush is physically overwhelming. The principle exists because haste in lethal operations produces exactly one outcome: mistakes that kill people who shouldn't die.
The Pentagon is measuring Maven by decisions-per-second. Targets generated. Strikes executed. Time compressed. Human analysts eliminated. By those metrics, Minab doesn't register as a system failure. The target was in the system. The strike was executed. The loop closed.
The metric that actually matters is judgment-per-decision. How much contextual evaluation, verification, and proportionality analysis was applied to each individual target before a weapon was launched? Maven has no score for that. The AIPCON demo didn't mention it. Cameron Stanley didn't quantify it. Because the system wasn't built to optimize for judgment. It was built to optimize for speed.
I'm not anti-military AI. I founded Daetra Technologies to build autonomous systems for defense applications. AI has a legitimate, valuable role in military operations: processing sensor data faster than humans can, identifying patterns across massive datasets, accelerating logistics and planning. But there is a categorical difference between using AI to help human analysts make better decisions faster and using AI to replace human analysts so decisions happen without them.
The first application makes the kill chain smarter. The second makes it faster and dumber simultaneously.
The Accountability Void
When Maven generates a target list, a human commander "approves" it, and a school is struck, who bears legal responsibility?
Under the law of armed conflict, the commander who ordered the strike bears responsibility for verifying the target and assessing proportionality. But when the system generates 1,000 targets in 24 hours and 20 analysts process them in rapid succession, the notion that each target received individualized legal review becomes fiction. No human being is conducting a meaningful proportionality assessment for 50 targets per analyst per day while simultaneously monitoring active strike operations.
The legal framework assumes human judgment at every stage of the targeting cycle. Maven was designed to compress that cycle. These two facts are irreconcilable, and nobody involved in the AIPCON demo attempted to reconcile them.
This gap should alarm anyone building AI systems for consequential decisions. I wrote in The AI Safety Gap No One Is Talking About about the distance between stated safety commitments and operational reality. The military context makes this gap lethal. The Pentagon's published policy is that AI assists human decision-making. The demonstrated reality is a system designed so that 20 humans approve what 2,000 used to deliberate.
The Redundancy Problem
Two weeks after Minab, the Pentagon designated Anthropic a supply chain risk and began purging Claude from classified systems. Anthropic, whose Claude was the only frontier AI model authorized for classified settings, had insisted on two contract restrictions: no mass domestic surveillance and no fully autonomous weapons.
Pentagon CTO Emil Michael described his reaction when he realized how dependent the military had become on a single AI vendor: "I'm like, holy shit, what if this software went down, some guardrail picked up, some refusal happened for the next fight like this one and we left our people at risk?"
His solution was to bring in alternatives: OpenAI, xAI, eventually Google. "I just want all of them," he said. "I need redundancy."
The irony is staggering. The Pentagon is demanding AI redundancy for its software supply chain while simultaneously eliminating human redundancy from its targeting chain. Emil Michael panicked when he realized the military depended on a single AI model. Nobody panicked when the military reduced its targeting analysts by 99%.
As I wrote in the Anthropic vendor concentration piece: never remove redundancy from a critical system. That principle doesn't apply only to software vendors. It applies to the human beings whose judgment stands between a targeting recommendation and a dead child.
Meanwhile, the company that had contractual restrictions against autonomous weapons got blacklisted. The system designed to remove human judgment from the kill chain got a standing ovation at AIPCON. And a top robotics engineer at OpenAI, Caitlin Kalinowski, resigned citing concerns about "surveillance of Americans without judicial oversight and lethal autonomy without human authorization."
She's not alone. Scientists from OpenAI and Google DeepMind, Anthropic's two biggest competitors, filed an amicus brief supporting Anthropic's legal challenge against the supply chain designation. When your competitors defend you in court, the principle at stake is bigger than any one company.
The Wrong Question
The defense establishment is asking: how do we make the kill chain faster? Palantir is answering: eliminate the humans. The Pentagon is celebrating: 900 strikes in 12 hours.
Nobody in that room asked the question that 165 dead civilians in Minab demand: how do we make the kill chain more accurate when the intelligence is wrong?
Speed is not judgment. Automation is not verification. A 99% reduction in human analysts is not a capability gain if the 1% remaining can't catch the mistakes the 99% used to catch. And a system that generates 1,000 targets in 24 hours with 20 people reviewing them hasn't solved the targeting problem. It has solved a throughput problem while creating a judgment catastrophe.
In EOD, the operators who got people killed weren't the slow ones. They were the ones who let speed override their training, who skipped verification steps because the system said they could, who trusted the tool more than their own assessment.
Maven is an extraordinary tool. The engineers who built it solved genuinely hard technical problems. But a tool that optimizes for speed in a domain where errors are measured in children's lives needs to be evaluated by the metric that matters: not how many targets it can generate, but how many mistakes it catches before a weapon is launched.
The AIPCON demo celebrated the former. Minab revealed the cost of ignoring the latter.
